U.S. patent application number 11/902789 was filed with the patent office on 2008-06-12 for cell-based detection and differentiation of disease states.
This patent application is currently assigned to Monogen, Inc.. Invention is credited to Adrian Hirsch, Kenneth S. Hirsch, Norman J. Pressman.
Application Number | 20080139402 11/902789 |
Document ID | / |
Family ID | 23049031 |
Filed Date | 2008-06-12 |
United States Patent
Application |
20080139402 |
Kind Code |
A1 |
Pressman; Norman J. ; et
al. |
June 12, 2008 |
Cell-based detection and differentiation of disease states
Abstract
The present invention provides a method for detecting and
differentiating disease states with high sensitivity and
specificity. The method allows for a determination of whether a
cell-based sample contains abnormal cells and, for certain
diseases, is capable of determining the histologic type of disease
present. The method detects changes in the level and pattern of
expression of the molecular markers in the cell-based sample. Panel
selection and validation procedures are also provided.
Inventors: |
Pressman; Norman J.;
(Glencoe, IL) ; Hirsch; Kenneth S.; (Redwood City,
CA) ; Hirsch; Adrian; (Redwood City, CA) |
Correspondence
Address: |
FOLEY AND LARDNER LLP;SUITE 500
3000 K STREET NW
WASHINGTON
DC
20007
US
|
Assignee: |
Monogen, Inc.
|
Family ID: |
23049031 |
Appl. No.: |
11/902789 |
Filed: |
September 25, 2007 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
10095298 |
Mar 12, 2002 |
|
|
|
11902789 |
|
|
|
|
60274638 |
Mar 12, 2001 |
|
|
|
Current U.S.
Class: |
506/9 ; 506/13;
506/23 |
Current CPC
Class: |
G01N 33/57492 20130101;
G01N 33/57423 20130101; G01N 33/56966 20130101 |
Class at
Publication: |
506/9 ; 506/13;
506/23 |
International
Class: |
C40B 40/10 20060101
C40B040/10; C40B 50/00 20060101 C40B050/00; C40B 30/04 20060101
C40B030/04 |
Claims
1. A panel for detecting a generic disease state or discriminating
between specific disease states using cell-based diagnosis,
comprising a plurality of probes each of which specifically binds
to a marker associated with a generic or specific disease state,
wherein the pattern of binding of the component probes of the panel
to cells in a cytology specimen is diagnostic of the presence or
specific nature of said disease state.
2. The panel of claim 1, wherein said generic disease state is
selected from the group consisting of cancer and infectious
diseases.
3. The panel of claim 2, wherein said cancer is selected from the
group consisting of epithelial cell-based cancers, solid
tumor-based cancers, secretory tumor based cancers, and blood based
cancers.
4. The panel of claim 2 wherein said infectious disease is selected
from the group consisting of cell-based diseases in which the
infectious organism is a virus, bacterium, protozoan, parasite, or
fungus.
5. The panel of claim 1, wherein said panel is optimized by using
weighting factors selected from the group consisting of cost,
prevalence of a generic disease state in a geographic location,
prevalence of a specific disease state in a geographic location,
availability of probes and commercial considerations.
6. The panel of claim 1, wherein each of said probes comprises a
detectable label.
7. The panel of claim 6, wherein said probes comprise
antibodies.
8. The panel of claim 6, wherein said label is selected from the
group consisting of a chromophore, a fluorophore, a dye, a
radioisotope and an enzyme.
9. The panel of claim 8, wherein said label is a chromophore
detected using electromagnetic radiation selected from the group
consisting of beta rays, gamma rays, X rays, ultraviolet radiation,
visible light, infrared radiation and microwaves.
10. The panel of claim 1, wherein said pattern of binding is
detected using photonic microscopy.
11. The panel of claim 10, wherein said photonic microscopy
utilizes at least one electromagnetic radiation selected from the
group consisting of gamma rays, X rays, beta rays, ultraviolet
radiation, visible light, infrared radiation and microwaves.
12. The panel of claim 1, wherein said detecting is for sexually
transmitted diseases and said discriminating is between chlamydia,
trichomonas, gonorrhea, herpes and syphilis.
13. A method of forming a panel for detecting a disease state or
discriminating between disease states in a patient using cell-based
diagnosis, comprising: (a) determining the sensitivity and
specificity of binding of probes each of which specifically binds
to a member of a library of markers associated with a disease
state; and (b) selecting a limited plurality of said probes whose
pattern of binding is diagnostic for the presence or specific
nature of said disease state.
14. The method of claim 13, wherein said determining comprises: (a)
separately contacting a histological or cytological sample from a
patient known to be suffering from said disease and a histological
or cytological sample from a patient known not to be suffering from
said disease with each of said probes; (b) measuring the amount of
specific binding of each probe with its complementary disease
marker at loci where said marker is known to be present in cells of
said samples; and (c) correlating each said amount with the
presence or specific nature of said disease.
15. The method of claim 13, wherein said selecting comprises one or
more of statistical analytical methods, pattern recognition methods
and neural network analysis.
16. The method of claim 13, where said selecting comprises the use
of weighting factors.
17. A method of detecting a disease or discriminating between
disease states comprising: (a) contacting a cytological sample
suspected of containing abnormal cells characteristic of a disease
state with a panel according to claim 1; and (b) detecting a
pattern of binding of said probes that is diagnostic for the
presence or specific nature of said disease state.
18. The method of claim 17, wherein said cytological sample is a
cellular sample collected from a body fluid, an epithelial
cell-based organ system, a fine needle aspiration or a biopsy.
19. The method of claim 18, wherein said cytological sample is
sputum.
20. A panel for detecting a generic disease state or discriminating
between specific disease states using cell-based diagnosis, wherein
said panel is formed according to the method of claim 14.
21. The panel of claim 1, wherein said disease marker is selected
from the group consisting of a morphologic biomarker, a genetic
biomarker, a cell cycle biomarker, a molecular biomarker and a
biochemical biomarker.
22. The panel of claim 3, wherein said epithelial cell-based cancer
is from the pulmonary, urinary, gastrointestinal or genital
tract.
23. The panel of claim 3, wherein said solid tumor-based cancer is
selected from the group consisting of a sarcoma, breast cancer,
pancreatic cancer, liver cancer, kidney cancer, thyroid cancer, and
prostate cancer.
24. The panel of claim 3, wherein said secretory tumor-based cancer
is selected from the group consisting of a sarcoma, breast cancer,
pancreatic cancer, liver cancer, kidney cancer, thyroid cancer, and
prostate cancer.
25. The panel of claim 3, wherein said blood-based cancer is
selected from the group consisting of leukemia and lymphoma.
26. The method of claim 18, wherein said body fluid is selected
from the group consisting of blood, urine, spinal fluid and
lymph.
27. The method of claim 18, wherein said epithelial cell based
organ system is selected from the group consisting of the pulmonary
tract, the urinary tract, the genital tract and the
gastrointestinal tract.
28. The method of claim 18, wherein said final needle aspiration is
from solid tissue types in organs and systems.
29. The method of claim 18, wherein said biopsy is from solid
tissue types in organs and systems.
30. The method of claim 28 or 29, wherein said organs and systems
are selected from the group consisting of breast, pancreas, liver,
kidney, thyroid, bone marrow, muscle, prostate and lung.
31. The panel of claim 21, wherein said morphologic biomarker is
selected from the group consisting of DNA ploidy, MACs, and
premalignant lesions.
32. The panel of claim 21, wherein said genetic biomarker is
selected from the group consisting of DNA adducts, DNA mutations
and apoptotic indices.
33. The panel of claim 21, wherein said cell cycle biomarker is
selected from the group consisting of cellular proliferation
markers, differentiation markers, regulatory molecules and
apoptosis markers.
34. The panel of claim 21, wherein said molecular biomarker or
biochemical biomarker is selected from the group consisting of
oncogenes, tumor suppressor genes, tumor antigens, growth factors
and receptors, enzymes, proteins, prostaglandins and adhesion
molecules.
Description
BACKGROUND OF THE INVENTION
[0001] The present invention relates to early detection of a
general disease state in a patient. The present invention also
relates to discrimination (differentiation) between specific
disease states in their early stages.
[0002] Early detection of a specific disease state can greatly
improve a patient's chance for survival by permitting early
diagnosis and early treatment while the disease is still localized
and its pathologic effects limited anatomically and
physiologically. Two key evaluative measures of any test or disease
detection method are its sensitivity (Sensitivity=True
Positives/(True Positives+False Negatives) and specificity
(Specificity=True Negatives/(False Positives+True Negatives), which
measure how well the test performs to accurately detect all
affected individuals without exception, and without falsely
including individuals who do not have the target disease.
Historically, many diagnostic tests have been criticized due to
poor sensitivity and specificity.
[0003] Sensitivity is a measure of a test's ability to detect
correctly the target disease in an individual being tested. A test
having poor sensitivity produces a high rate of false negatives,
i.e., individuals who have the disease but are falsely identified
as being free of that particular disease. The potential danger of a
false negative is that the diseased individual will remain
undiagnosed and untreated for some period of time, during which the
disease may progress to a later stage wherein treatments, if any,
may be less effective. An example of a test that has low
sensitivity is a protein-based blood test for HIV. This type of
test exhibits poor sensitivity because it fails to detect the
presence of the virus until the disease is well established and the
virus has invaded the bloodstream in substantial numbers. In
contrast, an example of a test that has high sensitivity is
viral-load detection using the polymerase chain reaction (PCR).
High sensitivity is achieved because this type of test can detect
very small quantities of the virus (see Lewis, D. R. et al.
"Molecular Diagnostics: The Genomic Bridge Between Old and New
Medicine: A White Paper on the Diagnostic Technology and Services
Industry" Thomas Weisel Partners, Jun. 13, 2001).
[0004] Specificity, on the other hand, is a measure of a test's
ability to identify accurately patients who are free of the disease
state. A test having poor specificity produces a high rate of false
positives, i.e., individuals who are falsely identified as having
the disease. A drawback of false positives is that they force
patients to undergo unnecessary medical procedures treatments with
their attendant risks, emotional and financial stresses, and which
could have adverse effects on the patient's health. A feature of
diseases which makes it difficult to develop diagnostic tests with
high specificity is that disease mechanisms often involve a
plurality of genes and proteins. Additionally, certain proteins may
be elevated for reasons unrelated to a disease state. An example of
a test that has high specificity is a gene-based test that can
detect a p53 mutation. A p53 mutation will never be detected unless
there are cancer cells present (see Lewis, D. R. et al. "Molecular
Diagnostics: The Genomic Bridge Between Old and New Medicine: A
White Paper on the Diagnostic Technology and Services Industry"
Thomas Weisel Partners, Jun. 13, 2001).
[0005] Cellular markers are naturally occurring molecular
structures within cells that can be discovered and used to
characterize or differentiate cells in health and disease. Their
presence can be detected by probes, invented and developed by human
beings, which bind to markers enabling the markers to be detected
through visualization and/or quantified using imaging systems. Four
classes of cell-based marker detection technologies are
cytopathology, cytometry, cytogenetics and proteomics, which are
identified and described below.
[0006] Cytopathology relies upon the visual assessment by human
experts of cytomorphological changes within stained whole-cell
populations. An example is the cytological screening and
cytodiagnosis of Papanicolaou-stained cervical-vaginal specimens by
cytotechnologists and cytopathologists, respectively. Unlike
cytogenetics, proteomics and cytometry, cytopathology is not a
quantitative tool. While it is the state-of-the-art in clinical
diagnostic cytology, it is subjective and the diagnostic results
are often not highly sensitive or reproducible, especially at early
stages of cancer (e.g., ASCUS, LSIL).
[0007] Tests that rely on morphological analyses involve observing
a sample of a patient's cells under a microscope to identify
abnormalities in cell and nuclear shape, size, or staining
behavior. When viewed through a microscope, normal mature
epithelial cells appear large and well differentiated, with
condensed nuclei. Cells characterized by dysplasia, however, may be
in a variety of stages of differentiation, with some cells being
very immature. Finally, cells characterized by invasive carcinoma
often appear undifferentiated, with very little cytoplasm and
relatively large nuclei.
[0008] A drawback to diagnostic tests that rely on morphological
analyses is that cell morphology is a lagging indicator. Since form
follows function, often the disease state has already progressed to
a critical stage by the time the disease becomes evident by
morphological analysis. The initial stages of a disease involve
chemical changes at a molecular level. Changes that are detectable
by viewing cell features under a microscope are not apparent until
later stages of the disease. Therefore, tests that measure chemical
changes on a molecular level, referred to as "molecular diagnostic"
tests, are more likely to provide early detection than tests that
rely on morphological analyses alone.
[0009] Cytometry is based upon the flow-microfluorometric
instrumental analysis of fluorescently stained cells moving in
single file in solution (flow cytometry) or the computer-aided
microscope instrumental analysis of stained cells deposited onto
glass microscope slides (image cytometry). Flow cytometry
applications include leukemia and lymphoma immunophenotyping. Image
cytometry applications include DNA ploidy, Malignancy-Associated
Changes (MACs) and S-phase analyses. The flow and image cytometry
approaches yield quantitative data characterizing the cells in
suspension or on a glass microscope slide. Flow and image cytometry
can produce good marker detection and differentiation results
depending upon the sensitivity and specificity of the cellular
stains and flow/image measurement features used.
[0010] Malignancy-Associated Changes (MACs) have been qualitatively
observed and reported since the early to mid-1900's (O C Gruner:
"Study of the changes met with leukocytes in certain cases of
malignant disease" in Brit J Surg 3: 506-522, 1916) (HE Neiburgs, F
G Zak, D C Allen, H Reisman, T Clardy: "Systemic cellular changes
in material from human and animal tissues" in Transactions,
7.sup.th Ann Mtg Inter Soc Cytol Council, pp 137-144, 1959). From
the mid-1900's through 1975, MACs were documented in independent
qualitative histology and cytology studies in buccal mucosa and
buccal smears (Nieburgs, Finch, Klawe), duodenum (Nieburgs), liver
(Elias, Nieburgs), megakaryocytes (Ramsdahl), cervix (Nieburgs,
Howdon), skin (Kwitiken), blood and bone marrow (Nieburgs),
monocytes and leukocytes (van Haas, Matison, Clausen), and lung and
sputum (Martuzzi and Oppen Toth). Before 1975 these qualitative
studies reported MAC-based sensitivities for specific disease
detection from 76% to 97% and specificities from 50% to 90%. In
1975 Oppen Toth reported a sensitivity of 76% and specificity of
81% in a qualitative sputum analysis study.
[0011] Quantitative observations regarding MAC-based probe analysis
began two to three decades ago (H Klawe, J Rowinski: "Malignancy
associated changes (MAC) in cells of buccal smears detected by
means of objective image analysis" in Acta Cytol 18: 30-33, 1974)
(G L Wied, P H Bartels, M Bibbo, J J Sychra: "Cytomorphometric
markers for uterine cancer in intermediate cells" in Analyt Quant
Cytol 2: 257-263, 1980) (G Burger, U Jutting, K Rodenacker:
"Changes in benign population in cases of cervical cancer and its
precursors" in Analyt Quant Cytol 3: 261-271, 1981). MACs were
documented in independent quantitative histology and cytology
studies in buccal mucosa and smears Klawe, Burger), cervix (Wied,
Burger, Bartels, Vooijs, Reinhardt, Rosenthal, Boon, Katzke,
Haroske, Zahniser), breast (King, Bibbo, Susnik), bladder and
prostate (Sherman, Montironi), colon (Bibbo), lung and sputum
(Swank, MacAulay, Payne), and nasal mucosa (Reith) studies with
MAC-based sensitivities from 70% to 89% and specificities from 52%
to 100%. Marek and Nakhosteen showed (1999, American Thoracic
Society annual meeting) the results from two quantitative pulmonary
studies showing (a) sensitivity of 89% and specificity of 92%, and
(b) sensitivity of 91% and specificity of 100%.
[0012] Clearly, Malignancy-Associated Changes (MACs) are
potentially useful probes that result from the image-cytometry
marker detection technology. MAC-based features from DNA-stained
nuclei can be used in conjunction with other molecular diagnostic
probes to create optimized molecular diagnostic panels for the
detection and differentiation of lung cancer and other disease
states.
[0013] Cytogenetics detects specific chromosome-based intracellular
changes using, for example, in situ hybridization (ISH) technology.
ISH technology can be based upon fluorescence (FISH), multi-color
fluorescence (M-FISH), or light-absorption-based chromogenics
imaging (CHRISH) technologies. The family of ISH technologies uses
DNA or RNA probes to detect the presence of the complementary DNA
sequence in cloned bacterial or cultured eukaryotic cells. FISH
technology can, for example, be used for the detection of genetic
abnormalities associated with certain cancers. Examples include
probes for Trisomy 8 and HER-2 neu. Other technologies such as
polymerase chain reactions (PCR) can be used to detect B-cell and
T-cell gene rearrangements. Cytogenetics is a highly specific
marker detection technology since it detects the causative or
"trigger" molecular event producing a pathology condition. It may
be less sensitive than the other marker detection technologies
because fewer events may be present to detect. In situ
hybridization (ISH) is a molecular diagnostic method uses
gene-based analyses to detect abnormalities on the genetic level
such as mutations, chromosome errors or genetic material inserted
by a specific pathogen. For example, in situ hybridization may
involve measuring the level of a specific mRNA by treating a sample
of a patient's cells with labeled primers designed to hybridize to
the specific mRNA, washing away unbound primers and measuring the
signal of the label. Due to the uniqueness of gene sequences, a
test involving the detection of gene sequences will likely have a
high specificity, yielding very few false positives. However,
because the amount of genetic material in a sample of cells may be
very low, only a very weak signal may be obtained. Therefore, in
situ hybridization tests that do not employ pre-amplification
techniques will likely have a poor specificity, yielding many false
negatives.
[0014] Proteomics depends upon cell characterization and
differentiation resulting from the over-expression,
under-expression, or presence/absence of unique or specific
proteins in populations of normal or abnormal cell types.
Proteomics includes not only the identification and quantification
of proteins, but also the determination of their localization,
modifications, interactions, chemical activities, and
cellular/extracellular functions. Immunochemistry
(immunocytochemistry in cells and immunohistochemistry (IHC) in
tissues) is the technology used, either qualitatively or
quantitatively (QIHC) to stain antigens (i.e., proteomes) using
antibodies. Immunostaining procedures use a dye as the detection
indicator. Examples of IHC applications include analyses for ER
(estrogen receptor), PR progesterone receptor), p53 tumor
suppressor genes, and EGRF prognostic markers. Proteomics is
typically a more sensitive marker detection technology than
cytogenetics because there are often orders of magnitude more
protein molecules to detect using proteomics than there are
cytogenetic mutations or gene-sequence alterations to detect using
cytogenetics. However, proteomics may have a poorer specificity
than the cytogenetic marker detection technology since multiple
pathologies may result in similar changes in protein
over-expression or under-expression. Immunochemistry involves
histological or cytological localization of immunoreactive
substances in tissue sections or cell preparations, respectively,
often utilizing labeled antibodies as probe reagents.
Immunochemistry can be used to measure the concentration of a
disease marker (specific protein) in a sample of cells by treating
the cells with an agent such as a labeled antibody (probe) that is
specific for an epitope on the disease marker, then washing away
unbound antibodies and measuring the signal of the label.
Immunochemistry is based on the property that cancer cells possess
different levels of certain disease markers than do healthy cells.
The concentration of a disease marker in a cancer cell is generally
large enough to produce a large signal. Therefore, tests that rely
on immunochemistry will likely have a high sensitivity, yielding
few false negatives. However, because other factors in addition to
the disease state may cause the concentration of a disease marker
to become raised or lowered, tests that rely on immunochemical
analysis of a specific disease marker will likely have poor
specificity, yielding a high rate of false positives.
[0015] The present invention provides a noninvasive disease state
detection and discrimination method with both high sensitivity and
high specificity. The method involves contacting a cytological
sample suspected of containing diseased cells with a panel of
probes comprising a plurality of agents, each of which
quantitatively binds to a specific disease marker, and detecting
and analyzing the pattern of binding of the probe agents. The
present invention also provides methods of constructing and
validating a panel of probes for detecting a specific disease (or
group of diseases) and discriminating among its various disease
states. Illustrative panels for detecting lung cancer and
discriminating among different types of lung cancer are also
provided.
[0016] A human disease results from the failure of the human
organism's adaptive mechanisms to neutralize external or internal
insults which result in abnormal structures or functions within the
body's cells, tissues, organs or systems. Diseases can be grouped
by shared mechanisms of causation as illustrated below, in Table
1.
TABLE-US-00001 TABLE 1 Classes of Diseases Examples of Disease
States Allergy Adverse reactions to foods and plants Cardiovascular
Heart failure, atherosclerosis Degenerative (neurological and
Alzheimer's and Parkinson's muscular) Diet Non-nutritional
substances and excess/imbalanced nutrition Hereditary Sickle cell
anemia, cystic fibrosis Immune HIV and autoimmune Infection Viral,
bacterial, fungal, parasitic Metabolic Diabetes Molecular and cell
biology Cancer (neoplasia) Toxic insults Alcohol, drugs,
environmental mutagens and carcinogens Trauma Bodily injury from
automobile collision
[0017] Disease states are either caused by or result in abnormal
changes (i.e., pathological conditions) at a subcellular, cellular,
tissue, organ, or human anatomic or physiological system level.
Many disease states (e.g., lung cancer) are characterized by
abnormal changes at a subcellular or cellular level. Specimens
(e.g., cervical PAP smears, voided urine, blood, sputum, colonic
washings) can be collected from patients with suspected disease
states to diagnose those patients for the presence and type of the
disease state. Molecular pathology is the discipline that attempts
to identify and diagnostically exploit the molecular changes
associated with these cell-based diseases.
[0018] Lung cancer is an illustrative example of a disease state in
which screening of high-risk populations and at-risk individuals
can be performed using diagnostic tests (e.g., molecular diagnostic
panel assays) to detect the presence of the disease state. Also,
for patients in which lung cancer or other disease states have been
detected by these means, related diagnostic tests can be employed
to differentiate the specific disease state from related or
co-occurring disease states. For example, in this lung cancer
illustration, additional molecular diagnostic panel assays may
indicate the probabilities that the patient's disease state is
consistent with one of the following types of lung cancer: (a)
squamous cell carcinoma of the lung, (b) adenocarcinoma of the
lung, (c) large cell carcinoma of the lung, (d) small cell
carcinoma of the lung, or (e) mesothelioma. Early detection and
differentiation of cell-based disease states is a hypothesized
means to improve patient outcomes.
[0019] Cancer is a neoplastic disease the natural course of which
is fatal. Cancer cells, unlike benign tumor cells, exhibit the
properties of invasion and metastasis and are highly anaplastic.
Cancer includes the two broad categories of carcinoma and sarcoma,
but in normal usage it is often used synonymously with carcinoma.
According to the World Health Organization (WHO), cancer affects
more than 10 million people each year and is responsible for in
excess of 6.2 million deaths.
[0020] Cancer is, in reality, a heterogeneous collection of
diseases that can occur in virtually any part of the body. As a
result, different treatments are not equally effective in all
cancers or even among the stages of a specific type of cancer.
Advances in diagnostics (e.g., mammography, cervical cytology, and
serum PSA testing) have, in some cases, allowed for the detection
of early-stage cancer when there are a greater number of treatment
options, and therapies tend to be more effective. In cases where a
solid tumor is small and localized, surgery alone may be sufficient
to produce a cure. However, in cases where the tumor has spread,
surgery may provide, at best, only limited benefits. In such cases
the addition of chemotherapy and/or radiation therapy may be used
to treat metastatic disease. While somewhat effective in prolonging
life, treatment of patients with metastatic disease rarely produces
a cure. Even through there may be an initial response, with time
the disease progresses and the patient ultimately dies from its
effects and/or from the toxic effects of the treatments.
[0021] While not proven, it is generally accepted that early
detection and treatment will reduce the morbidity, mortality and
cost of cancer. Early detection will, in many cases, permit
treatment to be initiated prior to metastasis. Furthermore, because
there are a greater number of treatment options, there is a higher
probability of achieving a cure or significant improvement in
long-term survival.
[0022] Developing a test that can be used to screen an "at-risk"
population has long been a goal of health practitioners. While
there have been some successes such as mammography for breast
cancer, PSA testing for prostate cancer, and the PAP smear for
cervical cancer, in most cases cancer is detected at a relatively
late stage where the patient is symptomatic and the disease is
almost always fatal. For most cancers, no test or combination of
tests has exhibited the necessary sensitivity and specificity to
permit cost-effective identification of patients with early stage
disease.
[0023] For a cancer screening program to be successful, and gain
acceptance by patients, physicians, and third party payers, the
test must have implied benefit (changes the outcome), be widely
available and be able to be carried out readily within the
framework of general healthcare. The test should be relatively
noninvasive, leading to adequate compliance, have high sensitivity,
and reasonable specificity and predictive value. In addition, the
test must be available at relatively low cost.
[0024] For patients who are suspected of having cancer, the
diagnosis must be confirmed and the tumor properly staged
cytologically and clinically in order for physicians to undertake
appropriate therapeutic intervention. Some tests currently being
used in the diagnosis and staging of cancer, however, either lack
sufficient sensitivity or specificity, are too invasive, or are too
costly to justify their use as a population-based screening test.
Shown below in Tables 2 and 3, for example, are estimates of
sensitivity and specificity of lung cancer diagnostics and
estimated costs for diagnostic tests used to detect lung
cancer.
TABLE-US-00002 TABLE 2 ESTIMATES OF SENSITIVITY AND SPECIFICITY OF
LUNG CANCER DIAGNOSTICS [1] SPECIFICITY DIAGNOSTIC TEST SENSITIVITY
(%) (%) Conventional Sputum Cytology 51.0 100.0 Chest X-ray 16-85*
90-95 White Light Bronchoscopy 48.0-80.0 91.1-96.8 LIFE
Bronchoscopy 72.0 86.7 Computed Tomography 63.0-99.9 80.0-61 PET
Scan 88.0-92.5 83.0-93.0 *Dependent upon the stage of the disease
at the time of diagnosis
TABLE-US-00003 TABLE 3 ESTIMATED COSTS FOR DIAGNOSTIC TESTS USED IN
LUNG CANCER [1] DIAGNOSTIC TEST COST ($) Sputum Cytology 90 Chest
X-ray 44 Bronchoscopy 725 Computed Tomography 378 PET Scan 800-3000
Open Biopsy 12,847-14,121
[0025] The chest radiograph (X-ray) is often used to detect and
localize cancer lesions due to its reasonable sensitivity, high
specificity and low cost. However, small lesions are often
difficult to detect and although larger tumors are relatively easy
to visualize on a chest film, at the time of detection most have
already metastasized. Thus, chest X-rays lack the necessary
sensitivity for use as an early detection method.
[0026] Computed tomography (CT) is useful in the confirmation and
characterization of pulmonary nodules and allows the detection of
subtle abnormalities that are often missed on a standard chest
X-ray [2]. CT, and Spiral CT methods in particular, remains the
test of choice for patients who present with a prior malignant
sputum cytology result or vocal chord paralysis. CT, with its
improved sensitivity over the conventional chest film, has become
the primary tool for imaging the central airway [3]. While capable
of examining large areas, CT is subject to artifacts from cardiac
and respiratory motion although improved resolution can be achieved
through the use of iodinated contrast material.
[0027] Spiral CT is a more rapid and sensitive form of CT that has
the potential to detect early cancer lesions more reliably than
either conventional CT or X-ray. Spiral CT appears to have greatly
improved sensitivity in diagnosing early disease. However, the test
has relatively low specificity with a 20% false positive rate [4].
Spiral CT is also less sensitive in detecting the central lesions
that represent one-third of all lung cancers. Furthermore, while
the cost of the initial test is relatively low ($300), the cost of
follow-up can be high. Cytology using molecular diagnostic panel
assays offers significant promise as an adjunctive test with Spiral
CT to improve the specificity of Spiral CT testing by minimizing
false positive results through the evaluation of fine needle
aspirations (FNAs) or biopsies (FNBs) from Spiral CT-suspicious
pulmonary nodules.
[0028] Fluorescence bronchoscopy provides increased sensitivity
over conventional white light bronchoscopy, significantly improving
the detection of small lesions within the central airway [5].
However, fluorescence bronchoscopy is unable to detect peripheral
lesions, it takes a long time for bronchoscopists to examine a
patient's airways, and it is an expensive procedure. Additionally,
the procedure is moderately invasive, creating an insurmountable
barrier to its use as a population-based screening test.
[0029] Positron Emission Tomography (PET) is a highly sensitive
test that utilizes radioactive glucose to identify the presence of
cancer cells within the lung [6-8]. The cost of establishing a
testing facility is high and there is the need for a cyclotron on
site or nearby. This, coupled with the high cost of the test, has
limited the use of PET scans to staging lung cancer patients rather
than for early detection of the disease.
[0030] Although used for some time as a means of screening for lung
cancer, sputum cytology has enjoyed only limited success due to its
low sensitivity and its failure to reduce disease-specific
mortality. In conventional sputum cytology, the pathologist uses
characteristic changes in cellular morphology to identify malignant
cells and make a diagnosis of cancer. Today only 15% of patients
who are "at-risk" or who are suspected of having lung cancer
undergo sputum cytology testing, and less than 5% undergo multiple
evaluations [9]. A number of factors including tumor size,
location, degree of differentiation, cell clumping, inefficiency of
clearing mechanisms to release cells and sputum to the external
environment, and the poor stability of cells within the sputum
contribute to the overall poor performance of the test.
[0031] Cancer diagnostics has traditionally relied upon the
detection of single molecular markers. Unfortunately, cancer is a
disease state in which single markers have typically failed to
detect or differentiate many forms of the disease. Thus, probes
that recognize only a single marker have been shown to be largely
ineffective. Exhaustive searches for "magic bullet" diagnostic
tests have been underway for many decades though no universal
successful magic bullet probes have been found to date.
[0032] A major premise of this invention is that cell-based cancer
diagnostics and the screening, diagnosis for, and therapeutic
monitoring of other disease states will be significantly improved
over the state-of-the-art that uses single marker/probe analyses
rather than kits of multiple, simulaneously labeled probes. This
multiplexed analytical approach is particularly well suited for
cancer diagnostics since cancer is not a single disease.
Furthermore, this multi-factorial "panel" approach is consistent
with the heterogeneous nature of cancer, both cytologically and
clinically.
[0033] Key to the successful implementation of a panel approach to
cell-based diagnostic tests is the design and development of
optimized panels of probes that can chemically recognize the
pattern of markers that characterizes and distinguishes a variety
of disease states. This patent application describes an efficient
and unique methodology to design and develop such novel and
optimized panels.
[0034] Improved methods for specimen collection (e.g.,
point-of-care mixers for sputum cytology) and preparation (e.g.,
new cytology preservation and transportation fluids, and
liquid-based cytology preparation instruments) are under
development and becoming commercially available. In conjunction
with existing and these emerging methods, a successful
implementation of this molecular diagnostics cell-based panel assay
will lead to (a) characterization of the molecular profile of
malignant tumors and other disease states, (b) improved methods for
early cancer and other disease state detection and differentiation,
and (c) opportunities for improved clinical diagnoses, prognoses,
customized patient treatments, and therapeutic monitoring.
SUMMARY OF THE INVENTION
[0035] The present invention is directed to a panel for detecting a
generic disease state or discriminating between specific disease
states using cell-based diagnosis. The panel comprises a plurality
of probes each of which specifically binds to a marker associated
with a generic or specific disease state, wherein the pattern of
binding of the component probes of the panel to cells in a cytology
specimen is diagnostic of the presence or specific nature of said
disease state. The present invention is also directed to a method
of forming a panel for detecting a disease state or discriminating
between disease states in a patient using cell-based diagnosis. The
method involves determining the sensitivity and specificity of
binding of probes each of which specifically binds to a member of a
library of markers associated with a disease state and selecting a
limited plurality of said probes whose pattern of binding is
diagnostic for the presence or specific nature of said disease
state. The present method is also directed to a method of detecting
a disease or discriminating between disease states comprising. The
method involves contacting a cytological sample suspected of
containing abnormal cells characteristic of a disease state with a
panel according to claim 1 and detecting a pattern of binding of
said probes that is diagnostic for the presence or specific nature
of said disease state.
BRIEF DESCRIPTION OF THE FIGURES
[0036] FIG. 1. Molecular markers that are preferable markers to be
included in a panel for identifying different histologic types of
lung cancer. The column labeled "%" indicates the percentage of
tumor specimens that express a particular marker.
[0037] FIG. 2. Potential ways in which different markers may be
used to discriminate between specific types of lung cancer. SQ
indicates squamous cell carcinoma, AD indicates adenocarcinoma, LC
indicates large cell carcinoma, SC indicates small cell carcinoma
and ME indicates mesothelioma. The numbers appearing in each cell
represent frequency of marker change in one cell type versus
another. To be included in the table, the ratio must be greater
than 2.0 or less than 0.5. A number larger than 100 generally
indicates that the second marker is not expressed. In such cases
the denominator was set at 0.1 for the purpose of the analysis.
Finally, empty cells represent either no difference in expression
or the absence of expression data.
[0038] FIG. 3. Comparisons between H-scores for probes 7 and 15 in
control tissue and in cancerous tissue. The x-axis shows the
H-scores while the y-axis shows the percent of cases.
[0039] FIG. 4. Correlation matrix, in which correlation measures
the amount of linear association between a pair of variables. All
markers in this matrix with a correlation number of 50% or higher
are considered correlate markers.
[0040] FIG. 5. Detection panel compositions, pair-wise
discrimination panel compositions and joint discrimination panel
compositions. Panel compositions using decision tree analysis,
stepwise LR and stepwise LD are shown.
[0041] FIG. 6. Detection panel compositions wherein probe 7 was not
included as a probe. Panel compositions using decision tree
analysis, stepwise LR and stepwise LD are shown.
[0042] FIG. 7. Detection panel compositions using only commercially
preferred probes. Panel compositions using decision tree analysis,
stepwise LR and stepwise LD are shown.
DETAILED DESCRIPTION OF THE INVENTION
[0043] 1. Introduction
[0044] The present invention provides a noninvasive disease state
detection and discrimination method with high sensitivity and
specificity. The method involves contacting a cytological sample
suspected of containing diseased cells with a panel comprising a
plurality of agents, each of which quantitatively binds to a
disease marker, and detecting a pattern of binding of the agents.
This pattern includes the localization and density/concentration of
binding of the component probes of the panel. The present invention
also provides methods of making a panel for detecting a disease and
also for discriminating between disease states as well as panels
for detecting lung cancer in early stages and discriminating
between different types of lung cancer. Panel tests have been used
in medicine. For example, panels are used in blood serum analysis.
However, because a cytology analysis involves imaging and
localization of specific markers within individual cells and
tissues, prior to the present invention it was not apparent that
the panel approach would be effective for cytology samples.
Additionally, it was not apparent which, if any statistical
analyses could be applied to design and develop an optimized
cell-based diagnostic panel of probes.
[0045] One of the few examples of a cytology-based screening
program is the PAP Smear, which screens for cervical cancer. For
over 50 years this method has been practiced and has greatly
contributed to the fact that today, almost no woman who has regular
PAP smears dies of cervical cancer. There are drawbacks, however,
to the PAP smear screening program. For example, PAP smears are
labor intensive and are not universally accessible. The present
molecular diagnostic cell-based screening method utilizing probe
panels does not suffer from these drawbacks. The method may be
fully automated and thereby made less expensive, increasing access
to this type of testing.
[0046] The present invention provides a method, having both high
specificity and high sensitivity, for detecting a disease state and
for discriminating between disease states. The invention is
applicable to any cell-based disease state, such as cancer and
infectious diseases.
[0047] The panel is diagnostic of the presence or specific nature
of the disease state. The present invention overcomes the
limitations and drawbacks of known disease state detection methods
by enabling quick, accurate, relatively noninvasive and easy
detection and discrimination of diseased cells in a cytological
sample while keeping costs low.
[0048] A feature of the inventive method for making a panel of the
present invention is the rapidity with which the panel may be
developed.
[0049] There are several benefits to using a panel of agents in a
method for detecting a disease state, and for discriminating
between types of disease states. One benefit is that a panel of
agents has sufficient redundancy to permit detection and
characterization of disease states thereby increasing the
sensitivity and specificity of the test. Given the heterogeneous
nature of many disease states, no single agent is capable of
identifying the vast majority of cases.
[0050] An additional benefit to using a panel is that use of a
panel permits discrimination between the various types of a disease
state based on specific patterns (probe localization and
density/concentration) of expression. As the various types of a
disease may exhibit dramatic differences in their rate of
progression, response to therapy, and lethality, knowledge of the
specific type can help physicians choose the optimal therapeutic
approach.
[0051] 2. The Panel
[0052] The panel of the present invention comprises a plurality of
agents, each of which quantitatively binds to a disease marker,
wherein the pattern (localization and density/concentration) of
binding of the component agents of the panel is diagnostic of the
presence or specific nature of a disease state. Therefore, the
panel may be a detection panel or a discrimination panel. A
detection panel detects whether a generic disease state is present
in a sample of cells, while a discrimination panel discriminates
among different specific disease states in a sample of cells known
to be affected by a disease state which comprises different types
of diseases. The difference between a detection panel and a
discrimination panel lies in the specific agents that the panels
comprise. A detection panel comprises agents having a pattern of
binding that is diagnostic of the presence of a disease state,
while a discrimination panel comprises agents having a pattern of
binding that allows for determining the specific nature (i.e., each
type) of the disease state.
[0053] A panel, by definition, contains more than one member. There
are several reasons why it is beneficial to use a panel of markers
rather than just one marker alone to detect a generic disease state
or to discriminate among specific disease states. One reason is the
unlikely existence of a probe for one single marker, that is
present in all diseased cells yet not present in healthy cells,
whose behavior can be measured with a high specificity and
sensitivity to yeild an accurate test result. If such a single
probe existed for detection of a particular disease with high
sensitivity and specificity, it would already have been utilized
for clinical testing. Rather, it is the directed selection of panel
tests, each consisting of multiple probes, that together can
provide the range of detection capability to ensure clinically
adequate testing.
[0054] If one nevertheless chooses to construct a panel test
comprising one or a very few probes, then the failure of any single
marker/probe combination to perform its labeling function for any
reason (for example, diminished reactivity of the specimen cells
due to biological variability; inherent variability between lots of
probe reagents; a weak, outdated or defective processing reagent;
improper processing time or conditions for that probe) could result
in a catastrophic failure of the test to detect or discriminate the
target disease. The inclusion of multiple, and even redundant
probes in each panel test greatly enhances the probability that a
failure of any one probe will not cause a catastrophic failure of
the test.
[0055] A probe is any molecular structure or substructure that
binds to a disease marker. The term "agent" as used herein, may
also refer to a molecular structure or substructure that binds to a
disease marker. Molecular probes are homing devices used by
biologists and clinicians to detect and locate markers indicative
of the specific disease states. For example, antibodies may be
produced that bind specifically to a protein previously identified
as a marker for small cell lung cancer. This antibody probe can
then be used to localize the target protein marker in cells and
tissues of patients suspected of having the disease by using
appropriate immunochemical protocols and incubations. If the
antibody probe binds to its target marker in a stoichiometric
(i.e., quantitative) fashion and is labeled with a chromogenic or
colored "tag", then localization and quantitation of the probe and,
indirectly, its target marker may be accomplished using an optical
microscope and image cytometry technology.
[0056] The present invention contemplates detecting changes in
molecular marker expression at the DNA, RNA or protein level using
any of a number of methods available to an ordinary skilled
artisan. Exemplary probes may be a polyclonal or monoclonal
antibody or fragment thereof or a nucleic acid sequences that is
complementary to the nucleic acid sequence encoding a molecular
marker in the panel. A probe may also be a stain, such as a DNA
stain. Many of the antibodies used in the present invention are
specific to a variety of cell surface or intracellular antigens as
marker substances. The antibodies may be synthesized using
techniques generally known to those of skill in the art. For
example, after the initial raising of antibodies to the marker, the
antibodies can be sequenced and subsequently prepared by
recombinant techniques. Alternatively, antibodies may be
purchased.
[0057] In embodiments of the present invention, the probe contains
a label. A probe containing a label is often referred to herein as
a "labeled probe". The label may be any substance that can be
attached to a probe so that when the probe binds to the marker a
signal is emitted or the labeled probe can be detected by a human
observer or an analytical instrument. This label may also be
referred to as a "tag". The label may be visualized using reader
instrumentation. The term "reader instrumentation" refers to the
analytical equipment used to detect a probe. Labels envisioned by
the present invention are any labels that emit a signal and allow
for identification of a component in a sample. Preferred labels
include radioactive, fluorogenic, chromogenic or enzymatic
moieties. Therefore, possible methods of detection include, but are
not limited to, immunocytochemistry, immunohistochemistry, in situ
hybridization, fluorescent in situ hybridization, flow cytometry
and image cytometry. The signal generated by the labeled probe is
of sufficient intensity to permit detection by a medical
practitioner.
[0058] A "marker", "disease marker" or "molecular marker" is any
molecular structure or substructure that is correlated with a
disease state or pathogen. The term "antigen" may be used
interchangeably with "marker". Broadly defined, a marker is a
biological indicator that may be deliberately used by an observer
or instrument to reveal, detect, or measure the presence or
frequency and/or amount of a specific condition, event or
substance. For example, a specific and unique sequence of
nucleotide bases may be used as a genetic marker to track patterns
of genetic inheritance among individuals and through families.
Similarly, molecular markers are specific molecules, such as
proteins or protein fragments, whose presence within a cell or
tissue indicates a particular disease state. For example,
proliferating cancer cells may express novel cell-surface proteins
not found on normal cells of the same type, or may over-express
specific secretory proteins whose increased or decreased abundance
(e.g., overexpression or underexpression, respectively) can serve
as markers for a particular disease state.
[0059] Suitable markers for cytology panels are substances that are
localized in or on the nucleus, cytoplasm or cell membrane. Markers
may also be localized in organelles located in any of these
locations in the cell. Exemplary markers localized in the nucleus
include but are not limited to retinoblastoma gene product (Rb),
Cyclin A, nucleoside diphosphate kinase/nm23, telomerase, Ki-67,
Cyclin D1, proliferating cell nuclear antigen (PCNA), p120
(proliferation-associated nucleolar antigen) and thyroid
transcription factor 1 (TTF-1). Exemplary markers localized in the
cytoplasm include but are not limited to VEGF, surfactant
apoprotein A (SP-A), nucleoside nm23, melanoma antigen-1 (MAGE-1),
Mucin 1, surfactant apoprotein B (SP-B), ER related protein p29 and
melanoma antigen-3 (MAGE-3). Exemplary markers localized in the
cell membrane include but are not limited to VEGF, thrombomodulin,
CD44v6, E-Cadherin, Mucin 1, human epithelial related antigen
(HERA), fibroblast growth factor (FGF), heptocyte growth factor
receptor (C-MET), BCL-2, N-Cadherin, epidermal growth factor
receptor (EGFR) and glucose transporter-3 (GLUT-3). An example of a
marker located in an organelle of the cytoplasm is BCL-2, located
(in part) in the mitochondrial membrane. An example of a marker
located in an organelle of the nucleus is p120
(proliferating-associated nucleolar antigen), located in the
nucleoli.
[0060] Preferred are markers where changes in expression: occur
early in disease progression, are exhibited by a majority of
diseased cells, allow for detection of in excess of 75% of a given
disease type, most preferably in excess of 90% of a given disease
type and/or allow for the discrimination between the nature of
different types of a disease state.
[0061] It is noted that the inventive panel may be referred to as a
panel of probes or a panel of markers, since the probes bind to the
markers. Therefore, the panel may comprise a number of markers or
it may comprise a number of probes that bind to specific markers.
For the sake of consistency, the present panel is referred to as a
panel of probes; however, it could also be referred to as a panel
of markers.
[0062] Markers can also include features such as
malignancy-associated changes (MACs) in the cell nucleus or
features related to the patient's family history of cancer.
Malignancy-associated changes, or MACs, are typically sub-visual
changes that occur in normal-appearing cells located in the
vicinity of cancer cells. These exceedingly subtle changes in the
cell nucleus may result biologically from changes in the nuclear
matrix and the chromatin distribution pattern. They cannot be
appreciated even by trained observers through the visual
observation of individual cells, but may be determined from
statistical analysis of cell populations using highly automated,
computerized high-speed image cytometry. Techniques for detection
of MACs are well known to those of skill in the art and are
described in more detail in: Gruner, O. C. Brit J. Surg. 3 506-522
(1916); Neiburgs, H. E. et al., Transaction, 7.sup.th Annual Mtg.
Inter. Soc. Cytol. Council 137-144 (1959); Klawe, H. Acta. Cytol.
18 30-33 (1974); Wied, G. L., et al., Analty. Quant. Cytol. 2
257-263 (1980); and Burger, G., et al., Analyt. Quant. Cytol. 3
261-271 (1981).
[0063] The present invention encompasses any marker that is
correlated with a disease state. The individual markers themselves
are mere tools of the present invention. Therefore, the invention
is not limited to specific markers. One way to classify markers is
by their functional relationship to other molecules. As used
herein, a "functionally related" marker is a component of the same
biological process or pathway as the marker in question and would
be known by a person of skill in the art to be abnormally expressed
together with the marker in question. For example, many markers are
associated with a cell proliferation pathway, such as fibrobast
growth factor (FGF), (vascular endothelial growth factor) VEGF,
CyclinA and Cyclin D1. Other markers are glucose transporters, such
as Glut-1 and Glut-3.
[0064] A person of ordinary skill in the art is well equipped to
determine a functionally related marker and may research various
markers or perform experiments in which the functional behavior of
a marker is determined. By way of non-limiting example, a marker
may be classified as a molecule involved in angiogenesis, a
transmembrane glycoprotein, a cell surface glycoprotein, a
pulmonary surfactant protein, a nuclear DNA-binding phosphoprotein,
a transmembrane Ca.sup.2+ dependent cell adhesion molecule, a
regulatory subunit of the cyclin-dependent kinases (CDK's), a
nucleoside diphosphate kinase, a ribonucleoprotein enzyme, a
nuclear protein that is expressed in proliferating normal and
neoplastic cells, a cofactor for DNA polymerase delta, a gene that
is silent in normal tissues yet when it is expressed in malignant
neoplasms is recognized by autologous, tumor-directed and specific
cytotoxic T cells (CTL's), a glycosylated secretory protein, the
gastrointestinal tract or genitourinary tract, a hydrophobic
protein of a pulmonary surfactant, a transmembrane glycoprotein, a
molecule involved in proliferation, differentiation and
angiogenesis, a proto-oncogene, a homeodomain transcription factor,
a mitochondrial membrane protein, a molecule found in nucleoli of a
rapidly proliferating cell, a glucose transporter, or an
estrogen-related heat shock protein.
[0065] Classes of biomarkers and probes include, but are not
limited to: (a) morphologic biomarkers, including DNA ploidy, MACs
and premalignant lesions; (b) genetic biomarkers including DNA
adducts, DNA mutations and apoptotic indices; (c) cell cycle
biomarkers including cellular proliferation, differentiation,
regulatory molecules and apoptosis markers, and; (d) molecular and
biochemical biomarkers including oncogenes, tumor suppressor genes,
tumor antigens, growth factors and receptors, enzymes, proteins,
prostaglandin levels and adhesion molecules.
[0066] A "disease state" may be any cell-based disease. In some
embodiments the disease state is cancer. In other embodiments, the
disease state is an infectious disease. The cancer may be any
cancer, including, but not limited to epithelial cell-based cancers
from the pulmonary, urinary, gastrointestinal, and genital tracts;
solid and/or secretory tumor-based cancers, such as sarcomas,
breast cancer, cancer of the pancreas, cancer of the liver, cancer
of the kidneys, cancer of the thyroid, and cancer of the prostate;
and blood-based cancers, such as leukemias and lymphomas. Exemplary
cancers which may be detected by the present invention are lung,
bladder, gastrointestinal, cervical, breast or prostate cancer.
Exemplary infectious diseases which may be detected are cell-based
diseases in which the infectious organism is a virus, bacteria,
protozoan, parasite, or fungus. The infectious disease, for
example, may be HIV, hepatitis, influenza, meningitis,
mononucleosis, tuberculosis and sexually transmitted diseases
(STDs), such as chlamydia, trichomonas, gonorrhea, herpes and
syphilis.
[0067] As used herein, the term "generic disease state" refers to a
disease which comprises several types of specific diseases, such as
lung cancer, sexually transmitted diseases and immune-based
diseases. Specific disease states are also referred to as
histologic types of diseases. For example, the term "lung cancer"
comprises several specific diseases, among which are squamous cell
carcinoma, adenocarcinoma, large cell carcinoma, small cell lung
cancer and mesothelioma. The term "sexually transmitted diseases"
comprises several specific diseases, among which are Gonorrhea,
Human Papilloma Virus (HPV), herpes and Syphilis. The term
"immune-based diseases" comprises several specific diseases, such
as systemic lupus erythematosus (Lupus), rheumatoid arthritis and
pernicious anemia.
[0068] As used herein, the term "high-risk population" refers to a
group of individuals who are exposed to disease causing agents,
e.g., carcinogens, either at home or in the workplace (i.e., a
"high risk population" for lung cancer might be exposed to smoking,
passive smoking and occupational exposure). Individuals in a
"high-risk population" may also have a genetic predisposition.
[0069] The term "at-risk" refers to individuals who are asymptotic
but, because of a family history or significant exposure are at a
significant risk of developing a disease state (i.e., an individual
at risk for lung cancer with a >30 pack-year history of smoking;
"pack-year" is a measurement unit computed by multiplying the
number of packs smoked per day, times the number of years for this
exposure).
[0070] Cancer is a disease in which cells divide without control
due to, for example, altered gene expression. In the methods and
panels of the present invention, the cancer may be any malignant
growth in any organ. For example, the cancer may be lung, bladder,
gastrointestinal, cervical, breast or prostate cancer. Each cancer
may comprise a collection of diseases or histological types of
cancer. The term "histologic type" refers to cancers of different
histology. Depending on the cancer there can be one or several
histologic types. For example, lung cancer includes, but is not
limited to, squamous cell carcinoma, adenocarcinoma, large cell
carcinoma, small cell carcinoma and mesothelioma. Knowledge of the
histologic type of cancer affecting a patient is very useful
because it helps the medical practitioner to localize and
characterize the disease and to determine the optimal treatment
strategy.
[0071] Infectious diseases include cell-based diseases in which the
infectious organism is a virus, bacteria, protozoan, parasite or
fungus.
[0072] Exemplary detection and discrimination panels are panels
that detect lung cancer, a general disease state, and panels that
discriminate a single lung cancer type, specific disease state,
against all other types of lung cancer and false positives. False
positives can include metastatic cancer of a different type, such
as metastasized liver, kidney or pancreatic cancer.
[0073] 3. Methods of Making a Panel
[0074] The method of making a panel for detecting a generic disease
state or discriminating between specific disease states in a
patient involves determining the sensitivity and specificity of
binding of probes to a library of markers associated with a generic
or specific disease state and selecting a plurality of said probes
whose pattern of binding (localization and density/concentration)
is diagnostic of the presence or specific nature of the disease
state. In some embodiments, optional preliminary pruning and
preparation steps are performed. The method of making a panel of
the present invention involves analyzing the pattern of binding of
probes to markers in known histologic pathology samples, i.e. gold
standards. The classifier designed on the gold standard data can
then be used to design a classifier for cytometry, especially
automated cytometry. Therefore, the set of marker probes selected
from the pathology analysis is used to prepare a new training data
set taken from a cytology sample, such as sputum, fine needle
aspirations, urine, etc. Cells shed from the specified lesions will
stain in a similar fashion to the gold standards. The method
described here eliminates the experimental error in selecting the
best features set because the integrity of the diagnosis based on
gold standard histologic pathology samples is high. Although it is,
in principle, possible to use cytology samples to produce a panel,
this is less desirable because cytology samples contain debris,
there may be deterioration of the cells in a cytology sample, and
the pathology diagnosis may be difficult to confirm clinically.
[0075] A library of markers is a group of markers. The library can
comprise any number of markers. However, in some embodiments the
number of markers in the library is limited by technical and/or
commercial practicalities, such as specimen size. For example, in
some embodiments, each specimen is tested against all of the
markers in the panel. Therefore, the number of markers must not be
larger than the number of samples into which the specimen may be
divided. Another technical practicality is time. Typically, the
library contains less than 60 markers. Preferably, the library
contains less than 50 markers. More preferably, the library
contains less than 40 markers. Most preferably the library contains
10-30 markers. It is preferable that the library of potential panel
members contain more than 10 markers so that there is opportunity
to optimize the performance of the panel. As used herein, the term
"about" means plus or minus 3 markers.
[0076] In some embodiments, a library is obtained by consulting
sources which contain information about various markers and
correlations between the markers and generic/specific disease
states. Exemplary sources include experimental results, theoretical
or predicted analyses and literary sources, such as journals,
books, catalogues and web sites. These various sources may use
histology or cytology and may rely on cytogenetics, such as in situ
hybridization; proteomics, such as immunohistochemistry; cytometry,
such as MACs or DNA ploidy; and/or cytopathology, such as
morphology. The markers may be localized anywhere in or on a cell.
For example, the markers may be localized in or on the nucleus, the
cytoplasm or the cell membrane. The marker may also be localized in
an organelle within any of the aforementioned localizations.
[0077] In some embodiments, the library may be of an unsuitable
size. Therefore, one or more pruning steps may be required prior to
initiating the basic method for making a panel. The pruning step
may involve one or several successive pruning steps. One pruning
step may involve, for example, setting an arbitrary threshold for
sensitivity and/or specificity. Therefore, any marker whose
experimental or predicted sensitivity and/or specificity falls
below the threshold may be removed from the library. Other
exemplary pruning steps, which may be performed alone or in
sequence with other pruning steps, may rely on detection technology
requirements, access constraints and irreproducibility of reported
results. With respect to detection technology requirements, it is
possible that the machinery required to detect a particular marker
is unavailable. With respect to access constraints, it is possible
that licensing restrictions make it difficult or impossible to
obtain a probe that binds to a particular marker. In some
embodiments, a due diligence study is performed on each marker.
[0078] In some embodiments, prior to beginning the basic method for
making a panel, it may be necessary to perform preparation steps.
Exemplary preparation steps include optimizing the protocols for
objective quantitative detection of the markers in the library and
collecting histology specimens. Optimization of the protocols for
objective quantitative detection of the markers is within the skill
of an ordinary artisan. For example, the necessary reagents and
supplies must be obtained, such as buffers, reagents, software and
equipment. It is possible that the concentration of reagents may
need to be adjusted. For example, if non-specific binding is
observed, a person of ordinary skill in the art may dilute the
concentration of the probe solution.
[0079] In some embodiments, the histology specimens are Gold
Standards. The term "Gold Standard" is known by a person of
ordinary skill in the art to mean that the histology and clinical
diagnosis of the specimen is known. The gold standards are often
referred to as a "training" data set. The gold standards comprise a
set of measurements, or reliable estimates, of all the features
that may contribute to the discriminating process. Such features
are collected from samples collected from a representative number
of patients with known disease states. The standard samples can be
cytology samples but this is less desirable for panel
selection.
[0080] The histology samples may be obtained by any technique known
to those of skill in the art, for example biopsy. In some
embodiments, it is necessary that the size of the specimen per
patient be large enough so that enough tissue sections can be
obtained to test each marker in the library.
[0081] In some embodiments, specimens are obtained from multiple
patients diagnosed with each specific disease state. One specimen
per patient may be obtained, or multiple specimens per patient may
be obtained. In embodiments in which multiple specimens are
obtained from individual patients, the expertise of the surgeon is
relied upon to establish that each specimen obtained from a single
patient is similar to the other specimens obtained from that
patient. Specimens are also obtained from a control group of
patients. The control group of patients may be healthy patients or
patients that are not suffering from the generic or specific
disease state that is being tested.
[0082] The first step of the basic method is determining the
sensitivity and specificity of binding of probes to a library of
markers associated with the desired disease state. In this step, a
probe that is specific for each marker in the library is applied to
a sample of the patients' specimens. Therefore, in some
embodiments, if there are, for example, 30 markers in the library,
each patient's specimen will be divided into 30 samples and each
sample will be treated with a probe that is specific for one of the
30 markers. The probe contains a label that may be visualized.
Therefore, the pattern and level of binding of the probe to the
marker can be detected. The pattern and level of binding may be
detected either quantitatively, i.e., by an analytical instrument,
or qualitatively, by a human, such as a pathologist.
[0083] In some embodiments, an objective and/or quantitative
scoring method is developed to detect the pattern and level of
binding of the probe to the markers. The scoring method may be
heuristically designed. Scoring methods are used to objectify a
subjective interpretation, for example, by a pathologist. It is
within the skill of an ordinary artisan to determine a suitable
scoring method. In some embodiments, the scoring method may
comprise categorizing features, such as the density of a marker
probe stain as: none, weak, moderate, or intense. In another
embodiment, these features may be measured with algorithms
operating on microscope slide images. An exemplary scoring method
is one in which the proportions and density are consolidated into a
single "H Score" obtained by grading the intensity as: none=0,
weak=1, moderate=2, intense=3, and the percentage cells as: 0-5%=0,
6-25%=1, 26-50%=2, 51-75%=3, >75%=4, and then multiplying the
two grades together. For example, 50% weakly stained plus 50%
moderately stained would score 6=(1.times.2)+(2.times.2). The "H
score" honors the late Kenneth Hirsch, one of the present
inventors.
[0084] An ordinary artisan is capable of addressing issues related
to minimizing potential biases related to pathologists and samples.
For example, randomizing may be used to minimize the chance of
having a systematic error. Blinding may be used to eliminate
experimental biases by the people conducting the experiments. For
example, in some embodiments, pathologist-to-pathologist variation
may be minimized by conducting a double blind study. As used
herein, the term "double blind study" is a well establish method
for avoiding biases, where the data collection and data analysis
are done independently. In other embodiments, sample-to-sample
variation is minimized by randomizing the samples. For example, the
samples are randomized before the pathologist analyzes them. There
is also randomization involved in the experimental protocols. In
some embodiments, each sample is analyzed by at least two
pathologists. For each patient, a reliable assessment of the
binding of the probe to the marker is obtained. In one embodiment,
this diagnosis is made by qualified pathologists, using two
pathologists per patient, to check for reliability.
[0085] A sufficient number of samples should be collected to
produce reliable designs and reliable statistical performance
estimates. It is within the skill of a normal artisan to determine
how many samples are sufficient to produce reliable designs and
reliable statistical performance estimates. Most standard
classifier design packages have methods for determining the
reliability of the performance estimates and the sample size should
be progressively increased until reliable estimates are achieved.
For example, sufficient estimates to produce reliable designs may
be achieved with 200 samples collected and 27 different features
estimated from each sample.
[0086] The second step is selecting a limited plurality of probes.
The selecting step may employ statistical analysis and/or pattern
recognition techniques. In order to perform the selecting step, the
data may be consolidated into a database. In some embodiments, the
probes may be numbered to render their method of action as unseen
during the analysis of their effectiveness and further minimize
biases. Rigorous statistical techniques are used because of the
large amount of data that is generated by this method. Any
statistical method may be used and an ordinary skilled statistician
will be able to identify which and how many methods are
appropriate.
[0087] Any number of statistical analysis and/or pattern
recognition methods may be employed. Since the structure of the
data is initially unknown, and since different classifier design
methods perform better for different structures, it is preferred to
use at least two design methods on the data. In some embodiments,
three different methodologies may be used. One of ordinary skill in
the art of statistical analysis and/or pattern recognition of data
sets would recognize from characteristics of the data set
structures that certain statistical methods would be more likely to
yield an efficient result than others, where efficient in this case
means achieving a certain level of sensitivity and specificity with
a desired number of probes. A person of ordinary skill in the art
would know that the efficiency of the statistical analysis and/or
method is data dependent.
Exemplary statistical analysis and/or pattern recognition methods
are described below:
[0088] a) A decision tree method, known as C4.5. C4.5 is public
domain software available via ftp from
http://www.cse.unsw.edu.au/.about.quinlan/. This is well suited to
data that can be best classified by sequentially applying a
decision threshold to specific features in turn. This works best
with uncorrelated data; it also copes with data with similar means
provided the variances differ. The C4.5 package was used to provide
the examples shown herein.
[0089] b) Linear Discriminant Analysis. This involves finding
weighted combinations of the features that give the best separation
of the classes. These methods work well with correlated data, but
not in data with similar means and different variances. Several
statistical packages were used (SPSS, SAS and R), depending on the
performance estimates and graphical outputs required. Fisher's
linear discriminant function was used to obtain the classifier that
minimized the error rate. A canonical discriminant function was
used to compute receiver operating characteristic (ROC) curves
showing the trade-off between sensitivity and selectivity as the
decision threshold is changed.
[0090] c) Logistic Regression. This is a non-linear transformation
of the linear regression model: the dependent variable is replaced
by a log odds ratio (logit). Linear regression, like discriminant
analysis, belongs to a class of statistical methods founded on
linear models. Such models are based on linear relationships
between the explanatory variables.
[0091] With a sufficient number of samples it is possible, using
the above techniques and software packages, to search for
combinations of features giving good discrimination between the
classes. Other exemplary statistical analysis and/or pattern
recognition methods are the linear Discriminant Function Method in
SPSS and Logistic Regression Method in R and SAS. SPSS is the full
product name and is available from SPSS, Inc., located at SPSS,
Inc. Headquarters, 233 S. Wacker Drive, 11th floor, Chicago, Ill.
60606 (www.spss.com). SAS is the full product name and is available
from SAS Institute, Inc., 100 SAS Campus Drive, Cary, N.C.
27513-2414, USA (www.sas.com). R is the full product name and is
available as Free Software under the terms of the Free Software
Foundation's GNU (General Public License).
[0092] http://www.r-project.org/.
[0093] In some embodiments, a correlation matrix is obtained.
Correlation measures the amount of linear association between a
pair of variables. A correlation matrix is obtained by correlating
the data obtained with one marker to data obtained with another
marker. A threshold correlation number may be set, for example, 50%
correlation. In this case, all markers with a correlation number of
50% or higher would be considered correlate markers.
[0094] In some embodiments of the present invention, user supplied
weighting factors may be used to obtain optimized panels. Weighting
may be related to any factor. For example, certain markers may be
weighted higher than others due to cost, commercial considerations,
misclassifications or error rates, prevalence of a generic disease
state in a geographic location, prevalence of a specific disease
state in a geographic location, redundancy and availability of
probes. Some factors related to cost that may encourage a user to
weight certain markers higher than others is the cost of the probe
and commercial access issues, such as license terms and conditions.
Some factors related to commercial considerations that may
encourage a user to weight certain markers higher than others are
Research and Development (R&D) time, R&D cost, R&D
risk, i.e., the probability that the probe will work, cost of final
analytical instrument, final performance and the time to market. In
a detection panel, for example, some factors related to
misclassifications or error rates that may encourage a user to
weight some markers higher than others is that it may be desirable
to minimize false negatives. In a discrimination panel, on the
other hand, it may be desirable to minimize false positives. Some
factors related to prevalence of a generic or specific disease
state in a geographic area that may encourage a user to weight some
probes higher than others are that in some geographic locations the
incidence of certain generic or specific diseases are more or less
prevalent. With respect to redundancies, in some instances it is
desirable to have redundancies in the panel. For example, if for
some reason one probe fails to be detected, due to the biological
variability of the markers in the panel, a disease state will still
be detected by the other markers. In some embodiments, markers that
are preferred redundant markers may be weighted more heavily.
[0095] The invention is flexible in being adaptable to the
availability of features where cost or supply problems may not
allow the very best combination. In one embodiment, the invention
can simply be applied to the available features to find an
alternative combination.
[0096] In another embodiment, the algorithm is used to select
features that allow cost weightings to be included in the selection
process to arrive at a minimum cost solution. In the examples,
marker performance estimates for combinations selected from all the
markers collected or for only a group of commercially preferred
probes are shown. The examples also demonstrate how the C4.5
package can be used to down weight certain probes on the basis of
their high cost. These probe combinations may not perform as well
as the optimum combination, but the performance might be acceptable
in circumstances where cost is a significant factor.
[0097] Some of the methods used allow weightings to be applied to
the classes. This is available in C4.5 where the tree design can
optimize the cost. Also, the Discriminant Function method gives a
single parameter output which can be used to give a desired false
positive or false negative probability. A plot of these parameters
for different threshold settings is known as the receiver operating
characteristic (ROC) curve. An ROC curve shows the estimated
percentage of false positive against true positive scores for
different threshold levels of a classifier.
[0098] Given the heterogeneous nature of many generic disease
states, the panels may be constructed with a degree of redundancy
to ensure that the tests have sufficient sensitivity, specificity,
positive predictive value (Positive Predictive Value=True
Positives/(True Positives+False Positives) and negative predictive
value (Negative Predictive Value=True negatives/(False
Negatives+True Negatives) to justify their use as a
population-based screen. However, local and regional differences
may dictate specific use of the tests in different segments of the
global market, and so may significantly influence the criteria used
to construct the final panel test for a given market. While the
optimization of clinical utility is of utmost importance, local
factors including affordability (cost), technical competence,
laboratory and healthcare provider resources, workflow issues,
manpower requirements, and availability of the probes and labels
will contribute to a final, local selection of the markers used in
the panel. Well known linear discriminant function analysis is used
to include and assess all potential selection factors, by which
each local factor is represented by a term in the equation, and
each is weighted according to its locally determined significance.
In this way, a panel test optimized for use in one world region may
differ from a panel test optimized for use in a different
region.
[0099] Once detection or discrmination panels have been designed
using the above described method, the next step is to validate the
panel using known cytology samples. Prior to validation, optional
optimization steps may be performed. In some embodiments, the
method for collecting cytology samples may be improved. This
encompasses methods of obtaining the sample from the patient as
well as methods for mixing the cytology sample. In other
embodiments, the cytology presentation methods may be improved. For
example, identifying optimal fixatives (preservation fluids) or
transportation fluids.
[0100] The cytology samples used to validate the panels produced
using the gold standard histology samples are cytology samples with
known diagnoses. These samples may be collected using any method
known by those of skill in the art. For example, sputum samples can
be collected by spontaneous production, induced production and
through the use of agents that enhance sputum production. The
sample is contacted with each probe in the panel and the level and
pattern of binding of the probes is analyzed to determine the
performance of the panel. In some embodiments, it may be necessary
to further optimize the panel. For example, it may be necessary to
remove a probe from the panel. Or, it may be necessary to add an
additional probe to the panel. Additionally, it may be necessary to
replace one probe on the panel with another probe. If a new probe
is added, this probe may be a correlate marker as determined from a
correlation matrix. Alternatively, the probe may be a functionally
similar marker. Once the panel is optimized, the panel may proceed
for further testing in clinical studies.
[0101] In other embodiments, it is not necessary to optimize the
panel. If the results with the cytology samples correlate with the
results from the histology samples, there may not be a need to
optimize the panel and the panel may proceed for further testing in
clinical studies.
[0102] 4. Methods of Use
[0103] Once a panel is obtained using the above described method,
it may be applied to cytologic samples. To illustrate the method,
cancer, especially lung cancer, will be exemplified. Similar steps
and procedures will be appliced for other disease states. It is to
be expected that cells shed from the specified lesions will stain
in a similar fashion and show in a cytologic sample, such as a fine
need aspiration, sputum, urine, in a similar fashion as in the
histologic pathology samples used to obtain the panel.
[0104] The basic method of the present invention typically involves
two steps. First, a cytological sample suspected of containing
diseased cells is contacted with a panel containing a plurality of
agents, each of which quantitatively binds to a disease marker.
Then, the level or pattern of binding of each agent to a disease
marker is detected. The results of the detection may be used to
diagnose the presence of a generic disease or to discriminate among
specific disease states. An optional preliminary step is
identifying an optimized panel of agents that will aid in the
detection of a disease or the discrimination between disease states
in a cytologic sample.
[0105] Cytology specimens may include, but are not limited to,
cellular samples collected from body fluids, such as blood, urine,
spinal fluids, and lymphatic systems; epithelial cell-based organ
systems, such as the pulmonary tract, e.g., lung sputum, urinary
tract, e.g., bladder washings, genital tract, e.g., cervical PAP
smears, and gastrointestinal tract, e.g., colonic washings; and
fine needle aspirations from solid tissue sites in organs and
systems such as the breast, pancreas, liver, kidneys, thyroid, bone
marrow, muscles, prostate, and lungs; biopsies from solid tissue
sites in organs and systems such as the breast, pancreas, liver,
kidneys, thyroid, bone marrow, muscles, prostate, and lungs; and
histology specimens, such as tissue from surgical biopsies.
[0106] An illustrative panel of agents according to the present
invention includes any number of agents that allows for accurate
detection of malignant cells in a cytological sample. Molecular
markers envisioned by the present invention may be any molecule
that aids in the detection of malignant cells. Markers may be
selected for inclusion in a panel based on several different
criteria relating to changes in level or pattern of expression of
the marker. Preferred are molecular markers where changes in
expression: occur early in tumor progression, are exhibited by a
majority of tumor cells, allow for detection of in excess of 75% of
a given tumor type, most preferably in excess of 90% of a given
tumor type and/or allow for the discrimination between histologic
types of cancer.
[0107] The first step of the basic method is the detection of
changes in the level or pattern of expression of the panel of
agents in a cytological sample. This step typically involves
contacting the cytologic sample with an agent, such as a labeled
polyclonal or monoclonal antibody or fragment thereof or a nucleic
acid probe, and observing the signal in individual cells. Detection
of cells where there is a change in signal is indicative of a
change in the level of expression of the molecular marker to which
the label probe is directed. The changes are based on an increase
or decrease in the level of expression relative to nonmalignant
cells obtained from the tissue or site being examined.
[0108] An analysis of the changes in the level or pattern of
expression of a panel of agents enables a skilled artisan to
determine, with high sensitivity and high specificity, whether
malignant cells are present in the cytologic sample. The term
"sensitivity" refers to the conditional probability that a person
having a disease will be correctly identified by a clinical test,
(the number of true positive results divided by the number of true
positive and false negative results). Therefore, if a cancer
detection method has high sensitivity, the percentage of cancers
detected is high e.g., 80%, preferably greater than 90%. The term
"specificity" refers to the conditional probability that a person
not having a disease will be correctly identified by a clinical
test, (i.e., the number of true negative results divided by the
number of true negative and false positive results). Therefore, if
a cancer detection method has high specificity, 80%, preferably
90%, more preferably 95%, the percentage of false positives the
method produces is low. A "cytologic sample" encompasses any sample
collected from a patient that contains that patient's cells.
Examples of cytological samples envisioned by the present invention
include body fluids, epithelial cell-based organ system washings,
scrapings, brushings, smears or effusions, and fine-needle
aspirates and biopsies.
[0109] Use of the markers described in this invention assumes that
it is possible to obtain an adequate cytologic sample routinely and
that the samples can be adequately preserved for subsequent
evaluation. The cytologic sample may be processed and stored in a
suitable preservative. Preferably, the cytologic sample is
collected in a vial containing the preservative. The preservative
is any molecule or combination of molecules known to maintain
cellular morphology and inhibit or block degradation of cellular
proteins and nucleic acids. To ensure proper fixation, the sample
may be mixed at the collection site at high speeds to disaggregate
the sample and/or break up obscuring material such as mucus,
thereby exposing the cells to the preservative.
[0110] Once a specimen is obtained, it is desirable to homogenize
it, using an appropriate mixing device. This permits using aliquots
for multiple purposes, including the possibility of sending
aliquots to more than one testing site, as well as preparing
multiple slides and/or multiple depositions on a slide. The initial
homogenization of the specimen and of each aliquot before use will
ensure that each individual slide will have substantially the same
distribution of cells, so that comparisons of results from one
slide to another will be meaningful.
[0111] Preparation of a specimen for analysis involves applying a
sample to a microscope slide using methods including, but not
limited to, smears, centrifugation, or deposition of a monolayer of
cells. Such methods may be manual, semi-automated, or fully
automated. The cell suspension may be aspirated depositing the
cells on a filter and a monolayer of cells transferred to a
prepared slide that may be processed for further evaluation. By
repeating this process additional slides may be prepared as
necessary. The present invention encompasses detection of one
molecular marker per slide. Detection of several molecular markers
per slide is also envisioned. Preferably, 1-6 markers are detected
per slide. In some embodiments 2 markers are detected per slide. In
other embodiments, 3 markers are detected per slide.
[0112] The present invention contemplates detecting changes in
molecular marker expression at the DNA, RNA or protein level using
any of a number of methods available to an ordinary skilled
artisan. Detection of the changes in the level or pattern of
expression of the molecular markers in a cytologic sample generally
involves contacting a cytologic sample with a polyclonal or
monoclonal antibody or fragment thereof or a nucleic acid sequence
that is complementary to the nucleic acid sequence encoding a
molecular marker in the panel, collectively "probes", and a label.
Typically, the probe and label components are operatively linked so
that when the probe reacts with the molecular marker a signal is
emitted (a "labeled probe"). Labels envisioned by the present
invention are any labels that emit or enable a signal and allow for
identification of a component in a sample. Preferred labels include
radioactive, fluorogenic, chromogenic or enzymatic moieties.
Therefore, possible methods of detection include, but are not
limited to, immunocytochemistry; proteomics, such as
immunochemistry; cytogenetics, such as in situ hybridization, and
fluorescence in situ hybridization; radiodetection, cytometry and
field effects, such as MACs and DNA ploidy (the quantitation of
stoichiometrically-stained nuclear DNA using automated computerized
cytometry) and; cytopathology, such as quantitative cytopathology
based on morphology. The signal generated by the labeled probe is
preferrably of sufficient intensity to permit detection by a
medical practitioner or technician.
[0113] Once the slide is prepared, a medical practitioner conducts
a microscopic review of the slides in order to identify cells that
exhibit a change in marker expression characteristic of a diagnosis
of cancer. The medical practitioner may use an image analysis
system and automated microscope to identify cells of interest.
Analysis of the data may make use of an information management
system and algorithms that will assist the physician in making a
definitive diagnosis and select the optimal therapeutic approach. A
medical practitioner may also examine the sample using an
instrument platform that is capable of detecting the presence of
the labeled agent.
[0114] A molecular diagnostic panel assay will result in one or
more glass microscope slides with labeled cells and/or tissue
sections. The challenge for human experts to assess these
(cyto)pathology multilabeled-cell preparations objectively and with
clinically meaningful results is a virtually insurmountable
detection and perception problem for any human being.
[0115] Computer-aided imaging systems (i.e., Photonic
Microscopes.TM.) can be developed and used to assess quantitatively
and reproducibly the amount and location of probe-labeled cells and
tissues. Such Photonic Microscopes.TM. combine robotic
slide-handling capabilities, data management systems (e.g., medical
informatics), and quantitative digital (optical and electronic)
image analysis hardware and software modules to detect and report
cell-based probe content and localization data that cannot be
obtained by human visualization with comparable sensitivity and
accuracy. These probe data can be used to characterize and
differentiate cellular samples based upon their related
characteristics and differences in their respective cell-based
markers for a variety of disease states.
[0116] The present methodology is a methodology whereby the
molecular diagnostic panels are applied to cell-based specimens and
samples, and whereby computer-aided imaging systems are
subsequently used to quantify and report the results of the
molecular diagnostic panel tests. Such imaging systems can be used
to evaluate cell-based samples in which multiple probes are used
simultaneously on a given slide-based sample, and in which the
probes can be separately analyzed, quantified, and reported because
the probes are differentiated by color on the microscope cytology
or histology slide.
[0117] The signals generated by a labeled agent in the sample may,
if they are of appropriate type and of sufficient intensity, be
detected by a human reviewer (e.g., pathologist) using a standard
microscope or a Computer-Aided Microscope [167]. The Computer-Aided
Microscope is an ergonomic, computer-interfaced microscope
workstation that integrates mouse-driven control of microscope
operation (e.g., stage movement, focusing) with computerized
automation of key functions (e.g., slide scanning patterns). A
centralized Data Management System stores, organizes and displays
relevant patient information as well as results from all specimen
screenings and pathologist reviews. An identification number that
is imprinted onto barcodes and affixed to each sample slide
uniquely identifies each sample in the database, and relates it to
the original specimen and the patient.
[0118] In a preferred embodiment the signals generated by a labeled
agent in the sample will be detected and quantitated using an
automated image analysis system, or Photonic Microscope, interfaced
to the centralized Data Management System. The Photonic Microscope
provides fully automated software control of the microscope
operations and incorporates detectors and other components
appropriate for quantitation even of signals not detectable by
human reviewers, such as very faint signals or signals from
radiolabeled moieties. The location of detected signals is stored
electronically for rapid relocation by automated instruments, and
for human review using a Computer-Aided Microscope [168].
[0119] The centralized Data Management System archives all patient
and sample data using the bar-coded identification number. The data
may be acquired asynchronously, from a multiplicity of sites, and
may be derived from multiple reviews and analyses by human
cytologists and/or automated analyzers. These data may include
results from multiple sample slides representing aliquots from a
single previously homogenized patient specimen. Part or all of the
data may be transferred to or from a hospital's Laboratory
Information System to meet reporting, archiving, billing or
regulatory requirements. A single, comprehensive report with
integrated results from panel tests and human reviews may be
generated and delivered to the physician in hardcopy, or
electronically through networked computers or the Internet.
[0120] In some embodiments, the instant method allows for
differential discrimination of different diseases, such as
different histologic types of cancers. The term "histologic type"
refers to specific disease states. Depending on the general disease
state there can be one or several histologic types. For example,
lung cancer includes, but is not limited to, squamous cell
carcinoma, adenocarcinoma, large cell carcinoma, small cell
carcinoma and mesothelioma. Knowledge of the histologic type of
cancer affecting a patient is very useful because it helps the
medical practitioner to localize and characterize the disease and
to determine the optical treatment strategy.
[0121] In order to determine the specific disease state, a panel of
markers is selected that allows for discrimination between specific
disease states. For example, within a panel of molecular markers, a
pattern of expression may be identified that is indicative of a
particular histologic type of cancer. The detection of the level of
expression of the panel of molecular markers is achieved by the
above-described methods. Preferably, a panel of 1-20 molecular
markers is employed to discriminate among the various histologic
types of lung cancer. However, most preferably, 4-7 markers are
used. Decision trees may be developed to aid in discriminating
between different histologic types based on patterns of marker
expression.
[0122] In addition to allowing for the detection of malignant cells
in a cytologic sample, the instant invention has utility in the
molecular characterization of the disease state. Such information
is often of prognostic significance and can assist the physician in
the selection of the optimal therapeutic approach for a particular
patient. In addition, the panel of markers described in this
invention may have utility in monitoring the patient for either
recurrence or to measure the efficacy of the therapy being used to
treat the disease.
[0123] By way of non-limiting example, the presence of lung cancer
may be detected by a lung cancer detection panel and the specific
type of lung cancer may be detected by a discrimination panel. If
the medical practitioner determines that malignant cells are
present in the cytologic sample, a further analysis of the
histologic type of lung cancer may be performed. The histologic
type of lung cancer encompassed by the present invention includes
but is not limited to squamous cell carcinoma, adenocarcinoma,
large cell carcinoma, small cell carcinoma and mesothelioma. FIG. 1
illustrates molecular markers that are preferable markers to be
included in a panel for identifying different histologic types of
lung cancer. The column labeled "%" indicates the percentage of
tumor specimens that express a particular marker.
[0124] In determining the various histologic types of lung cancer,
the relative level of expression of a marker is analyzed. FIG. 2
illustrates how different markers may be used to discriminate among
different histologic types of cancer. In this table, SQ indicates
squamous cell carcinoma, AD indicates adenocarcinoma, LC indicates
large cell carcinoma, SC indicates small cell carcinoma and ME
indicates mesothelioma. The numbers appearing in each cell
represent frequency of marker change in one cell type versus
another. To be included in the table, the ratio must be greater
than 2.0 or less than 0.5. A number larger than 100 generally
indicates that the second marker is not expressed. In such cases
the denominator was set at 0.1 for the purpose of the analysis.
Finally, empty cells represent either no difference in expression
or the absence of expression data.
[0125] One method for analyzing the data collected is to construct
decision trees. Schemes 1-4 are examples of decision trees that may
be constructed to enable a differential determination of a
histologic type of lung cancer using the patterns of expression.
The present invention is in no way limited to the decision trees
presented in Schemes 1-4. The relative level of expression of a
marker can be higher, lower, or the same (ND) as the level of
expression of the molecular marker in a malignant cell of a
different histologic type. Each scheme enables a distinction
between five histologic types of lung cancer through the use of the
indicated panel of molecular markers.
[0126] For example, in Scheme 1 the panel consists of HERA, MAGE-3,
Thrombomodulin and Cyclin D1. First the sample is contacted with a
labeled probe directed toward HERA. If the expression of HERA is
lower than the control, the test indicates that the histologic type
of lung cancer is mesothelioma (ME). If, however, the expression is
higher or the same as the control, the sample is contacted with a
probe directed toward MAGE-3. If the expression of MAGE-3 is lower
than the control, the sample is contacted with a labeled probe
directed toward Cyclin D1 and a determination of small cell
carcinoma (SC) or adenocarcinoma (AD) is possible. If the
expression of MAGE-3 is higher than or the same as the control, the
sample is contacted with a labeled probe directed toward
Thrombomodulin and a determination of squamous cell carcinoma (SC)
or large cell carcinoma (LC) is possible.
##STR00001##
[0127] In Scheme 2 the panel consists of E-Cadherin, Pulmonary
Surfactant B and Thrombomodulin. First the sample is contacted with
a labeled probe directed toward E-Cadherin. If the expression of
E-Cadherin is lower than the control, the test indicates that the
histologic type of lung cancer is mesothelioma (ME). If, however,
the expression is higher or the same as the control, the sample is
contacted with a probe directed toward Pulmonary Surfactant B. If
the expression of Pulmonary Surfactant B is lower than the control,
the sample is contacted with a labeled probe directed toward
Thrombomodulin and a determination of squamous cell carcinoma (SQ)
or large cell carcinoma (LC) is possible. If the expression of
Pulmonary Surfactant B is higher than or the same as the control,
the sample is contacted with a labeled probe directed toward CD44v6
and a determination of adenocarcinoma (AD) and small cell carcinoma
(SC) is possible. (See Schemes 3 and 4 for more examples of
decision trees).
##STR00002##
##STR00003##
##STR00004##
[0128] A preferred method involves using panels of molecular
markers where differences in the pattern of expression permits the
discrimination between the various histologic type of lung
cancer.
[0129] Many different decision trees may be constructed to analyze
the patterns of marker expression. This information may be used by
physicians or other healthcare providers to make patient management
decisions and select an optimal treatment strategy.
[0130] 5. Reporting of Results of Panel Analysis
[0131] The results from the panel analysis may be reported in
several ways. For example, the results may be reported as a simple
"yes or no" result. Alternatively, the result may be reported as a
probability that the test results are correct. For example, the
results from a detection panel study may indicate whether a patient
has a generic disease state or not. As the panel also reports the
specificity and sensitivity, the results may also be reported as
the probability that the patient has a generic disease state. The
results from a discrimination panel analysis will discriminate
among specific disease states. The results may be reported as a
"yes or no" with respect to whether the specific disease state is
present. Alternatively, the results may be reported as a
probability that a specific disease state is present. It is also
possible to perform several discrimination panel analyses on a
specimen from one patient and report a profile of the probabilities
that the disease state present is a specific disease state with
respect to the other possibilities. The other possibilities may
also include false positives.
[0132] In embodiments in which a profile of the probabilities of
each specific disease state being present is produced, there are
several possible outcomes. For example, it is possible that all of
the probabilities will be a very small probability. In this
instance, it is possible that the doctor will conclude that the
patient's specimen diagnosis is a false positive. It is also
possible that all of the probabilities will be low except for one
that is above 80-90%. In this instance, it is possible that the
doctor will conclude that the test verifies that the patient has
the specific disease state that indicated the high probability. It
is also possible that most of the probabilities will be low, but
similarly high probabilities are reported for two specific disease
states. In this case, a doctor may recommend more extensive panel
testing to ensure that the correct disease state is identified.
Another possibility is that all of the probabilities reported will
be low, with one being slightly higher than the rest but not high
enough to be in the 80-90% range. In this case, a doctor may
recommend more extensive panel testing to ensure that the correct
disease state is identified and/or to rule out metastatic cancer
from a remote primary tumor of a different cancer type.
[0133] The following Example is illustrative of the method of the
invention for selecting a disease detection panel, disease
discrimination panels, validation of the panels and use of the
panels in the clinic to screen for a disease and to discriminate
among different subtypes of the disease. Lung cancer was selected
for this illustrative example, in part because of its importance to
world health, but it will be appreciated that similar procedures
will apply to other types of cancer, as well as to infectious,
degenerative and autoimmune diseases, according to the foregoing
general disclosure.
ILLUSTRATIVE EXAMPLE
[0134] The present method was used to develop lung cancer detection
panels as well as single lung cancer type specific discrimination
panels. Lung cancer is an extremely complex collection of diseases
that can be segregated into two main classes. Non-small cell lung
carcinoma (NSCLC) that accounts for approximately 70 to 80% of all
lung cancers can be further subdivided into three main histologic
types including squamous cell carcinoma, adenocarcinoma, and large
cell carcinoma. The remaining 20 to 30% of lung cancer patients
present with small cell lung carcinoma (SCLC). In addition,
malignant mesothelioma of the pleural space, can develop in
individuals exposed to asbestos and will often spread widely
invading other thoracic structures. Different forms of lung cancer
tend to localize in different regions of the lung, have different
prognoses, and respond differently to various forms of therapy.
[0135] According to the latest statistics from the World Health
Organization (Globocan 2000), lung cancer has become the most
common fatal malignancy in both men and women with an estimated
1.24 million new cases and 1.1 million deaths each year. In the
U.S. alone, the National Cancer Institute reports that there are
approximately 186,000 new cases of lung cancer and each year
162,000 people die of the disease, accounting for 25% of all
cancer-related deaths. In the U.S., overall 1-year survival for
patients with lung cancer is 40%, however, only 14% live 5 years.
In other parts of the world, 5-year survival is significantly lower
(5% in the UK). The high mortality of lung cancer can be attributed
to the fact that most patients (85%) are diagnosed with advanced
disease when treatment options are limited and the disease is
likely to have metastasized. In these patients, 5-year survival is
between 2-30% depending of the stage at the time of diagnosis. This
is in sharp contrast to cases where patients are diagnosed early
and 5-year survival is greater than 75%. While it is true that a
number of new chemotherapeutic agents have been introduced into
clinical practice for the treatment of advanced lung cancer, to
date, none have yielded a significant improvement in long-term
survival. Even though patients with early stage disease can
presumably be cured by surgery, they remain at significant risk, as
there is a high probability that they will develop a second
malignancy. Thus, for the lung cancer patient, early detection and
treatment followed by aggressive monitoring provides the best
chance of achieving significant improvements in long-term survival
along with a reduction in morbidity and cost.
[0136] At the present time, a patient is suspected of having lung
cancer either because of a suspicious lesion on X-ray or because
the patient becomes symptomatic. As a result, most patients are
diagnosed with relatively late stage disease. In addition, because
most methods lack sufficient sensitivity with respect to the
detection of early stage disease, the current policy of the U.S.
National Cancer Institute (NCI), National Institutes of Health,
recommends against screening for lung cancer even in populations of
patients who are at significant risk. In this embodiment of the
present invention, however, sputum cytology is employed to provide
a relatively noninvasive, more effective and cost-effective means
for the early detection of lung cancer.
[0137] The specificity of sputum cytology is relatively high.
Recent studies have indicated that experienced cytotechnologists
are able to recognize malignant or severely dysplastic cells with a
high degree of accuracy and reliability [10]. While the detection
rate can be as high as 80 to 90% when samples are collected from
patients with a relatively advanced disease [11,12], overall,
sputum cytology has a sensitivity of only 30-40% [13,14]. The low
sensitivity of sputum cytology is particularly important given that
obtaining and preparing the specimen can be relatively expensive.
Furthermore, failing to detect a malignancy can significantly delay
treatment thereby reducing the chance of achieving a cure.
[0138] The selection of an "at-risk" population can also influence
the value of sputum cytology as a screening tool. Individuals who
are at significant risk include those with a prior diagnosis of
lung cancer, long-term smokers or former smokers (>30 pack
years) and individuals with long-term exposure to asbestos or
pulmonary carcinogens. People with a genetic predisposition or
familial history are also included in an "at-risk" population. Such
individuals are likely to benefit from testing. While the inclusion
of individuals with lower risk may result in an increase in the
absolute number of cases detected, it would be hard to justify the
substantial increase in healthcare costs.
[0139] Other factors that contribute to the relatively poor
performance of conventional sputum cytology include the location of
the lesion, tumor size, histologic type, and the quality of the
sample. Squamous-cell carcinoma accounts for 31% of all primary
pulmonary neoplasms. Most of these tumors arise from segmental
bronchi and extend to the proximal lobar and distal subsegmental
branches [15]. For this reason, sputum cytology is reasonably
effective (79%) in detecting these lesions. Currently, squamous
cell carcinoma is viewed as the only type of lung cancer that is
amenable to cytologic detection in an in situ and radiologically
occult stage [15], as sloughed cells are more likely to be
available for evaluation. In one large study where patients were
followed with both chest X-ray and sputum cytology, 23% of all lung
cancers were detected by cytology alone, suggesting that the tumors
were early stage and radiologically occult [16]. In another study
[17], sputum cytology detected 76% of patients with radiologically
occult tumors.
[0140] In the case of adenocarcinoma, 70% of tumors occur in the
periphery of the lung making it less likely that malignant cells
will be found in a conventional sputum specimen. For this reason,
adenocarcinomas are rarely detected by sputum cytology (45%)
[12,18,19], an important consideration, since the incidence of
adenocarcinoma appears to be increasing, particularly in women
[20-22].
[0141] Tumor size can also affect the likelihood of achieving a
correct diagnosis, a factor that is particularly important when
considering a screening test for the detection of disease in
asymptomatic individuals. While there is only a 50% chance that
tumors <24 mm will be read as a true positive, the probability
of detecting a larger lesion is in excess of 84% [12].
[0142] Recent reports also indicate that the cellularity of the
specimen will affect the sensitivity of sputum cytology [14,23]. In
general, patients with squamous cell carcinoma produce specimens
with significant numbers of tumor cells, thereby increasing the
likelihood of a correct diagnosis [14,23]. For patients with
adenocarcinoma, the presence of tumor cells in a sputum specimen is
reported to be less than 10% in 95% of the specimens and less than
2% in 75% of specimens, making the diagnosis significantly more
difficult.
[0143] The degree of differentiation can also influence the ability
of a pathologist to detect malignant cells, particularly in cases
of adenocarcinoma. Well-differentiated tumor cells frequently
resemble normeoplastic respiratory epithelial cells. In the case of
small-cell lung carcinoma, sputum samples often contain nests of
loosely aggregated cells that have a distinct appearance. However,
techniques currently used to process sputum samples tend to
disaggregate the cells, making a diagnosis more difficult.
[0144] Sample quality is another factor that can contribute to the
low sensitivity of sputum cytology. Recent reports suggest that it
is possible to obtain adequate samples from 70-85% of subjects.
However, achieving this measure of success often requires that
patients provide multiple specimens [13]. This procedure is
inconvenient, time-consuming and costly. Patient compliance is also
generally low, as patients are frequently asked to collect over
several days [13]. Of equal importance is the observation that
former smokers, while at significant risk for developing lung
cancer, often fail to produce an adequate specimen. Sample
preservation and processing is another critical factor that can
affect the value of sputum cytology as a diagnostic test.
[0145] Lastly, even if adequate samples could be obtained and
optimally prepared, cytotechnologists generally still have to
review 2-4 slides per specimen, each typically taking up to four
minutes [24]. Given the low sensitivity, high technical complexity
and labor intensity of conventional sputum cytology, it is not
surprising that this test has been almost universally rejected as a
population-based screen for the early detection of lung cancer
[25].
[0146] Even if these technical issues were resolved, the low
sensitivity of sputum cytology remains a significant problem. The
high incidence of false negative results can significantly delay
the patient receiving potentially curative therapy. While it may be
possible to develop tests with greater sensitivity, such
improvements must not come at the cost of specificity. An increase
in the number of false positive results would subject patients to
unnecessary, often invasive and costly, follow-up and would have a
negative impact on the patient's quality of life. The present
invention overcomes many of the limitations associated with
previous methods of early cancer detection, including those related
to the use of sputum cytology for the early detection of lung
cancer.
[0147] Lung cancer is a heterogeneous collection of diseases. To
ensure that a test has the necessary level of sensitivity and
specificity to justify its use as a population based screen, the
present invention envisions using, for example, a library of 10 to
30 cellular markers to develop panels. Selection of the library of
this invention was based on a review and reanalysis of the relevant
scientific literature where, in most cases, marker expression was
measured in biopsy specimens taken from patients with lung cancer
in an attempt to link expression with prognosis.
[0148] For example, a preferred panel for early detection,
characterization, and/or monitoring of lung cancer in a patient's
sputum may include molecular markers for which a change in
expression occurred in at least 75% of tumor specimens. An
exemplary panel includes markers selected from VEGF,
Thrombomodulin, CD44v6, SP-A, Rb, E-Cadherin, cyclin A, nm23,
telomerase, Ki-67, cyclin D1, PCNA, MAGE-1, Mucin, SP-B, HERA,
FGF-2, C-MET, thyroid transcription factor, Bcl-2, N-Cadherin,
EGFR, Glut-1, ER-related (p29), MAGE-3 and Glut-3. A most preferred
panel includes molecular markers for which a change in expression
occurs in more than 85% of tumor specimens. An exemplary panel
includes molecular markers selected from Glut-1, HERA, Muc-1,
Telomerase, VEGF, HGF, FGF, E-cadherin, Cyclin A, EGF Receptor,
Bcl-2, Cyclin D1 and N-cadherin. With the exception of Rb and
E-cadherin, a diagnosis of lung cancer is associated with an
increase in marker expression. A brief description of the library
of probes/markers utilized in the present example is provided below
in Table 4. It is noted that the numbering of the antibodies in the
table below is consistent with the number of the
antibodies/probes/markers throughout this example.
TABLE-US-00004 TABLE 4 Probes and Markers for Lung Panel No. Marker
Abbreviation Full Name of Antibody Probe Target Marker
Name/Description 1 VEGF anti-VEGF Vascular Endothelial Growth
Factor protein 2 Thrombomodulin anti-Thrombomodulin trams-membrane
glycoprotein 3 CD44v6 anti-CD44v6 cell surface glycoprotein (CD44
variant 6 gene): cell adhesion molecule 4 SP-A anti-Surfactant
Apoprotein A pulmonary surfactant apoprotein 5 Retinoblastoma
anti-Retinoblastoma gene product phosphoprotein 6 E-Cadherin
anti-E-Cadherin transmembrane Ca** dependent cell adhesion molecule
7 Cyclin A anti-Cyclin A protein subunit of cyclin-dependent kinase
enzymes: for cell cycle (regulatio 8 nm23 anti-nm23 2 closely
related proteins produced by nm23-H1 and -H2 genes 9 Telomerase
anti-Telomerase ribonucleoprotein enzyme for chromosome repair 10
Mib-1 (Ki-67) anti-Ki-67 nuclear protein: expressed in
proliferating cells 11 Cyclin D1 anti-Cyclin D1 protein subunit of
cyclin-dependent kinase enzymes: for cell cycle regulatio 12 PCNA
anti-Proliferating Cell Nuclear Antigen protein cofactor for DNA
polymerase delta 13 MAGE-1 anti-Melanoma-Associated Antigen 1 cell
recognition protein coded by MAGE family of genes 14 Mucin 1
(MUC-1) anti-Mucin 1 cell surface and secreted mucin (highly
glycosylated protein) 15 SP-B anti-mature Surfactant Apoprotein B
pulmonary surfactant apoprotein 16 HERA anti-Human Epithelial
Related cell surface antigen (transmembrane protein) Antigen
(MOC-31) 17 FGF-2 (basic FGF) anti-Fibroblast Growth Factor protein
that binds to cell surface 18 c-MET anti-c-MET trans-membrane
receptor protein for Hepatocyte Growth Factor (HGF) 19 Thyroid
Transcription anti-TTF-1 regulator of thyroid-specific genes: also
expressed in lung Factor 1 20 BCL-2 anti-BCL2 intracellular
membrane-bound protein encoded by BCL2 gene 21 P120 anti-p120
Proliferation-Associated Nucleolar Antigen protein 22 N-Cadherin
anti-N-Cadherin transmembrane Ca** dependent cell adhesion molecule
23 EGFR anti-EGFR Epidermal Growth Factor Receptor, transmembrane
glycoprotein 24 Glut 1 anti-Glut 1 Glucose-transporting,
transmembrane Glut family of proteins 25 ER-related (p29)
anti-ER-related P29: anti-HSP 27 Estrogen Receptor-related p29
protein: Heat Shock protein 27 26 Mage 3 anti-Melanoma-Associated
Antigen 3 cell recognition protein coded by MAGE family of genes 27
Glut 3 anti-Glut 3 Glucose-transporting, transmembrane Glut family
of proteins 28 PCNA (higher dilution) anti-Proliferating Cell
Nuclear Antigen protein cofactor for DNA polymerase delta
[0149] Each molecular marker in the preferred panel is described
below. Table 5, reciting the percentage of expression of the
markers in tissue for each type of lung cancer is provided at the
end of this section.
[0150] Glucose Transporter Proteins (Glut 1 and Glut 3) [26-28]
[0151] Glucose Transporter-1 (Glut 1) and Glucose Transporter-3
(Glut-3) are a ubiquitously expressed high affinity glucose
transporter. Tumor cells often display higher rates of respiration,
glucose uptake, and glucose metabolism than do normal cells, and
the elevated uptake of glucose in tumor cells is thought to be
mediated by glucose transporters. Overexpression of certain types
of GLUT isoforms has been reported in lung cancer. The cellular
localization of Glut 1 is in the cell membrane. GLUT-1 and GLUT-3
are disease markers useful for detection of a disease state.
[0152] Malignant cells exhibit an increase in glucose uptake that
appears to be mediated by a family of glucose transporter proteins
(Gluts). Oncogenes and growth factors appear to regulate the
expression of these proteins as well as their activities. Members
of the Glut family of proteins exhibit different patterns of
distribution in various human tissues and rapid proliferation is
often associated with their overexpression. Recent evidence
suggests that Glut1 is expressed by a large percentage of NSCLC and
by a majority of SCLC.
[0153] While the expression of Glut 3 is relatively low in both
NSCLC and SCLC a significant percentage (39.5%) of large cell
carcinomas express the protein. In stage I tumors, 83% express
Glut1 at some level with 75-100% of cells staining in 25% of cases.
These data would suggest that Glut1 overexpression is a relatively
early event in tumor progression. Glut1 immunoreactivity has also
been detected in >90% of stage II and IIIA cancers. There also
appears to be an inverse correlation between Glut1 and Glut3
immunoreactivity and tumor differentiation. Tumors expressing high
levels of Glut1 appear to be particularly aggressive that are
associate with a poor prognosis. In cases were tumors were negative
for the proteins better survival was observed.
[0154] Human Epithelial Related Antigen (HERA) [29,30]
[0155] HERA is a transmembrane glycoprotein with an, as yet,
unknown function. HERA is present on most normal and malignant
epithelia. Recent reports suggest that the while HERA expression is
high in all histologic types of NSCLC making it useful as a
detection marker. In contrast HERA expression is absent in
mesothelioma and thus suggesting would have utility as a
discrimination marker. The cellular localization of HERA is the
cell surface.
[0156] Basic Fibroblast Growth Factor (FGF) [31-34]
[0157] Basic Fibroblast Growth Factor (FGF) is a polypeptide growth
factor with a high affinity for heparin and other
glycosaminoglycans. In cancer, FGF functions as a potent mitogen,
plays a role in angiogenesis, differentiation, and proliferation,
and is involved in tumor progression and metastasis. FGF
overexpression frequently occurs in both SCLC and squamous cell
carcinoma. In many cases (62%), the cells also express the FGF
receptor suggesting the presence of an autocrine loop. Forty-eight
percent of Stage 1 tumors overexpress FGF. The frequency of FGF in
Stage II lung cancer is 84%. Expression of either the growth factor
or its receptor was associated with the poor prognosis. Five-year
survival rates for those patients with stage I disease were 73% for
those expressing FGF versus 80% for those who were FGF negative.
The cellular localization is the cell membrane.
[0158] Telomerase [35-42]
[0159] Telomerase is a ribonucleoprotein enzyme that extends and
maintains telomeres of eukaryotic chromosomes. It consists of a
catalytic protein subunit with reverse transcriptase activity and
an RNA subunit with reverse transcriptase activity and an RNA
subunit that serves as the template for telomere extension. Cells
that do not express telomerase have successively-shortened
telomeres with each cell division, which ultimately leads to
chromosomal instability, aging and cell death. The cellular
localization of telomerase is nuclear.
[0160] Expression of telomerase appears to occur in immortalized
cells and enzyme activity is a common feature of the malignant
phenotype. Approximately 80-94% of lung tumors exhibit high levels
of telomerase activity. In addition, 71% of hyperplasia, 80% of
metaplasia, and 82% of dysplasia express enzyme activity. All the
carcinoma in situ (CIS) specimens exhibit enzyme activity. The low
levels of expression in premaligant tissues is probably related to
the fact that only a small percentage of cells (5 and 20%) in the
sample express enzyme activity. This is in contrast to tumors where
20-60% of cells may express enzyme activity. Based on a limited
number of samples it would appear expression of telomerase activity
is also common in SCLC.
Proliferating Cell Nuclear Antigen (PCNA) [43-51]
[0161] PCNA functions as a cofactor for DNA polymerase delta. PCNA
is expressed in both S phase of the cell cycle and during periods
of DNA synthesis associated with DNA repair. PCNA is expressed in
proliferating cells in a wide range of normal and malignant
tissues. The cellular localization of PCNA is nuclear. Expression
of PCNA is a common feature of rapidly dividing cells and is
detected in 98% of tumors. Immunohistochemical staining is nuclear
with moderate to intense staining detected in 83% of NSCLC. Intense
PCNA staining was observed in 51% of p53-negative tumors. However,
when both PCNA (>50% of cells staining) and p53 are
overexpressed (>10% of cells stained) the prognosis tends to be
poorer with a shorter time to progression. Although frequently
detected in all stages of lung cancer, intense staining for PCNA is
more common in metastatic disease. Thirty-one percent of CIS also
overexpress PCNA.
[0162] CD44 [51-58]
[0163] CD44v6 is a cell surface glycoprotein that acts as a
cellular adhesion molecule. It is expressed on a wide range of
normal and malignant cells in epithelial, mesothelial and
hematopoietic tissues. The expression of specific CD44 splice
variants has been shown to be associated with metastasis and poor
prognosis in certain human malignancies. It is expected to be used
for detection and discrimination between squamous cell carcinoma
and adenocarcinoma. CD44 is a cell adhesion molecule that appears
to play a role in tumor invasion and metastasis. Alternative
splicing results in the expression of several variant isoforms.
CD44 expression is generally lacking in SCLC and is variably
expressed in NSCLC. Highest levels of expression occur in squamous
cell carcinoma, thus making it valuable in discriminating between
tumor types. In non-neoplastic tissue, CD44 staining is observed in
bronchial epithelial cells, macrophages, lymphocytes, and alveolar
pneumocytes. There was no significant correlation between CD44
expression and tumor stage, recurrence, or survival particularly
when overexpression occurs in early stage disease. In metastatic
lesions 100% of squamous cell carcinoma and 75% of adenocarcinoma
showed strong CD44v6 positivity. These data would tend to indicate
that changes in CD44 expression occur relatively late in tumor
progression that could limit its value as an early detection
marker. Recent findings suggest that the CD44v8-10 variant is
expressed by a majority of NSCLC making it a possible candidate
marker.
[0164] Cyclin A [59-62]
[0165] Cyclin A is a regulatory subunit of the cyclin-dependent
kinases (CDK's) which control the transition points at specific
phases of the cell cycle. It is detectable in S phase and during
progression into G2 phase. The cellular localization of Cyclin A is
nuclear.
[0166] Protein complexes consisting of cyclins and cyclin-dependent
kinases function to regulate cell cycle progression. Changes in
cyclin expression are associated with genetic alterations affecting
the CCDN1 gene. While the cyclins act as regulatory molecules, the
cyclin-dependent kinases function as catalytic subunits activating
and inactivating Rb.
[0167] Immunohistochemical analysis has revealed that the
overexpression of the cyclins is associated with an increase in
cellular proliferation as indicated by a high Ki-67 labeling index.
Cyclin overexpression occurs in 75% of NSCLC and appears to occur
relatively early in tumor progression. Recent reports indicate that
66.7% of stage I/II and 70.9% of stage III tumors overexpress
Cyclin A. Nuclear staining is common in poorly differentiated
tumors. Expression of cyclin A is often associated with a decrease
in mean survival time and a tendency towards the development of
drug resistance. However, increased expression has also been
associated with a greater response to doxorubicin.
[0168] Cyclin D1 [63-73]
[0169] Cyclin D1, as with Cylcin A, is a regulatory subunit of the
cyclin-dependent kinases (CDK's) which control the transition
points at specific phases of the cell cycle. Cyclin D1 regulates
the entry of cells into S phase of the cell cycle. This gene is
frequently amplified and/or its expression deregulated in a wide
range of human malignancies. The cellular localization of Cyclin D1
is nuclear.
[0170] Like Cyclin A, cyclin D1 functions to regulate cell cycle
progression. Staining of cyclin D1 is predominately cytoplasmic and
independent of histologic type. Reports suggest that cyclin D1
overexpression occurs in 40-70% of NSCLC and 80% of SCLC. Cyclin
D1, staining was observed in 37.9% of stage I, 60% stage II, and
57.9% of stage III tumors. Cyclin D1 expression has also been seen
in dysplastic and hyperplastic tissue providing evidence that these
changes occur relatively early in tumor progression. Patients who
overexpress cyclin D1 exhibit shorter mean survival time and lower
five-year survival rate.
[0171] Hepatocyte Growth Factor Receptor (C-MET) [74-77]
[0172] C-MET is a proto-oncogene that encodes a transmembrane
receptor tyrosine kinase for HGF. HGF is a mitogen for hepatocytes
and endothelial cells, and exerts pleitrophic activity on several
cell types of epithelial origin. The cellular localization of C-MET
is the cell surface.
[0173] Hepatocyte growth factor/scatter factor (HGF/SF) stimulates
a broad spectrum of epithelial cells causing them to proliferate,
migrate, and carry out complex differentiation programs including
angiogenesis. HGF/SF binds to a receptor encoded by the c-MET
oncogene. While both normal and malignant tissues express the HGF
receptor, expression of HGF/SF appears to be limited to malignant
tissue.
[0174] While the human lung generally expresses low levels of
HGF/SF, expression increases markedly in NSCLC. Using Western blot
analysis, 88.5% of lung cancers exhibited an increase in the
protein expression. All histologic types of tumors expressed the
protein at increased concentrations. While increased levels of
protein occur in all stages of the disease, recent evidence
suggests that in addition to the cancer cells, stromal cells and/or
inflammatory cells may be responsible for the production of the
growth factor.
[0175] Mucin--MUC-1 [78-82]
[0176] Mucin-1 comes from a family of highly glycosylated secretory
proteins which comprise the major protein constituents of the
mucous gel which coats and protects the tracheobronchial tree,
gastrointestinal tract and genitourinary tract. Mucin-1 is
atypically expressed in epithelial tumors. The cellular
localization of Mucin-1 is cytoplasm and the cell surface.
[0177] Mucins are a family of high molecular weight glycoproteins
that are synthesized by a variety of secretory epithelial cells
that are either membrane bound or secreted. Within the respiratory
tract, these proteins contribute to the mucus gel that coats and
protects that tracheobronchial tree. Changes in mucin expression
commonly occur in conjunction with malignant transformation
including lung cancer. Evidence exists suggesting at these changes
may contribute to alterations in cell growth regulation,
recognition by the immune system, and the metastatic potential of
the tumor.
[0178] Although normal lung tissue expresses MUC-1, significantly
higher levels of expression are found in lung cancer with highest
levels occurring in adenocarcinoma. Staining appears to occur
independently of stage and is more common in smokers than in former
smokers or nonsmokers. Some premalignant lesions also exhibit
increased MUC-1 expression.
Thyroid Transcription Factor-1 (TTF-1) [83,84]
[0179] TTF-1 belongs to a family of homeodomain transcription
factors that activate thyroid-specific and pulmonary-specific
differentiation genes. The cellular localization of TTF-1 is
nuclear.
[0180] TTF-1 is a protein originally found to mediate the
transcription of thyroglobulin. Recently, TTF-1 expression was also
found in the diencephalon and brohchioloalveolar epithelium. Within
the lung TTF-1 functions as a transcription factor regulating the
synthesis of surfactant proteins and clara secretory protein.
Overexpression of TTF-1 occurs in a large proportion of lung
adenocarcinomas and can aid in distinguishing between primary lung
cancer and cancers that metastasize to the lung. Adenocarcinomas
that express TTF-1 and are cytokeratin 7 positive and cytokeratin
20 negative can be detected with 95% sensitivity. Vascular
Endothelial Growth Factor (VEGF) [33,61,85-89]
[0181] VEGF plays an important role in angiogenesis, which promotes
tumor progression and metastasis. There are multiple forms of VEGF;
the two smaller isoforms are secreted proteins and act as
diffusible agents, whereas the larger two remain cell associated.
The cellular localization of VEGF is cytoplasmic, cell surface, and
extracellular matrix.
[0182] Vascular Endothelial Growth Factor (VEGF) is an important
angiogenesis factor and endothelial cell-specific mitogen.
Angiogenesis is an important process in the latter stages of
carcinogenesis, tumor progression and is particularly important in
the development of distant metastasis. VEGF binds to a specific
receptor Flt that is often present in the tumors expressing the
growth factor suggesting the presence of an autocrine loop.
[0183] Immunohistochemical analysis reveals that cells expressing
VEGF exhibit a pattern of staining that is diffuse and cytoplasmic.
While not expressed by normeoplastic cells, VEGF is present in the
majority of NSCLC and in a smaller percentage of SCLC. Several
reports have shown high levels of VEGF in early stage lung cancer.
Expression of VEGF has been associated with an increased frequency
of metastasis. Studies have shown that VEGF expression is
indicative of a poor prognosis and shorter disease-free interval in
adenocarcinoma but not in squamous cell carcinoma. Three year and
five year survival rates in the group expressing high levels of
VEGF were 50% and 16.7% as compared to 90.9 and 77.9% respectively
for the low VEGF group.
Epidermal Growth Factor Receptor (EGFR) [90-104]
[0184] Epidermal Growth Factor Receptor (EGFR) is a transmembrane
glycoprotein, which can bind and become activated by various
ligands. Binding initiates a chain of events that result in DNA
synthesis, cell proliferation, and cell differentiation. EGFR has
been demonstrated in a broad spectrum of normal tissues, and EGFR
overexpression is found in a variety of neoplasms. Increased
expression has been observed in adenocarcinomas of the lung and
large cell carcinomas but not in small cell lung carcinomas. The
cellular localization of EGFR is the cell surface.
[0185] The EGFR plays an important role in cell growth and
differentiation. The EGFR is uniformly present in the basal cell
layer but not in more the superficial layers of histologically
normal bronchial epithelium. With this exception, there is no
consistent staining of normal tissue. Recent evidence suggests that
the overexpression of the EGF receptor may not be an absolute
requirement for the development of invasive lung cancer. However,
it appear that in cases where EGFR overexpression occurs it is a
relatively early event with greater staining intensity in more
advanced disease.
[0186] For patients with invasive carcinomas, 50-77% of tumors
stain for EGF. Overexpression of the EGFR is more common in
squamous cell carcinoma than in adenocarcinoma and common in SCLC.
Highest levels of EGFR occur in conjunction with late stage and
metastatic disease that have approximately twice the concentration
of EGFR as that seen in stage I/II tumors. Estimates suggest that
the level of the EGFR observed in stage I tumors is approximately
twice that seen in normal tissue. In addition, 48% of bronchial
lesions also show EGFR staining including, metaplasia, atypia,
dysplasia, and CIS. In the "normal" bronchial mucosa, of these same
cancer patients, overexpression of the EGFR was observed in 39% of
cases but was absent in the bronchial epithelium of the non-cancer.
In addition, overexpression of the EGFR occurs more frequently in
the tumors of smokers than in nonsmokers, particularly in the case
of squamous cell carcinoma.
[0187] While several studies have suggested that overexpression of
the EGFR is associated with the poor prognosis, other studies have
failed to make this correlation.
[0188] Nucleoside Diphosphate Kinase/nm23 [105-111]
[0189] Nucleoside diphosphate kinase (NDP kinase)/nm23 is a
nucleoside diphosphate kinase. Tumor cells with high metastatic
potential often lack or express only a low amount of nm23 protein,
hence the nm23 protein has been described as a metastasis
suppressor protein. The cellular localization of nm23 is nuclear
and cytoplasmic.
[0190] Expression of nm23/nucleoside diphosphate/kinase A (nm23) is
a marker of tumor progression where there is an inverse
relationship between expression and metastatic potential. In cases
where stage I tumors overexpress nm23, no evidence of metastasis
was seen during an average follow-up period of 35 months.
Immunohistochemical analysis reveals staining that is diffuse,
cytoplasmic and generally limited to malignant cells. Alveolar
macrophages also express the protein. Given that high levels of
expression are associated with a low metastatic potential, there is
currently no explanation as to why normal epithelial cells do not
express nm23.
[0191] Intense staining has been observed in high percentage of
NSCLC particularly large cell lung cancer and 74% of SCLC
suggesting that this protein plays an important role in tumor
progression. With the exception of squamous cell carcinoma,
staining intensity tends to increase with stage. Based on the
available evidence, it would appear that nm23 is a prognostic
factor in both SCLC and NSCLC.
[0192] Bcl-2 [101,112-125]
[0193] Bcl-2 is a mitochondrial membrane protein that plays a
central role in the inhibition of apoptosis. Overexpression of
bcl-2 is a common feature of cells in which programmed cell death
has been arrested. The cellular localization of Bcl-2 is the cell
surface.
[0194] Bcl-2 is a protooncogene believed to play a role in
promoting the terminal differentiation of cells, prolonging the
survival of non-cycling cells and blocking apoptosis in cycling
cells. Bcl-2 can exist as a homodimers or can form a heterodimer
with Bax. As a homodimer, Bax functions to induce apoptosis.
However, the formation of a Bax-bcl-2 complex blocks apoptosis. By
blocking apoptosis, bcl-2 expression appears to confer a survival
advantage upon affected cells. Bcl-2 expression may also play a
role in the development of drug resistance. The expression of bcl-2
is negatively regulated by p53.
[0195] Immunohistochemistry analysis of bcl-2 reveals a
heterogeneous pattern of cytoplasmic staining. In adenocarcinoma,
expression of bcl-2 was significantly associated with smaller
tumors (<2 cm) and lower proliferative activity. The expression
of bcl-2 appears to be more closely associated with neuroendocrine
differentiation and occurs in a large percentage of SCLC.
[0196] Overexpression of bcl-2 is not present in preneoplastic
lesions suggesting that changes in bcl-2 occur relatively late in
tumor progression. In addition to tumor cells, bcl-2 immunostaining
also occurs in basal cells and on the luminal surfaces of normal
bronchioles but is generally not detected in more differentiated
cell types.
[0197] Association of bcl-2 immunoreactivity with improved
prognosis in NSCLC is controversial. Several reports of suggested
that patients with tumors expressing bcl-2 have a superior
prognosis and a longer time to recurrence. Several reports indicate
that bcl-2 expression tends to be lower in those patients who
develop metastatic disease. For patients with squamous cell
carcinoma, expression of bcl-2 has been linked to an improvement in
5-year survival. However, in three relatively large studies there
was no survival benefit linked to bcl-2 expression, particularly
for patients with early stage disease.
[0198] Estrogen Receptor-Related Protein (p29) [126]
[0199] ER related protein p29 is an estrogen-related heat shock
protein that has been found to correlate with the expression of
estrogen-receptor. The cellular localization of p29 is
cytoplasmic.
[0200] Estrogen-dependent intracellular processes are important in
the growth regulation of normal tissue and may play a role in the
regulation of malignancies. In one study expression of p29 was
detected in 109 (98%) of 111 lung cancers. The relation between p29
expression and survival time was different for men and women.
Expression of p29 was associated with poorer survival particularly
in women with Stage I and II disease. There was no correlation
between p29 expression and long-term survival in men.
[0201] Retinoblastoma Gene Product (Rb) [68,73,123,127-141]
[0202] Retinoblastoma Gene Product (Rb) is a nuclear DNA-binding
phosphoprotein. Under phosphorylated Rb binds oncoproteins of DNA
tumor viruses and gene regulatory proteins thus inhibiting DNA
replication. Rb protein may act by regulating transcription; loss
of Rb function leads to uncontrolled cell growth. The cellular
localization of Rb is nuclear.
[0203] Retinoblastoma protein (pRb) is a protein that is encoded by
the retinoblastoma gene and is phosphorylated and dephosphorylated
in a cell cycle dependent manner. pRb is considered an important
tumor suppressor gene that functions to regulate the cell cycle at
G0/G1. In its hypophosphorylated state, pRb inhibits the transition
from G1 to S. During G1, inactivation of the growth suppressive
properties of pRb occurs when the cyclin dependent kinases (CDK's)
phosphorylate the protein. The hyperphosphorylation of pRb prevents
it from forming a complex with E2F that functions as a
transcription factor proteins that are required for DNA
synthesis.
[0204] Inactivation of the retinoblastoma (Rb) gene has been
documented in various types of cancer, including lung cancer.
Small-cell carcinomas fail to stain for pRb indicating loss of Rb
function. Overall, 17.6% of the tumors fail to express pRb with no
correlation being seen with respect to stage or nodal status. A
reduction in staining has also seen in 31% dysplastic bronchial
biopsies. However, there appears to be no correlation between pRb
expression and the severity of dysplasia. In contrast, normal
bronchial epithelium and cells taken from areas adjacent to tumors
expressed pRb positive nuclei. These data suggest that alterations
in the expression of the Rb protein may arise early in the
development of some lung cancers.
[0205] Patients with Rb-positive carcinomas tend to have a somewhat
better prognosis but, in most studies, the difference is not
significant. However, patients with adenocarcinoma whose tumors are
both pRb negative and either p53 or ras positive exhibit a decrease
in 5-year survival. A similar relationship does not occur in
squamous cell carcinoma. pRb negative tumors have been reported to
be more likely to exhibit resistant to doxorubicin than Rb-positive
carcinomas.
[0206] Thrombomodulin [142-147]
[0207] Thrombomodulin is a transmembrane glycoprotein. Through its
accelerated activation of protein C (which in turn acts as an
anticoagulant by binding protein S and thrombin), synthesis of TM
is one of several mechanisms important in reducing clot formation
on the surface of endothelial cells. The cellular localization of
thrombomodulin is the cell surface.
[0208] Aggregation of host platelets by circulating tumor cells
appears to play an important role in the metastatic process.
Thrombomodulin plays an important role in the activation of the
anticoagulant protein C by thrombin and is an important modulator
of intravascular coagulation. In addition to its expression in
normal squamous epithelium, expression of thrombomodulin also
occurs in squamous metaplasia, carcinoma in situ, and invasive
squamous cell carcinomas. Although present in 74% of primary
squamous cell carcinomas, only 44% of metastatic lesions stained
for thrombomodulin. These data suggest that, with progression,
there is a decrease in thrombomodulin expression. Higher levels of
expression tend to occur in well and moderately differentiated
tumors when compared to poorly differentiated tumors.
[0209] Patients with thrombomodulin-negative squamous cell
carcinoma tend to have a worse prognosis. Eighteen percent of
patients with thrombomodulin-negative have a five-year survival as
compared to 60% in cases where the tumors stained positive for the
protein. Progression to metastatic disease was also more common in
thrombomodulin-negative tumors (69% vs. 37%) and there was a
greater tendency for these tumors to develop at extrathorasic
sites. Thus, loss of thrombomodulin expression appears to be
prognostic in cases of squamous cell carcinoma. The observation
that changes in thrombomodulin expression occur in later stages of
NSCLC and that the protein is expressed by normal bronchial
epithelial cells would tend to limit its utility as a marker for
early detection. However, since a majority of mesotheliomas and
only a small percentage of adenocarcinomas express thrombomodulin,
the marker has potential utility in discriminating between these
two tumor types.
[0210] E-Cadherin & N-Cadherin [148-151]
[0211] E-cadherin is a transmembrane Ca2+ dependent cell adhesion
molecule. It plays an important role in the growth and development
of cells via the mechanisms of control of tissue architecture and
the maintenance of tissue integrity. E-cadherin contributes to
intercellular adhesion of epithelial cells, the establishment of
epithelial polarization, glandular differentiation, and
stratification. Down-regulation of E-cadherin expression has been
observed in a number of carcinomas and is usually associated with
advanced stage and progression. The cellular localization of
E-cadherin is the cell surface.
[0212] E-cadherin is a calcium-dependent epithelial cell adhesion
molecule. A decrease in E-cadherin expression has been associated
with tumor dedifferentiation and metastasis and decreased survival.
Reduced expression has been observed in moderately and poorly
differentiated squamous cell carcinoma and in SCLC. There was no
change in E-cadherin expression in adenocarcinoma. Furthermore,
while adenocarcinomas express E-cadherin theses tumors fail to
express N-cadherin which is in contrast to mesotheliomas that
express N-cadherin but not E-cadherin. Thus, these markers can be
used to discriminate between adenocarcinoma and mesothelioma.
[0213] Expression of E-cadherin can also be used to assess the
prognosis of patients with squamous cell carcinoma. Whereas 60% of
patients with tumors expressing E-cadherin survived three-year
survival, only 36% of patients exhibiting a reduction in expression
survived 3 years.
[0214] MAGE-1 and MAGE-3 [152-156]
[0215] Melanoma Antigen-1 (MAGE-1) and Melanoma Antigen-3 (MAGE-3)
are members of a family of genes that are normally silent in normal
tissues but when expressed in malignant neoplasms are recognized by
autologous, tumor-directed and specific cytotoxic T cells (CTL's).
The cellular localization of MAGE-1 and MAGE-3 is cytoplasmic.
[0216] MAGE-1, MAGE-3 and MAGE 4 gene products are tumor-associated
antigens that are recognized by cytotoxic T lymphocytes. As such,
they could have utility as targets for immunotherapy in NSCLC. MAGE
proteins are also expressed by some SCLCs but not by normal cells.
While the frequency of MAGE expression falls below the level
necessary for use as a detection marker, differences in the pattern
of expression between histologic types suggest that MAGE expression
may have utility as differentiation markers. This utility is also
supported by the observation that, in 50% of squamous cell
carcinoma greater than 90% of tumor cells showed evidence of MAGE-3
overexpression with 30% to tumors exhibiting overexpression in at
least 50% of cells.
[0217] Nucleolar Protein (p120) [157]
[0218] p120 (proliferation-associated nucleolar antigen) is found
in the cells of nucleoli of rapidly proliferating cells during
early G1 phase. The cellular localization of p120 is nuclear.
[0219] Nucleolar protein p120 is a proliferation-associated protein
whose function has yet to be elucidated. Strong staining has been
detected in tumor tissue but not in macrophages or normal tissue.
Overexpression of p120 was more common in squamous cell carcinoma
that in adenocarcinoma or large cell carcinoma raising the
possibility that this marker may have utility in discriminating
between tumor types.
[0220] Pulmonary Surfactants [83,158-166]
[0221] Pulmonary surfactants are a phospholipid-rich mixture that
functions to reduce the surface tension at the alveolar-liquid
interface, thus providing the alveolar stability necessary for
ventilation. Surfactant proteins appear to be expressed exclusively
in the airway and are produced by alveolar type II cells. In the
non-neoplastic lung, pro-surfactant-B immunoreactivity is detected
in normal and hyperplastic alveolar type II cells and some
non-ciliated bronchiolar epithelial cells. Sixty percent of
adenocarcinomas contained strong cytoplasmic immunoreactivity with
10-50% of tumor cells exhibiting staining the majority of cases.
Squamous cell carcinoma and large cell carcinoma failed to stain
for pro-surfactant-B.
[0222] Surfactant Apoprotein B (SP-B) is one in four hydrophobic
proteins that make up the pulmonary surfactant, which is a
phospholipid and protein complex secreted by type II alveolar
cells. Squamous cell and large cell carcinomas of the lung and
nonpulmonary adenocarcinomas do not express SP-B. The cellular
localization of SP-B is cytoplasmic.
[0223] SP-A is a pulmonary surfactant protein that plays an
essential role in keeping alveoli from collapsing at the end of
expiration. SP-A is a unique differentiation marker of pulmonary
alveolar epithelial cells (type II pneumocytes); the antigen is
preserved even in the neoplastic state. The cellular localization
of SP-A is cytoplasmic.
[0224] Pulmonary surfactant A appears to be specific for
non-mucinous bronchoiolo-alveolar carcinoma with 100% staining as
compared to none of the of mucinous type. Pulmonary surfactants
potentially have utility in discriminating lung cancer from other
cancers metastasized to lung. In addition to tumor cells,
non-neoplastic pheumocytes also stain for pulmonary surfactant A.
As with pulmonary surfactant B staining for pulmonary surfactant A
is relatively common in adenocarcinoma but not in other forms of
NSCLC or in SCLC. Mesothelioma also fails to express pulmonary
surfactant A leading to the suggestion that pulmonary surfactant A
may have utility in the discrimination between adenocarcinoma and
mesothelioma.
[0225] Ki-67
[0226] Ki-67 is a nuclear protein that is expressed in
proliferating normal and neoplastic cells and is down-regulated in
quiescent cells. It is present in G1, S, G2, and M phases of the
cell cycle, but is absent in G0 phase. Commonly used as a marker of
proliferation. The cellular localization of Ki-67 is nuclear.
TABLE-US-00005 TABLE 5 Squamous Cell Large Cell Small Cell Marker
Carcinoma Adenocarcinoma Carcinoma Carcinoma Mesothelioma Glut1
100.0.sup.+ 64.5 80.5 64.0 NDA* Glut3 17.5 16.0 39.5 9.0 NDA* HERA
100.0 100.0 100.0 NDA 4.5 Basic FGF 83.0 48.7 50.0 100.0 NDA
Telomerase 82.3 86.3 93.0 66.7 NDA PCNA 80.0 69.8 87.7 51.0 NDA
CD44v6 79.3 34.8 44.2 0.0 NDA Cyclin A 79.0 68.0 83.5 97.0 NDA
Cyclin D1 42.7 36.0 62.0 90.0 NDA Hepatocyte Growth 75.5 78.3 100.0
NDA 100.0 Factor/Scatter Factor MUC-1 55.5 90.0 100.0 100 NDA TTF-1
38.0 76.0 NDA 83.0 NDA VEGF 61.8 68.3 100.0 43.5 NDA EGF Receptor
63.1 45.3 96.0 Frequently NDA nm23 68.0 52.6 83.5 73.5 NDA Bcl-2
45.5 43.3 42.5 92.0 NDA Loss of pRb Expression 20.1 25.8 35.4 85.3
NDA Thrombomodulin 66.8 12.2 4.0 0.0 81.0 E-cadherin 69.0 85.0 NDA
100.0 0.0 N-cadherin NDA 4.0 NDA NDA 94.0 MAGE 1 45.0 35.0 NDA 16.5
NDA MAGE 3 72.0 33.3 NDA 33.5 NDA MAGE 4 45.5 11.0 NDA 50.0 NDA
Nucleolar Protein (p120) 68.0 35.0 30.0 NDA NDA Pulmonary
Surfactant B 0.0 61.5 0.0 NDA NDA Pulmonary Surfactant A 12.0 52.9
17.5 20 0.0 .sup.+percent of tumors exhibiting a change in marker
expression *No Data Available
[0227] a. Obtaining a Library of Marker of a Suitable Size
[0228] Preliminary pruning steps were required in order to obtain a
suitable size library of markers that were correlated with lung
cancer. More than a hundred markers correlated to lung cancer are
known in the literature. A partial listing of candidate probes
identified in the literature and evaluated for potential inclusion
in panels tests include antibodies to: bax, Bcl-2, c-MET (HGFr),
CD44S, CD44v4, CD44v5, CD44v6, cdk2 kinase, CEA (carcino-embryonic
antigen), Cyclin A, Cyclin D1 (bcl-1), E-cadherin, EGFR, ER-related
(p29), erbB-1, erbB-2, FGF-2 (bFGF), FOS, Glut-1, Glut-2, Glut-3,
Glut-4, Glut-5, HERA (MOC-31), HPV-16, HPV-18, HPV-31, HPV-33,
HPV-51, integrin VLA2, integrin VLA3, integrin VLA6, JUN, keratin,
keratin 7, keratin 8, keratin 10, keratin 13, keratin 14, keratin
16, keratin 17, keratin 18, keratin 19, A-type lamins (A; C),
B-type lamins (B1; B2), MAGE-1, MAGE-3, MAGE-4, melanoma-associated
antigen clone NKI/C3, mdm2, mib-1 (Ki-67), mucin 1 (MUC-1), mucin 2
(MUC-2), mucin 3 (MUC-3), mucin 4 (MUC-4), MYC, N-cadherin, NCAM
(neural cell adhesion molecule), nm23, p120, p16, p21, p27, p53,
P-cadherin, PCNA, Retinoblastoma, SP-A, SP-B, Telomerase,
Thrombomodulin, Thyroid Transcription Factor 1, VEGF, vimentin, and
waf1. The initial list of markers was pruned by initially
assessing, from the literature, the apparent effectiveness of the
probes in detecting early stage cancer cells, discriminating
between cells of differing cancer states, and localizing the label
to the target cancer cells. This list of markers was further pruned
by removing markers whose utilization would be difficult to reduce
to practice because they are difficult to produce or obtain, have
unsuitable detection technology requirements or poor
reproducibility of reported results. After all of the pruning steps
were complete, a library of 27 markers was obtained.
[0229] b. Optimizing Protocols and Obtaining Gold Standard Lung
Cancer Samples
[0230] Preliminary preparation steps were also required prior to
obtaining the panels. The probes containing appropriate labels were
available from commercial vendors. The protocols of the probes were
analyzed for optimum objective quantitative detection. For example,
it was determined that the concentration of PCNA was too low.
Originally, PCNA was diluted 1:4000 in S809 buffer. A second
dilution was made, which was 1:3200 in S809. The optimized
protocols for each marker is shown in below. It is noted that the
second column is labeled "Antibody Name". Except for MOC-31, the
probes in this list are listed by the marker name because many of
the vendors refer to the antibody by the name of the marker. It is
noted that an alternative way these reagents might be listed is,
for example, anti-VEGF, anti-Thrombomodulin, anti-CD44v6, etc.
[0231] Gold standard tissue specimens were obtained from UCLA.
Tissue specimens were received from two sources. Cases had been
diagnosed using standard procedures including review of hematoxylin
and eosin (H&E)-stained slides and the clinical history.
Specimen slides were coded and labeled with arbitrary numbers to
blind the study pathologists to the historical diagnosis and
antibody marker and to protect patient confidentiality.
[0232] Specimen slides with tissue sections from cancerous and non
cancerous (control) tissues were used. A total of 175 separate
cases were analyzed. Within this set, the following diagnoses,
located in Table 6 were present with the following frequencies:
TABLE-US-00006 TABLE 6 Diagnosis Number of occurrences Cancer
Adenocarcinoma 25 Large Cell Carcinoma 18 Mesothelioma 26 Small
Cell Lung Cancer 20 Squamous Cell Carcinoma 24 Control Emphysema 34
Granulomatous Disease 3 Interstitial Lung Disease 25
[0233] c. Determination of the Level of Expression of the Panel of
Molecular Markers
[0234] Sufficient specimen slides were prepared for each case so
that only one probe was tested per slide. In general, a microscope
slide is prepared which contains the cytologic sample contacted
with one or more labeled probes that are directed at particular
molecular markers. Independently, each study pathologists examined
an H&E-stained slide to make a diagnosis for each case, and
then examined each probe-reacted and immunochemically-stained slide
to assess the level of probe binding, recording the results on a
standardized data form.
[0235] In greater detail, the immunohistochemical staining was
performed on formalin fixed, paraffin embedded (FFPE) tissue.
Tissue sections were cut at 4 microns thick on poly-L-Lysine coated
slides and dried at room temperature overnight. De-paraffinization
and rehydration of the tissue sections were performed as follows:
To completely remove all of the embedding medium from the specimen
the slides were incubated in two consecutive Xylene-substitute
(Histoclear) baths for five minutes each. All liquid was tapped off
the slides before incubation in two consecutive baths of 100%
reagent grade alcohol for three minutes each. Once again all excess
liquid was tapped off the slides before being incubated in two
final baths of 95% reagent grade alcohol for three minutes each.
After the last bath of 95% the slides were rinsed in tap water and
held in wash buffer (Tris-buffered saline wash buffer containing
0.05% Tween 20 corresponding to a 1:10 dilution of DAKO Autostainer
Wash buffer, code S3306). Table 7, below, presents a complete list
of the reagents used in this study along with corresponding product
code numbers. Detection systems used in the study were DAKO
EnVision+HRP mouse (code K4007) or rabbit (code K4003) and LSAB+HRP
(code K0690). The protocols for immunoassaying were followed
according to the package inserts. The kits contained liquid two
component DAB+ substrate chromogen (code
TABLE-US-00007 TABLE 7 Reagents used in the Study Reagents Code #
National Diagnostics HistoClear HS-200 Mallinckrodt Reagent
Alchohol Absolute 7019-10 DAKO Antibody Diluent S809 DAKO
Background Reducing Antibody Diluent S3022 DAKO Autostainer Buffer
10.times. S3306 DAKO Target Retrieval Solution S1700 DAKO Hi pH
Target Retrieval Solution S3307 DAKO Proteinase K S3020 Rite Aid
Hydrogen Peroxide 3% None DAKO Protein Block Serum Free X0909 DAKO
Goat Serum X0501 DAKO Swine Serum X0901 DAKO EnVision+ Mouse K4007
DAKO EnVision+ Rabbit K4003 DAKO LSAB+ K0690 DAKO DAB+ K3468 DAKO
Hematoxylin S3302 Dakomount Mounting Media S3025 Instruments Serial
Numbers DAKO Autostainers 3400-6613-03 3400-6142R-03 Autostainer
IHC Software Version V3.0.2
[0236] Pretreatments were critical in optimizing these antibodies
on lung tissue. For antibodies requiring enzyme digestion, DAKO
Proteinase K (code S3020) was used for 5 minutes at room
temperature. Antibodies requiring heat induced target retrieval
received pretreatment using either DAKO Target Retrieval Solution
(code S1700) or DAKO High pH Target Retrieval Solution (code
S3307). Tissues were placed in a pre-heated Target Retrieval
Solution and incubated in a 95.degree. C. water bath for 20 or 40
minutes depending on the specific protocol. Tissue sections were
then allowed to cool at room temperature for an additional 20
minutes.
[0237] After de-paraffinization, rehydration and tissue
pretreatment, all specimens were incubated in a solution of 3%
hydrogen peroxide to quench endogenous peroxidase activity.
Blocking reagents were used specifically for the two antibodies FGF
and Telomerase in order to minimize nonspecific background.
[0238] As shown in Table 8, below, tissue specimens were incubated
for a specified length of time with 200 micro liters of the
optimally diluted primary antibody. It is noted that the numbering
of the markers/antibodies in Table 8 is consistent with the
numbering of the antibody probes and markers throughout this
document. Slides were then washed in DAKO 1.times. Autostainer
Buffer (code S3306). Depending on the antibody, the correct
detection system was applied. The steps and total incubation times
for the DAKO EnVision+HRP and LSAB+HRP detection systems are shown
in Table 9, below. The color reaction is developed using
3,3'-diaminobenzidine (DAB) resulting in a brown color precipitate
at the site of the reaction.
TABLE-US-00008 TABLE 8 Antibodies for Lung Panel Antibody to #
Marker: Pretreatment Block Dilution Primary Inc Detection Sys Clone
Vendor Code# 1 VEGF Hi pH TRS 20 None 1:15 in S809 30 minutes
EnV+mouse JH121 NeoMarkers MS-350-P min S3307 2 Thrombomodulin None
None 1:100 in S809 30 minutes EnV+mouse 1009 DAKO M0617 3 CD44v6
TRS 20 min S1700 None RTU 30 minutes EnV+mouse VFF-7 NeoMarkers
MS-1093-R7 4 SP-A None None 1:200 in S809 30 minutes EnV+mouse PE10
DAKO M4501 5 Retinoblastoma TRS 40 min S1700 None 1:25 in S809 30
minutes EnV+mouse Rb1 DAKO M7131 6 E-Cadherin TRS 20 min S1700 None
1:100 in S809 30 minutes EnV+mouse NCH-38 DAKO M3612 7 Cyclin A TRS
20 min S1700 None 1:25 in S809 30 minutes EnV+mouse 6E6 Novocastra
NCL 117205 8 nm23 Hi pH TRS 20 min None 1:50 in S809 30 minutes
EnV+ rabbit Polyclonal DAKO A0096 S3307 9 Telomerase TRS 20 min
S1700 Prot Block 1:400 in S809 Overnight EnV+ rabbit Polyclonal
Alpha EST21-A X0909, 30 min Diagnostic w/5% goat serum X0501 10
Ki-67 TRS 40 min S1700 None 1:200 in S809 30 minutes EnV+mouse
IVAK-2 DAKO M7240 11 Cyclin D1 Hi pH TRS 20 min None 1:200 in S3022
30 minutes EnV+mouse DCS-6 DAKO M7155 S3307 12 PCNA Dilution 1 TRS
20 min S1700 None 1:4000 in S809 30 minutes EnV+mouse PC10 DAKO
M0879 13 MAGE-1 Hi pH TRS 20 min None 1:250 in S809 30 minutes
EnV+mouse MA454 NeoMarkers MS 1067 S3307 14 Mucin 1 TRS 20 min
S1700 None 1:40 in S809 30 minutes EnV+mouse VU4H5 Santa Cruz
Sc-7313 Biotech 15 SP-B TRS 20 min S1700 None 1:100 in S809 30
minutes EnV+mouse SPB02 NeoMarkers MS-1300-P1 16 HERA TRS 40 min
S1700 None 1:50 is S809 30 minutes EnV+mouse MOC-31 DAKO M3525 17
FGF-2 None Prot Block 1:50 in S809 Overnight EnV+mouse bFM-2
Upstate #05-118 X0909, 30 min Biotech w/5% swine serum X0901 18
C-Met Incomplete None Incomplete Incomplete EnV+mouse 8F11
Novocastra 118406 19 TTF-1 TRS 40 min S1700 None 1:25 in S809 30
minutes EnV+mouse 8G7G3/1 DAKO M3575 20 Bcl-2 Hi pH TRS 20 None
1:75 in S809 30 minutes EnV+mouse 124 DAKO M0887 min S3307 21 p120
TRS 20 min S1700 None 1:10 in S809 30 minutes EnV+mouse FB-2
Biogenex MU196-UC 22 N-Cadherin TRS 40 min S1700 None 1:75 in S809
30 minutes EnV+mouse 6G4 & DAKO N/A 6G11 23 EGFR Prot K 1:25
for None 1:1500 in S809 30 minutes EnV+mouse 2-18C9 DAKO K1492 5
min 24 Glut 1 TRS 40 min S1700 None 1:200 in S809 30 minutes LSAB+
Polyclonal Santa Cruz SC 1605 Biotech 25 ER-related (p29) TRS 40
min S1700 None 1:200 in S809 30 minutes EnV+mouse G3.1 Biogenex
MU171-UC 26 Mage 3 TRS 40 min S1700 None 1:20 in S809 30 minutes
EnV+mouse 57B G. Spagnoli N/A 27 Glut 3 TRS 20 min S1700 None 1:80
in S809 30 minutes LSAB+ Polyclonal Santa Cruz SC 7581 Biotech 28
PCNA Dilution 2 TRS 20 min S1700 None 1:3200 in S809 30 minutes
EnV+mouse PC10 DAKO M0879
TABLE-US-00009 TABLE 9 Detection Systems Used in the Study Steps 1
Deparafinization and rehydration 2 baths of Histoclear for 5 mins
each 2 baths of 100% alchohol for 3 mins each 2 baths of 95%
alchohol for 3 mins each Water Rinse 2 Pretreatments TRS 40 or 20
mins High pH TRS 20 mins Proteinase K for 5 mins Water Rinse 3
Peroxidase block Peroxide bath for 5 mins Water Rinse Buffer for 5
mins Protein Block for 30 mins after H2O2 Block 4 Primary Ab 30
mins or Overnight at room temp 5 Detection System EnV+ Systems
Labelled Polymer OR LSAB+ System Secondary Reagent 30 mins 15 mins
Secondary Ab link Tertiary Reagent 15 mins SA-HRP 6 Chromogen
Chromogen Chromogen 10 mins DAB+ 5 minsDAB+
[0239] Following immunostaining all slides were incubated in DAKO
Hematoxylin (code S3302) for 3 minutes and coverslipped using
DAKOMount Mounting Media (S3025). All protocols were run on DAKO
Autostainers (serial #'s 3400-6612-03 & 3400-6142R-03) using
the IHC software version 3.0.2.
[0240] Immunostaining was viewed under a light microscope to
determine that controls were correctly stained and tissues were
intact. Slides were labeled, boxed and sent to designated
pathologists for results interpretation. Trained pathologists
identified the type of cancer or other lesion seen in the samples.
Trained pathologists assessed the sensitivity to the marker probe
by estimating the staining density and proportion of cells stained.
These scores were entered in a data sheet for that patient. The
pathologists were blinded to the original diagnosis and antibody
marker used in the immunostaining. Each slide was read by at least
two pathologists and results recorded on a data collection form. To
provide additional integrity to the process, the method is repeated
with a second or third pathologist. The scores obtained can then be
matched to identify data entry errors. The additional data also
facilitates a better classifier design.
[0241] For each case, up to 27 slides were analyzed, each stained
for a marker coded with numbers 1 through to 17, 19 through to 28.
Staining for marker 18 (C-MET) could not be optimized and the
marker/probe was therefore not used. Pathologist 1 scored slides
from all 175 cases. Pathologist 2 scored slides from 99 of the
cases. Pathologist 3 scored slides from 80 of the cases.
[0242] Table 10 below shows how many cases of each diagnosis each
pathologist scored slides from:
TABLE-US-00010 TABLE 10 Pathol- Pathol- Diagnosis ogist 1 ogist 2
Pathologist 3 Cancer Adenocarcinoma 25 12 14 Large Cell Carcinoma
18 9 9 Mesothelioma 26 14 8 Small Cell Lung Cancer 20 12 6 Squamous
Cell Carcinoma 24 13 11 Control Emphysema 34 23 13 Granulomatous
Disease 3 3 2 Interstitial Lung Disease 25 13 17
[0243] For the purposes of some selected statistical analysis
techniques, it was necessary to consider only those cases that had
scores for all 27 slides present. Table 11 below shows how many
cases of each diagnosis were complete in terms of having scores
from all 27 slides.
TABLE-US-00011 TABLE 11 Pathol- Pathol- Diagnosis ogist 1 ogist 2
Pathologist 3 Cancer Adenocarcinoma 14 10 8 Large Cell Carcinoma 12
9 3 Mesothelioma 17 13 3 Small Cell Lung Cancer 7 9 1 Squamous Cell
Carcinoma 12 13 4 Control Emphysema 32 21 1 Granulomatous Disease 2
1 0 Interstitial Lung Disease 23 7 3
[0244] From this table, it can be calculated that each pathologist
scored the following total number of complete cases. Pathologist 1
scored all 27 slides for 119 of the cases Pathologist 2 scored all
27 slides for 83 of the cases. Pathologist 3 scored all 27 slides
for 23 of the cases.
[0245] The total number of cancer data points is 172. This
comprises 113 data points from Pathologist 1 and 60 data points
from Pathologist 2. The total number of control data points is 101.
This comprises 62 data points from Pathologist 1 and 39 data points
from Pathologist 2.
[0246] FIG. 3 shows a comparisons between H-scores for probes 7 and
15 in control tissue and in cancerous tissue. The x-axis shows the
H-scores while the y-axis shows the percent of cases with that
particular H-score. The difference in H-scores is apparent.
[0247] For each patient the scores were entered electronically into
a Pathology Review Form which consolidates the scores into a data
base showing the patient identifier together with diagnosis,
proportion of cells stained, and staining density. The proportions
and density were consolidated into a single "H-Score" obtained by
grading the intensity as: none=0, weak=1, moderate=2, intense=3,
and the percentage cells as: 0-5%=0, 6-25%=1, 26-50%=2, 51-75%=3,
>75%=4, and then multiplying the two grades together. For
example, 50% weakly stained plus 50% moderate stained would score
10=2.times.2+2.times.3. This is the standard scoring system
throughout the analysis, except for the section 3(f), below, titled
"Effect of Using other (non-H-score) objective scoring parameters",
which investigates alternative scoring systems.
[0248] Standard classification procedures were used to find the
best combination of probes. Typically these use a search procedure
such as the "Branch and Bound Algorithm" to find a hierarchy of the
best features, ranked according to a test of discriminating power,
and truncated according to a test of significance. This process
also defines the decision rule or rules for best
classification.
[0249] The performance of a classifier designed with these features
can be estimated from the data used to design the classifier. The
straightforward application of all the design data to the
classifier gives a very unsound estimate of performance.
[0250] The analysis of the data collected in the present example
provide the optimum selection of probes which provided the best
separation of classes. Therefore, panels were obtained that only
needed a few probes to perform the analysis. However the data
showed that near-optimum performance could be obtained with other
combinations of probes. Hence, the invention is flexible in being
adaptable to the availability of probes where cost or supply
problems may not allow the very best combination. In some cases,
the invention can simply be applied to the available features to
find an alternative combination. In other cases, the algorithm may
be used to select features which allows cost weightings to be
included in the selection process to arrive at a low cost
solution.
[0251] The design of data collection and analysis experiment was
chosen to avoid biases through the well established double blind
procedures where data collection and data analysis were done
independently.
[0252] In the first case the pathologists reviewed slides with
conventional staining to allow a diagnosis to be made. This
diagnosis was entered on the Pathology Review form. The
pathologists were then presented, in random order, with slides
stained by the marker probes for scoring the percentage of cells
stained and the relative intensity of the staining. The slides were
numbered to exclude information about the probe from the
pathologist. To allow data integrity to be checked two pathologists
reviewed all patients.
[0253] Data were consolidated into a database that was then
reviewed by a team of statisticians. Probes were numbered to render
their method of action as unseen during the analysis of their
effectiveness.
[0254] The first stage of the analysis was to check the integrity
of the data by comparing entries for each patient. Where large
differences were found, the data entries were checked and any
obvious errors were corrected. Unexplained differences were left in
the data.
[0255] The data were then separately analyzed by four
statisticians, using different techniques in recognition of the
fact that different statistical methodologies are suited to
different types of discriminating information in the data.
[0256] The first step in the process of selecting the best probe
combination is to divide the data into two sets, one for designing
a classifier and one for testing the performance of the classifier.
By selecting the design made with the design (train) set, but
showing the best performance evaluated on the test set, it can be
concluded with confidence that the classifier has generalized to
the structure of the data and not adapted to particular cases seen
in the training set.
[0257] In order to test for reliability the analysis was typically
repeated with many randomly selected sets of training data and test
data. This approach is generally accepted as giving good estimates
of the classifier performance. Where these tests showed
inconsistent selections of probes such probe selections were
discounted as unreliable.
[0258] d. Statistical Analysis and/or Pattern Recognition
[0259] 1. Introduction to Data Analysis
[0260] a. Input Data
[0261] i. Raw Data
[0262] For each patient the scores were entered electronically into
a Pathology Review Form that consolidates the scores into a
database showing the patient identifier together with diagnosis,
proportion of cells stained, and staining density.
[0263] ii. Computed Data
[0264] The efficiency of the score for each probe used in the
analysis is computed from the intensity/percentage tables. The
proportions and density are consolidated into a single "H-Score"
with a simple rule H=proportion stained.times.(3 if intense+2 if
moderate+1 if weakly stained). This is the feature value associated
with that probe.
[0265] iii. Alternative Computed Data Parameters
[0266] The H-score described above was heuristically derived, a
simple analysis to find a better way of combining percentages and
intensity failed to show a significant improvement over H-score
(Section 3(f), titled "Effect of Using other (non-H-score)
objective scoring parameters"). A larger data base may allow the
extraction of a better rule in future.
[0267] iv. User Supplied Weighting Criteria Per Marker
[0268] The invention is flexible in being adaptable to the
availability of features where cost or supply problems may not
allow the very best combination. For example, the invention can
simply be applied to the available features to find and alternative
combination. Alternatively, the algorithm used to select features
allows cost weightings to be included in the selection process to
arrive at a minimum cost solution. Marker performance estimates are
shown for combinations selected from all the markers collected or
only those from one supplier. It is also shown how the C4.5 package
can be used to down weight certain probes, say on the basis of
their high cost. These probe combinations do not perform as well as
the optimum combination, but the performance might be acceptable in
circumstances where cost is a significant factor.
[0269] v. User Supplied Weighting Criteria Per Class
[0270] Some of the methods used allow weightings to be applied to
the classes. This is available in C4.5 where the tree design can
optimize the cost. Also the Discriminant Function method gives a
single parameter output which can be used to give a desired false
positive or false negative probability. A plot of these parameters
for different threshold settings is known as the Receiver Operating
Curve.
[0271] vi. Detection Panels--Assumptions
[0272] A low probability of a false negatives was assumed to be
desirable for the cancer detection process (to avoid positive
patients being missed at the cost of an increased number of false
positives who would require re-screening). It was also assumed that
the cancer discrimination process would require a lower false
positive score (to minimize patients receiving the wrong
treatment).
[0273] It was assumed that detection panels requiring 6 or more
probes to achieve an acceptable performance would not be cost
effective. It was also assumed that a detection panel with a false
negative error rate of more than 5% would not be acceptable. Panels
falling outside this box are not accepted. This assumption
acknowledges that cytometric panels are likely to have a worse
performance than the histology based panels analyzed here. The
ultimate aim will be a cytometric panel which performs better than
20% error rate, this being approximately the performance of
cervical PAP smear screeners.
[0274] vii. Discrimination Panels--Assumptions
[0275] It was assumed that panels requiring 6 or more probes are
not cost effective and it was assumed that an error rate of better
than 20% is required. Panels falling outside this box were not
accepted.
[0276] b. Output Data
[0277] Outputs provided by the present analysis included:
[0278] Confusion Matrices, showing how data from the test set was
classified as either true positive, false positive, true negative
or false negative. These may be shown as actual counts or as
percentages. Confusion matrices are discussed in section 2(d)
titled "Performance Metrics". A confusion matrix shows how data
from a test set was classifiefd as either true positive, false
positive, true negative or false negative. An exemplary confusion
matrix, obtained from data analyzed by decision trees, is shown
below in table 12 for simultaneous discrimination of
adenocarcinoma, squamous cell carcinoma, large cell carcinoma,
mesothelioma and small cell carcinoma
TABLE-US-00012 TABLE 12 Large Small Adeno Squamous Cell
Mesothelioma Cell Adeno 67.74% 6.45% 19.40% 0.00% 6.45% Squamous
Cell 2.94% 76.47% 11.67% 0.00% 8.82% Large Cell 28.00% 8.00% 44.00%
8.00% 12.00% Mesothelioma 0.00% 25.64% 51.28% 89.74% 2.56% Small
Cell 0.00% 3.85% 23.08% 3.85% 69.23%
[0279] Error Rates, summarizing data in the confusion matrix as the
sum of all false classifications divided by the total number of
classifications made expressed as a percentage [0280] Receiver
Operating Characteristic (ROC) curves show the estimated percentage
(or per unit probability) of false positive and false negative
scores for different threshold levels in the classifier. An
indifferent classifier, unable to discriminate better than random
choice would present a ROC curve with equal true and false
readings. The area under this curve would be 50% (0.5 probability).
[0281] Area Under the Curve (AUC) is often used as an overall
estimate of classifier performance and most standard discriminant
function packages provide this AUC figure. A perfect classifier
would have 100% Area Under the Curve, and a useless classifier
would have an AUC near 50% (0.5). [0282] Sensitivity and
specificity (can be derived from the confusion matrix). See section
2(d)(iii) titled "Sensitivity and Specificity". [0283] Marker
correlation matrices. See FIG. 4.
[0284] i. Detection Panels: Composition
[0285] These panels are trained on data divided into two classes,
patients with any of the five cancers and patients with none of the
cancers. Not all probes were present for all patients. Where one or
more probes were missing for a particular analysis these cases were
excised from the data. Hence, where analysis was undertaken on
reduced numbers of probes the data set might include slightly more
cases.
[0286] The number of probes included in the analysis was 27.
Although in many cases a false probe was added where the data
entered for that probe was from a random number generator set to
generate numbers uniformly between zero and 12. This false probe
was included in much of the early analysis to ensure integrity in
the probe selection process. This false probe was also used in one
approach to progressively eliminate probes from the analysis.
Probes that contributed less information than the false probe could
be readily identified and excluded from the selection process.
Early elimination of such probes speeds the analysis and renders
the analysis less vulnerable to variations in results (noise)
caused by these probes.
[0287] ii. Detection Panel Performance
[0288] As outputs from this study, the probe combinations selected
by the different methodologies and their performance estimates in
terms of the confusion matrix, % error rate, and AUC are
reported.
[0289] iii. Detection Panels--Alternative Compositions
[0290] Detection panels were also selected from reduced sets of
probes. In one set of panels, performance measures of panels
weighted for commercially preferred markers were obtained. The
performances obtained when the best probe was removed from the
analysis to find a new combination of discriminating probes was
also analyzed. The performance of a single probe acting on its own
was found to be very high (probe 7). However, as shown below in the
performance diagrams, Table 13, evaluated using linear discriminant
analysis, the performance was improved as more markers were added.
The best subsets of probes were determined using best subsets
logistic regression. The improvement is statistically
significant.
TABLE-US-00013 TABLE 13 Cancer Control Probe 7 Cancer 87.93% 12.07%
Control 0.00% 100.00% Probes 7 and 16 Cancer 93.10% 6.90% Control
1.16% 98.84% Probes 7, 15 and 16 Cancer 90.52% 9.48% Control 1.16%
98.84% Probes 1, 7, 15, and 16 Cancer 90.52% 9.48% Control 0.00%
100.00% Probes 1, 4, 7, 15, and 16 Cancer 92.24% 7.76% Control
1.16% 98.84%
[0291] The best and second best subsets of probes (determined using
best subsets logistic regression) and evaluated using logistic
regression is shown below. AUC=Area under ROC curve. It is noted
that mean AUC is the average from 100 trials on random train and
test partitions (70%:30%). The results are shown below, in Table
14.
TABLE-US-00014 TABLE 14 Probes Mean AUC 7 94.28 28 80.14 7, 16 95
7, 15 94.59 7, 15, 16 95.94 1, 7, 16 95.33 1, 7, 15, 16 95.61 4, 7,
15, 16 95.34 1, 4, 7, 15, 16 95.3 1, 7, 11, 15, 16 95.57
[0292] iv. Discrimination Panels--Composition
[0293] For this part of the study five classifiers were designed
and tested, each designed to detect the presence of one of the
cancer from all patients with cancer. The application of this five
way pair-wise system allows doubtful cases to appear more than once
in the analysis, or not at all. Such cases can be identified and
subjected to closer scrutiny, re-testing or alternative testing
regimes.
[0294] Again the number of probes in the study was 27, with a false
probe used in the early stage to reduce the numbers in the
analysis
[0295] v. Discriminant Panels--Performance
[0296] The performance estimators described above were used to show
the performance of the best probe combinations discovered by the
different techniques
[0297] vi. Discriminant Panels--Alternative Composition
[0298] The analysis was repeated for a probe combination comprising
commercially preferred probes. Performance was degraded, but not
unusable for several reduced-set classifiers. Below, the best
subsets of probes without probe 7, determined using best subsets
logistic regression), is shown, as Table 15. The data was evaluated
using linear discriminat analysis.
TABLE-US-00015 TABLE 15 Cancer Control Probe 28 Cancer 0.706897
0.293103 Control 0.093023 0.906977 Probes 10 and 28 Cancer 0.793103
0.206897 Control 0.034884 0.965116 Probes 10, 15 and 28 Cancer
0.810345 0.189655 Control 0.011628 0.988372 Probes 1, 10, 15 and 28
Cancer 0.827586 0.172414 Control 0.011628 0.988372 Probes 1, 10,
15, 16 and 28 Cancer 0.827586 0.172414 Control 0.011628
0.988372
[0299] The best and second best subsets of probes with probe 7
(determined using best subsets logistic regression) and evaluated
using logistic regression is shown below. AUC=Area under ROC curve.
It is noted that mean AUC is the average from 100 trials on random
train and test partitions (70%:30%). The results are shown below,
in Table 16.
TABLE-US-00016 TABLE 16 Probes Mean AUC 28 79.36% 10 82.28% 10, 28
94.21% 15, 28 88.68% 10, 15, 28 92.90% 1, 10, 28 93.59% 1, 10, 15,
28 92.99% 8, 10, 15, 28 93.20% 1, 10, 15, 16, 28 93.13% 1, 8, 10,
15, 28 93.57%
[0300] 2. Data Analysis Methodology
[0301] In this section, the process of gaining an initial
understanding of the structure of the data as a guide to
interpreting results from the different methodologies used is
described.
[0302] a. Analysis of Variance
[0303] i. Pathologist-to-Pathologist Variability and Pooling
Pathologist Scores.
[0304] (1) t-Test
[0305] Two pathologists reviewed each patient's slides in this
clinical trial. Pathologist 1 reviewed all patients, Pathologist 2
also reviewed approximately half of this set and Pathologist 3
reviewed the remainder. With two independent estimates of the
H-score, the consistency of pathologist performance could be
tested.
[0306] A readily available statistical tool was used to test the
variability between pathologists. This is the paired-sample t-test.
This takes the difference between each pair of estimates, averages
these and expresses this as a proportion of the overall variances.
The t-test then converts this ratio into a probability estimating
the likelihood that the two samples sets came from the same
population (the P value).
[0307] This test was applied to the scores for each marker probe,
for all cases reviewed by Pathologist 1 and Pathologist 2, and also
for all cases reviewed by Pathologist 1 and Pathologist 3. Since
there were 27 tests applied (to cover all probes) a low value of
P=0.01 was selected as the "significant threshold". Results,
showing the P scores for each probe, and for the two pairs of
pathologists, are shown below, in Tables 17, 18, 19 and 20. It is
clear that Pathologist 1 and Pathologist 2 were more consistent
than Pathologist 1 and Pathologist 3.
TABLE-US-00017 TABLE 17 Pathologist 1, Pathologist 2 scores: X1 X2
X3 X4 X5 X6 X7 0.5875446 0.01051847 0.4659704 0.4659704 0.3772894
0.2307273 0.01001357 X8 X9 X10 X11 X12 X13 X14 0.004131056
0.7703014 0.1640003 0.2374452 0.9580652 0.1587876 0.001200265 X15
X16 X17 X18 X19 X20 X21 0.19742 0.3860899 0.3829022 NA 0.544601
0.08873848 0.1686243 X22 X23 X24 X25 X26 X27 X28 0.5428451
0.1912477 0.4031977 0.2477236 0.5673386 0.9174037 0.00339071
TABLE-US-00018 TABLE 18 Pathologist 1, Pathologist 2 scores
thresholded at 0.01 (.alpha. = 1% level of significance): X1 X2 X3
X4 X5 X6 X7 TRUE TRUE TRUE TRUE TRUE TRUE TRUE X8 X9 X10 X11 X12
X13 X14 FALSE TRUE TRUE TRUE TRUE TRUE FALSE X15 X16 X17 X18 X19
X20 X21 TRUE TRUE TRUE NA TRUE TRUE TRUE X22 X23 X24 X25 X26 X27
X28 TRUE TRUE TRUE TRUE TRUE TRUE FALSE
TABLE-US-00019 TABLE 19 Pathologist 2, Pathologist 3 scores: X1 X2
X3 X4 X5 X6 X7 3.814506e-09 0.0399131 0.1954867 5.671062e-05
0.01856276 0.2757166 0.2292583 X8 X9 X10 X11 X12 X13 X14
2.044038e-12 0.004166467 0.00983267 0.003710155 0.01461007
0.03312421 0.0003367823 X15 X16 X17 X18 X19 X20 X21 0.0005162036
0.2276537 0.002987705 4.267708e-06 0.007287372 0.1654067 X22 X23
X24 X25 X26 X27 X28 0.02400127 0.0009497766 2.478456e-07 0.1591684
0.08318303 3.122143e-05 1
TABLE-US-00020 TABLE 20 Pathologist 1, Pathologist 3 scores
thresholded at 0.01 (.alpha. = 1% level of significance):: X1 X2 X3
X4 X5 X6 X7 FALSE TRUE FALSE FALSE TRUE TRUE TRUE X8 X9 X10 X11 X12
X13 X14 FALSE FALSE FALSE FALSE TRUE TRUE FALSE X15 X16 X17 X18 X19
X20 X21 FALSE TRUE FALSE FALSE FALSE FALSE TRUE X22 X23 X24 X25 X26
X27 X28 TRUE FALSE FALSE TRUE TRUE FALSE TRUE
[0308] Because the H score is subjective it is prone to scale
factor differences and noise at marginal cases. So, in spite of the
three features which showed statistically different scores between
Pathologist 1 and Pathologist 2, this joint data was accepted as
representative of a measuring instrument. Pathologist 1 and
Pathologist 2 were combined into a single data set for the analysis
process. The results for Pathologist 3 were withheld for
independent testing purposes. Such tests using the Pathologist 3
data would be biased towards showing an under-performance because
of the significant differences.
[0309] The data from Pathologist 1 and Pathologist 2 were combined
by considering them as separate cases, with the variability giving
a degree of independence between the results for any one case. When
testing with such data the performance estimates will be biased
towards a more optimistic value. This is because samples coming
from the same patient may occur simultaneously in the training a
test subsets. This does not however invalidate the processes used
to find the best combination of features, it merely biases the
estimate of performance.
[0310] (2) Analysis of Variance of H-Scores
[0311] (a) Background
[0312] Within each probe, the H-scores may vary due to many
reasons. To the extent they vary consistently due to the type of
disease this is useful, variation due to which pathologist read the
slide is instructive, whereas random variation sets a limit on the
detection of the previous two sources of variation.
[0313] Analysis of Variance (ANOVA) is a standard technique for
splitting up the sources of variation in data and for testing its
statistical significance. ANOVA summarizes the total variation of a
set of data as a sum of terms which can be attributed to specific
sources, or causes, of variation.
[0314] ANOVA is available in many statistical packages. The public
domain package "R" was chosen ("The R Project for Statistical
Computing", http://www.R-project.org/).
[0315] (b) Aim
[0316] To perform ANOVA analyses on the H-score data from
pathologists 1 and 2 and to consider whether this data can be
safely merged into a single consistent set for further analysis for
the selection of panels.
[0317] (c) Methodology
[0318] From the database, data was selected from pathologists 1 and
2. Only data which was complete for a given probe was used in the
ANOVA for that probe.
[0319] The control categories of Emphysema, Granulomatous Disease,
and Interstitial Lung Disease were grouped together and called
"Normal" giving 6 levels within factor Disease.
[0320] Pathologist was coded as a factor with 2 levels (Pathologist
1, Pathologist 2).
[0321] An R script was written to perform a standard ANOVA analysis
for each probe in turn, using the factors: Disease, Pathologist,
and the interaction term Disease:Pathologist. The results are shown
in below, in Table 21. "Df" is defined as the degrees of freedom.
In a dataset of n observations, knowing n-1 deviations from the
mean, the nth is automatically determined. N-1 is the number of
degrees of freedom. Sum Sq and mean Sq are measures of variation. F
is a test statistic concerning the equality of two variances based
on the F distribution. Pr(>F) is the probability used to
determine whether or not the variability is statistically
significant.
TABLE-US-00021 TABLE 21 Analysis of Variance of H-Scores Df Sum Sq
Mean Sq F value Pr(>F) Probe 1 Disease 5 443.56 88.71 15.8202
3.690e-13 *** Pathologist 1 0.66 0.66 0.1174 0.7323
Disease:Pathologist 5 15.34 3.07 0.5470 0.7405 Residuals 204
1143.930 5.61 Probe 2 Disease 5 1067.39 213.48 24.1234 <2e-16
*** Pathologist 1 13.02 13.02 1.4709 0.2263 Disease:Pathologist 5
27.98 5.60 0.6324 0.6752 Residuals 249 2203.50 8.85 Probe 3 Disease
5 1098.49 219.70 21.0751 <2e-16 *** Pathologist 1 6.73 6.73
0.6458 0.4224 Disease:Pathologist 5 29.72 5.94 0.5703 0.7227
Residuals 243 2533.16 10.42 Probe 4 Disease 5 631.8 126.4 9.3707
3.454e-08 *** Pathologist 1 6.6 6.6 0.4869 0.4860
Disease:Pathologist 5 13.1 2.6 0.1939 0.9647 Residuals 246 3317.1
13.5 Probe 5 Disease 5 754.30 150.86 25.2826 <2e-16 ***
Pathologist 1 14.25 14.25 2.3875 0.1236 Disease:Pathologist 5 7.54
1.51 0.2528 0.9381 Residuals 248 1479.80 5.97 Probe 6 Disease 5
721.91 144.38 11.8515 2.771e-10 *** Pathologist 1 1.91 1.91 0.1568
0.6925 Disease:Pathologist 5 47.82 9.56 0.7850 0.5613 Residuals 246
2996.93 12.18 Probe 7 Disease 5 1171.47 234.29 77.6802 <2e-16
*** Pathologist 1 8.84 8.84 2.9294 0.08847 . Disease:Pathologist 5
46.36 9.27 3.0742 0.01063 * Residuals 209 630.37 3.02 Probe 8
Disease 5 209.82 41.96 6.4352 1.201e-05 *** Pathologist 1 12.66
12.66 1.9407 0.16483 Disease:Pathologist 5 71.20 14.24 2.1838
0.05654. Residuals 251 1636.76 6.52 Probe 9 Disease 5 197.21 39.44
8.4348 2.015e-07 *** Pathologist 1 7.33 7.33 1.5681 0.2116
Disease:Pathologist 5 24.56 4.91 1.0505 0.3884 Residuals 265
1239.17 4.68 Probe 10 Disease 5 1113.46 222.69 39.0730 <2e-16
*** Pathologist 1 1.01 1.01 0.1778 0.67371 Disease:Pathologist 5
62.45 12.49 2.1916 0.05635 . Residuals 213 1213.96 5.70 Probe 11
Disease 5 320.15 64.03 9.5553 2.416e-08 *** Pathologist 1 1.28 1.28
0.1918 0.6618 Disease:Pathologist 5 10.04 2.01 0.2996 0.9128
Residuals 245 1641.76 6.70 Probe 12 Disease 5 832.26 166.45 27.8793
<2e-16 *** Pathologist 1 0.18 0.18 0.0307 0.8610
Disease:Pathologist 5 15.16 3.03 0.5079 0.7701 Residuals 248
1480.68 5.97 Probe 13 Disease 5 46.594 9.319 7.8408 8.674e-07 ***
Pathologist 1 0.044 0.044 0.0368 0.8481 Disease:Pathologist 5
10.143 2.029 1.7069 0.1343 Residuals 210 249.584 1.188 Probe 14
Disease 5 1305.69 261.14 23.9460 <2e-16 *** Pathologist 1 28.66
28.66 2.6279 0.10630 Disease:Pathologist 5 142.90 28.58 2.6208
0.02492 * Residuals 243 2649.98 10.91 Probe 15 Disease 5 401.02
80.20 21.268 <2e-16 *** Pathologist 1 13.17 13.17 3.493 0.0630 .
Disease:Pathologist 5 6.17 1.23 0.327 0.8963 Residuals 214 807.02
3.77 Probe 16 Disease 5 2520.26 504.05 65.5572 <2e-16 ***
Pathologist 1 0.15 0.15 0.0194 0.8892 Disease:Pathologist 5 24.29
4.86 0.6318 0.6757 Residuals 247 1899.12 7.69 Probe 17 Disease 5
530.64 106.13 13.0178 2.426e-11 *** Pathologist 1 8.42 8.42 1.0325
0.31050 Disease:Pathologist 5 109.96 21.99 2.6975 0.02131 *
Residuals 266 2168.55 8.15 Probe 19 Disease 5 1670.86 334.17
29.1960 <2e-16 *** Pathologist 1 2.17 2.17 0.1895 0.6637
Disease:Pathologist 5 32.61 6.52 0.5698 0.7231 Residuals 248
2838.56 11.45 Probe 20 Disease 5 964.71 192.94 34.2760 <2e-16
*** Pathologist 1 8.83 8.83 1.5687 0.2116 Disease:Pathologist 5
19.60 3.92 0.6963 0.6267 Residuals 245 1379.12 5.63 Probe 21
Disease 5 6.927 1.385 2.0604 0.07076 . Pathologist 1 0.464 0.464
0.6906 0.40670 Disease:Pathologist 5 1.576 0.315 0.4687 0.79945
Residuals 263 176.830 0.672 Probe 22 Disease 5 640.16 128.03
31.7250 <2e-16 *** Pathologist 1 1.64 1.64 0.4058 0.5247
Disease:Pathologist 5 18.78 3.76 0.9305 0.4617 Residuals 247 996.81
4.04 Probe 23 Disease 5 1915.62 383.12 46.5565 <2e-16 ***
Pathologist 1 10.77 10.77 1.3092 0.2537 Disease:Pathologist 5 20.92
4.18 0.5084 0.7698 Residuals 246 2024.39 8.23 Probe 24 Disease 5
516.06 103.21 24.0786 <2e-16 *** Pathologist 1 9.52 9.52 2.2210
0.1376 Disease:Pathologist 5 12.48 2.50 0.5823 0.7135 Residuals 216
925.87 4.29 Probe 25 Disease 5 1761.26 352.25 34.5245 <2e-16 ***
Pathologist 1 11.51 11.51 1.1285 0.2891 Disease:Pathologist 5 41.49
8.30 0.8134 0.5411 Residuals 248 2530.33 10.20 Probe 26 Disease 5
399.85 79.97 13.6548 1.428e-11 *** Pathologist 1 0.30 0.30 0.0517
0.8204 Disease:Pathologist 5 14.81 2.96 0.5056 0.7719 Residuals 214
1253.31 5.86 Probe 27 Disease 5 117.92 23.58 6.2551 1.956e-05 ***
Pathologist 1 0.64 0.64 0.1695 0.6810 Disease:Pathologist 5 25.52
5.10 1.3539 0.2431 Residuals 212 799.31 3.77 Probe 28 Disease 5
1634.60 326.92 38.171 <2e-16 *** Pathologist 1 8.40 8.40 0.981
0.3229 Disease:Pathologist 5 16.15 3.23 0.377 0.8643 Residuals 267
2286.76 8.56 Signif. codes: 0 {acute over ( )}***' 0.001 {acute
over ( )}**' 0.01 {acute over ( )}*' 0.05 {acute over ( )}.' 0.1
{acute over ( )} ' 1
[0322] (d) Analysis of Results
[0323] In all cases (except for probe 21) the response of the
probes was related to disease. This is not surprising since the
probes have presumably been selected for this purpose. In no case
is the response of the probe related to pathologist (at the p=0.05
level). This indicates that it would be safe to merge this data and
use the two pathologists as two measurements on the data.
[0324] In a few cases, probes 7, 14, 17, there is some evidence of
an interaction term gaining significance. This indicates that there
may be some difference between pathologists in their scoring of
some diseases. Some of these cases may well be due to an occasional
outlier in the data.
[0325] (e) Conclusions
[0326] The results indicate that it is safe to merge this data for
further analysis. The data indicate that the slight interactions in
some cases between pathologist and disease appear to be attributed
to random sources.
[0327] ii. Patient to Patient Variability
[0328] The variability from patient to patient was measured by the
disease:disease variability of section 2(a)(i)(2) (see above,
"Analysis of Variance of H-Scores").
[0329] iii. Marker-to-Marker Variability
[0330] Histograms were plotted (PathologistData.xls, worksheet:
Histograms) showing the distribution of marker scores for each
probe for Control vs. Cancer.
[0331] b. Marker Correlation Matrix Analyses
[0332] The population correlation coefficient ("Applied
Mulitvariate Statistical Analysis", R. A. Johnson and D. W.
Wichern, 2nd Ed, 1988, Prentice-Hall, N.J.) measures the amount of
linear association between a pair of random variables. Typically
the distributions and associated parameters of the random variables
are not known and the population correlation coefficient cannot be
directly computed. In this case it is possible to compute the
sample correlation coefficient from sample data. See FIG. 4. The
sample correlation coefficient is, however, only an estimate of the
population correlation coefficient. Moreover, because it is
calculated on the basis of sample data it is possible, purely by
chance, that it may indicate a strong positive or negative
correlation when in reality there may be no actual relationship
between the corresponding random variables ("Modern Elementary
Statistics", J. E. Freund, 6th Ed, 1984, Prentice-Hall, N.J.).
[0333] The correlation coefficient measures the ability of one
variable to predict the other. A strong linear association does
not, however, imply a causal relationship. The square of the
correlation coefficient is called the coefficient of determination.
The coefficient of determination computed for a bivariate data set
measures the proportion of the variability in one variable that can
be accounted for by its linear relationship to the other. When
dealing with several variables, the correlation coefficient can be
calculated for each pair in turn and the set of coefficients can be
written as a matrix called the correlation matrix. See FIG. 4.
[0334] The H-scores for the individual markers can be modeled as
random variables. The sample correlation matrix for this
multivariate data set can be computed from the input data described
in the section titled "Input Data", above.
[0335] c. Pattern Recognition
[0336] Statistical pattern recognition is an approach to
classifying signals or geometric objects on the basis of
quantitative measurements (called features). Statistical pattern
recognition essentially reduces to the problem of dividing the
n-dimensional feature space into regions that correspond to the
categories or classes of interest.
[0337] Three different classifier methodologies employed in this
study are sensitive to different structural forms within the
data.
[0338] For the Decision Tree method a preliminary analysis of
different data combinations identified markers which were never
used by C4.5 for the detection panel. These were removed from the
analysis and this resulted in more consistent results, symptomatic
of the left-out probes only contributing noise to the selection
process.
[0339] Similarly a preliminary analysis of probes used in the
detection panels identified the noisy probes for removal prior to
the detailed analysis.
[0340] The Linear Discriminant Function method in SPSS has built-in
stepwise processes for reducing the numbers of markers in the
analysis. Typically, this reduced the probes used in the analysis
to between 2 and 7.
[0341] The Logistic Regression method in R and SAS implement
stepwise procedures for variable selection. In SAS, a best subsets
variable selection option is also provided. In R, the stepwise
methodology was used in conjunction with multiple random trials to
develop a heuristic method for selecting variables based on the
number of times a given feature was used in 100 random selections
of training and test data (split 70%:30% respectively). Features
with counts comparable to the count for artificial random feature
were progressively eliminated until a minimal consistent set of
features was obtained over 100 runs.
[0342] i. Statistical Methods
[0343] From the point of view of multivariate statistical analysis,
the problem is one of estimating density functions in
high-dimensional space (and partitioning this space into the
regions of interest). Assuming that the distributions of random
(feature) vectors are known, the theoretically best classifier is
the Bayes classifier because it minimizes the probability of
classification error (K. Fukunaga, "Statistical Pattern
Recognition", 2.sup.nd Ed., Academic Press 1990, p. 3).
Unfortunately the implementation of the Bayes classifier is
difficult because of its complexity, especially when the
dimensionality of the feature space is high. In practice, simpler
parametric classifiers are used. Parametric classifiers are based
on assumptions about the underlying density or discriminant
functions. The most common such classifiers are linear and
quadratic classifiers. In multivariate statistical analysis such
classifiers fall under the heading of discriminant analysis.
Discriminant analysis techniques are closely related to
multivariate linear regression models and generalized linear models
(encompassing logistic and multinomial regression).
[0344] (1) Logistic Regression with a Binomial Response
[0345] (a) Background
[0346] The problem of selecting a set of markers to be used on a
detection panel can be formulated as a logistic regression problem
with a binomial response. The response variable is a factor with
two levels: normal (no cancer) and abnormal (cancer). The
explanatory variables are the marker H-scores.
[0347] The problem of selecting a set of markers to be used on a
cancer discrimination panel can also be formulated as a logistic
regression problem with a binomial response. The response variable
is a factor with two levels: normal (not the cancer of interest)
and abnormal (cancer of interest). The explanatory variables are
the marker H-scores.
[0348] Stepwise variable selection can be used to select a subset
of the original variables (markers) for use in discriminating
between the two classes. This is a computationally expensive
exercise and is best suited to a computer. Several commercial and
public domain software packages--e.g., R, S-plus, and
SAS--implement stepwise logistic regression.
[0349] Two different approaches to feature selection were
investigated based on the stepwise variable selection procedures
found in R and SAS respectively.
[0350] (b) Experimental Data
[0351] The data used for the present analysis consists of the
H-scores for markers 1-17, and 19-28 for the cases examined by
Pathologist 1 and Pathologist 2 and described elsewhere in this
report. In addition, a dummy marker, 18, was added to the data set.
The dummy marker consists of integer values from 0 to 12 selected
at random from a uniform distribution.
[0352] (c) Method 1: Using the R package (version 1.4.1)
[0353] Computerized model fitting procedures generally cannot deal
with missing data. This is the case for the glm (glm stands for
generalized linear model) procedure used in R. Consequently when
fitting a model using glm it was necessary to exclude all the cases
for which there are one or more missing values. When fitting the
initial full model, containing the 27 real markers and the single
dummy marker, this reduces the data set to only 202 cases. With so
few observations it was decided that the best way to perform
variable selection, to train a classifier using the selected
variables, and to assess its performance was to undertake 100
trials on random partitions of the data into train and test
sets.
[0354] (i) Partitioning the Data into Train and Test Sets
[0355] At the start of each trial, the data is partitioned into a
test set and a training set. This is done by randomly choosing 30%
of the abnormals and 30% of the normals to form the test set, and
using the remaining observations to form the training set.
[0356] (ii) Variable (Marker) Selection
[0357] At the start of each trial, the full model, which includes
all of the variables (markers), is fitted to the training data. In
R the logistic regression model is fitted using glm. The code
fragment used is as follows:
TABLE-US-00022 my.model <- Class ~ X1 + X2 + X2 + X3 + X4 + X5 +
X6 + X7 + X8 + X9 + X10 + X11 + X12 + X13 + X14 + X15 + X16 + X17 +
X18 + X19 + X20 + X21 + X22 + X23 + X24 + X25 + X26 + X27 + X28
my.glm <- glm(my.model, family=binomial(link=logit),
data=training.data)
[0358] The procedure stepAIC is then used to perform stepwise
variable selection based on the Akaike Information Criterion (AIC).
This procedure is part of the publicly available MASS library. The
library and the procedure are described in "Modern Applied
Statistics with S-PLUS" (W. N. Venables and B. D. Ripley,
Springer-Verlag, Pathologist 3ew York, 1999). The R code fragment
to do this is as follows:
my.step <- stepAIC(my.glm, direction=both)
[0359] The resulting model is then assessed on the test data. The
code fragment used is as follows:
TABLE-US-00023 probability_is_abnormal <-
predict(my.step,testing.data,type="response")
[0360] The performance of the classifier is recorded in terms of
the actual error rate of misclassification (AER) and the area under
the ROC curve (AUC). After the 100 trials, 100 models and their
associated AERs and AUCs remain. A frequency table is constructed,
recording the number of times each variable made an appearance in
the 100 models. An example is shown in Table 22:
TABLE-US-00024 TABLE 22 Variable 1 2 3 4 5 6 7 8 9 11 12 14 Fre- 2
6 4 1 4 16 100 40 3 10 43 1 quency Variable 15 16 17 18 19 20 22 23
24 25 28 Frequency 4 28 2 1 3 2 2 18 10 84 4
[0361] This table is used to decide which markers to discard.
First, all of the markers that have a frequency less than or equal
to 10 are discarded. Next a cut-off frequency is chosen based on
the frequency of the dummy marker (typically this is 1 or 1.5 times
that of the dummy marker). All markers with a frequency less than
this cut-off value are discarded. The remaining markers, along with
the dummy marker, are then used as the full model for another 100
trials and the pruning process is repeated. If necessary, the
severity of the pruning can be increased to force one or more
markers out of the model. If necessary, the remaining markers can
be used as the full model for yet another 100 trials. Pruning stops
when the desired number of panel members is reached or the average
AUC for the current model is less than that for the preceding
model.
[0362] To illustrate the pruning process consider the table above.
The table was obtained using the detection panel data. The shaded
entries indicate those markers that are retained after pruning.
Another 100 trials is performed using the following full model:
my.model <- Class.about.X6+X7+X8+X12+X16+X18+X23+X25
[0363] Again, a frequency table, Table 23 is constructed:
TABLE-US-00025 TABLE 23 Variable 6 7 8 12 16 18 23 25 Frequency 63
100 51 48 30 47 66 98
[0364] The shaded entries show the markers retained after pruning
(using a cutoff of 47). Another 100 trials is performed using the
following full model:
my.model <- Class.about.X6+X7+X8+X12+X18+X23+X25
[0365] Again, a frequency table, Table 24 is constructed:
TABLE-US-00026 TABLE 24 Variable 6 7 8 12 18 23 25 Frequency 96 100
23 73 3 88 98
[0366] At this point a cut-off of 50 is chosen. The shaded entries
show the remaining markers for use on a 5 member panel. In each
step, the average AUC increases:
94.37%.fwdarw.95.45%.fwdarw.95.78%.
[0367] (iii) Assessing the Performance of the Panel
[0368] To assess the performance of the panel, 100 trials were
performed, as before, but without the stepwise selection procedure.
For each trial, the AUC, sensitivity, and specificity are recorded.
For the detection panel example above, the results are:
TABLE-US-00027 >my.model <- Class ~X7 + X25 + X6 + X23 + X12
Min. 1st Qu. Median Mean 3rd Qu. Max. >summary(AUC) 0.9289
0.9590 0.9615 0.9601 0.9630 0.9630 >summary(sensitivity) 0.8519
0.9630 0.9630 0.9737 1.0000 1.0000 >summary(specificity) 0.8378
0.9730 0.9730 0.9749 1.0000 1.0000
[0369] In summary, the panel has a sensitivity of 97.37% and a
specificity of 97.49%. The area under the ROC is 96.01%.
[0370] (d) Method 2: Using SAS (Version 8.2)
[0371] Logistic regression can be performed in SAS using the
procedure LOGISTIC. When the response variable is a two-level
factor, the procedure fits a binary logit model (equivalent to glm
in R with family-binomial and link=logit). SAS automatically
excludes all of the missing multivariate observations for the model
specified. Unlike R, SAS is able to perform a best subsets variable
selection procedure. The code fragment in SAS needed to do this is
as follows:
TABLE-US-00028 PROC LOGISTIC DATA=WORK.panel; CLASS Class; MODEL
Class = X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17
X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 /SELECTION= SCORE
BEST=28; RUN;
[0372] This procedure is applied to the entire data set. The
parameter BEST=28 directs SAS to find the best 28 single-variable
models, the best 28 two-variable models, the best 28 three-variable
models, up to the best 28 28-variable models.
[0373] (i) Assessing the Performance of the Panels
[0374] The procedure described in method 1 is used to assess the
performance of each of the panels. The following, Table 25, was
generated from the detection panel data. It lists results only for
the two best one-, two-, three-, four-, and five-marker panels.
TABLE-US-00029 TABLE 25 Panel Panel members Sensitivity Specificity
Area under ROC 1 7 94.28% 2 28 80.14% 3 7, 16 95.00% 4 7, 15 94.59%
5 7, 15, 16 95.94% 6 1, 7, 16 95.33% 7 1, 7, 15, 16 95.61% 8 4, 7,
15, 16 95.34% 9 1, 4, 7, 15, 16 95.30% 10 1, 7, 11, 15, 16
95.57%
[0375] (2) Linear Discriminant Analysis
[0376] (a) Background
[0377] The commercial statistical package SPSS has procedures
allowing simple linear discriminant functions to be design and
tested.
[0378] A commonly used method is Fisher's Linear discriminant
function. This finds the hyper-plane in feature space which gives a
good separation of classes. For a two class problem where the class
distributions have different means, but similar multivariate
Gaussian distributions, this classifier gives optimum performance.
The method can be extended heuristically to multi-class problems,
but this was not applied in the study.
[0379] The method is simplistic in its approach but robust to
problems associated with data sets containing a large number of
features (the probes in our case number 27, giving problem for a
data set comprising only some two hundred exemplars (cases)).
[0380] This package has a procedure for identifying the features
which contribute well to the discrimination process. This "stepwise
method" first finds the most discriminating feature. Other features
are then sequentially added and evaluated against the classifier.
Combinations are explored so the final solution may exclude
features initially selected if better combinations are found. The
number of features is gradually increased until a statistical test
shows the remaining features do not contribute reliably to the
classification process.
[0381] An estimate of the performance is gained by using the leave
one out method. This removes one sample from the data set to form
the training set. The left out sample is retained as the test set,
applied to the classifier, and the resulting classification
accumulated in the confusion matrix. The procedure is repeated for
case in the data. This procedure gives an unbiased estimate of
performance, but the estimate will have a high variance.
Method
[0382] In SPSS select the appropriate data set for analysis, select
"Analyze", select "Classify", select "Discriminant . . . ", on the
table select "Fishers method", "leave one out testing" and "use
stepwise method". Enter the diagnosis as the grouping variable and
enter all the features as the independents. Enter "OK" to complete
the analysis. Pre-set values for other parameters were left as
set.
[0383] The analysis output includes a list of the features used in
the analysis, the canonical discriminant function and a confusion
matrix and the correct-classification rate (I-error rate).
[0384] In order to compute an ROC curve the Canonical discriminant
function is applied to the selected features to generate a new
feature. In SPSS use Graphs, ROC to plot this curve
[0385] ii. Hierarchical Methods: Decision Trees
[0386] (1) Background
[0387] Decision tree learning is one of the most widely used and
practical methods for inductive inference. It is a method for
classification that is robust to noisy data and capable of learning
disjunctive expressions (Tom M. Mitchell, "Machine Learning",
McGraw-Hill, New York, N.Y., 1997.)
[0388] The most popular and accessible machine learning package is
"C4.5" the source code of which is published in: (J. Ross Quinlan,
"C4.5: Programs for Machine Learning", Morgan Kaufmann, San Mateo
Calif., 1993).
[0389] When a decision tree is being trained (on training data),
the algorithm decides at each node of the tree which single
attribute of the data to use at this node to best make a decision.
Therefore when the tree is completely constructed, it will have
selected some set of attributes to use and ignored others. In our
application, using decision trees to process measurements gained
from molecular probes, the decision tree has effectively chosen a
panel of probes, and a method of combining the probe scores, which
best explains the classification of the data. To obtain an unbiased
estimate of the panel performance, the resulting tree must be
evaluated on data which was not used in the training. One standard
technique for doing this is cross-validation. A 10-fold
cross-validation was employed.
[0390] Cross-validation is a technique for making the very best use
of limited data. In 10-fold cross-validation the data is randomly
split into 10 nearly-equal sized partitions, taking care to have
approximately the same number of cases in a class across each
partition. Then, the decision tree is trained on partitions 2-9
combined and tested on partition 1, then trained on partitions
1,3-9 combined and tested on partition 2, and so on for 10 trials
rotating the held-out test set through the data once. In this
manner tests are only ever performed on held-out data and so are
unbiased, and all data is tested exactly once so an aggregate error
rate across the whole data set can be computed.
[0391] Trees are usually constructed until they are a very good fit
to the training data, then they are "pruned" back by clipping off
"noisy" branches and leaves. This improves the generalization
ability of the decision tree on unseen data and is essential to
obtain good performance. The C4.5 package includes two methods for
pruning trees first a standard tree pruning algorithm, second a
rule extraction algorithm. In general, the tree based method was
found to give superior results on this data. Therefore, the
rule-based method is not reported.
[0392] (2) Data Preparation
[0393] Data on the response of various probes to normal tissue and
five different cancers (Adenocarcinoma, Large Cell Carcinoma,
Mesothelioma, Small Cell Lung Cancer, and Squamous Cell Carcinoma)
was obtained as described elsewhere. The H-scores for probes 1-28,
and pathologists Pathologist 1 and Pathologist 2 were extracted
from the database and put into a flat data file. For the decision
tree analysis each data point (even by two pathologists on a same
physical slide) was taken to be an independent observation of the
effect of disease on staining. This may slightly positively bias
the performance of classification but should have no effect on
panel selection. [0394] The control categories of Emphysema,
Granulomatous Disease, and Interstitial Lung Disease were grouped
together and called "Normal". [0395] For the detection panel all
the cancers were grouped together and called "Abnormal" making this
a 2-class problem. [0396] For the single discrimination panel, the
Normal cases were removed from the data to form a 5-class problem.
[0397] For the hold-out discrimination panels, each cancer was held
out in turn and the remaining cancers grouped into "Other" to give
a set of five 2-class problems.
[0398] C4.5 requires a ".names" file which describes the data and
the attributes to be included in the analysis. An example names
file for the discrimination panel is, Table 26:
TABLE-US-00030 TABLE 26 C4.5 Names file for MonoGen ZF21 diag data
Adenocarcinoma, Large Cell Carcinoma, Mesothelioma, Small Cell Lung
Cancer, Squamous Cell Carcinoma. | classes P1 continuous. P2
continuous. P3 continuous. P4 continuous. P5 continuous. P7
continuous. P8 continuous. P9 continuous. P10 continuous. P11
continuous. P12 continuous. P13 continuous. P14 continuous. P15
continuous. P16 continuous. P17 continuous. P18 ignore. P19
continuous. P20 continuous. P21 continuous. P22 continuous. P23
continuous. P24 continuous. P25 continuous. P26 continuous. P27
continuous. P28 continuous.
[0399] Probe 18 was missing from the data and was set to "ignore"
in all the designs. Setting attributes to "ignore" in the names
file is an easy and effective way of trimming probes from the
panels and is used in the data analysis.
[0400] (3) Data Analysis
[0401] Ten-fold cross validation was run on each data set using the
"xval.sh" script supplied with C4.5. Standard (default) parameters
for the package were used. Cross validation is a technique
developed for classifier training and testing on small data sets.
It involves randomly splitting the data into N equal sized
partitions. The clasifier is then trained on N-1 partitions
together and tested on the remaining partition. This is repeated N
times.
[0402] Since the decision tree trained in one cross-validation(CV)
trial may differ from the tree obtained in another (different in
both probes selected, and tree coefficients) the number of times
each probe was selected by the tree in 10 trials was computed.
[0403] The first cull of probes was done by setting to ignore any
probe which did not occur in a pruned tree 5 or more times out of
the 10 CV trials.
[0404] Then the cross-validation was repeated with this smaller set
of candidate probes. The second cull of probes was done by setting
to ignore any probe which did not occur in a pruned tree 5 or more
times out of the 10 CV trials. If any further probes dropped out, a
third CV run was done.
[0405] The panels were selected by the various runs, and their
estimated error performance are shown in the results tables. The
panel performance for decision tree analysis is shown below, in
Table 27.
TABLE-US-00031 TABLE 27 Panel Performance - Decision Trees Cancer
Control Detection Panel Cancer 99.42% 0.58% Probes: 3, 7, 19, 25
and 28 Control 17.82% 82.18% Adeno Others Pair-wise Discrimination
Adeno 67.74% 32.26% 4, 6, 14, 19 and 23 Others 11.20% 88.80%
Squamous Others Pair-wise Discrimination Squamous 70.59% 29.41% 3,
6, 17, 19 and 25 Others 4.07% 95.93% Large Cell Others Pair-wise
Discrimination Large Cell 36.36% 63.64% 1, 5, 10, 13, 21, 27 and 28
Others 7.37% 92.63% Mesothelioma Others Pair-wise Discrimination
Mesothelioma 82.05% 17.95% 3, 12 and 16 Others 5.00% 95.00% Small
Cell Others Pair-wise Discrimination Small Cell 69.23% 30.77% 12,
17, 20, 23 and 25 Others 1.49% 98.51% Cancer Control Detection
(without probe 7) Cancer 89.60% 10.40% 6, 10, 16 and 19 Control
3.30% 96.70% Cancer Control Detection (only commercially Cancer
92.80% 7.20% preferred probes) 5, 6, Control 5.49% 94.51% 10, 16,
19 and 23
[0406] An example decision tree structure is shown in below, in
Tables 28 and 29, for discriminating between Small Cell Lung Cancer
and the remaining four types of cancer.
C4.5 Output Format:
TABLE-US-00032 [0407] TABLE 28 P23 <= 3 : | P25 <= 2 : Small
Cell Lung Cancer (18.0) | P25 > 2 : | | P17 <= 5 : Small Cell
Lung Cancer (2.0) | | P17 > 5 : | | | P20 <= 11 : Other (9.0)
| | | P20 > 11 : Small Cell Lung Cancer (2.0) P23 > 3 : | P12
> 7 : Other (120.0) | P12 <= 7 : | | P20 <= 2 : Other
(5.0) | | P20 > 2 : Small Cell Lung Cancer (4.0) Tree saved
Evaluation on training data (160 items): Before Pruning After
Pruning Size Errors Size Errors Estimate 13 0(0.0%) 13 0(0.0%)
(5.2%) <<
TABLE-US-00033 TABLE 29 Pictorial format: ##STR00005##
[0408] The panel performance for stepwise linear discriminant is
shown below, in Table 30:
TABLE-US-00034 TABLE 30 Panel Performance - Stepwise LD Cancer
Control Detection Panel Cancer 92.24% 7.76% 1, 4, 7, 15 and 16
Control 1.16% 98.84% Adeno Others Pair-wise Discrimination Adeno
91.67% 8.33% 4, 5, 14, 19, 20, 25 and 27 Others 5.43% 94.57%
Squamous Others Pair-wise Discrimination Squamous 88.00% 12.00% 1,
2, 3, 24, 25 and 26 Others 6.59% 93.41% Large Cell Others Pair-wise
Discrimination Large Cell 80.95% 19.05% 1 and 7 Others 26.32%
73.68% Mesothelioma Others Pair-wise Discrimination Mesothelioma
96.67% 3.33% 3, 12 and 16 Others 4.65% 95.35% Small Cell Others
Pair-wise Discrimination Small Cell 93.75% 6.25% 12, 19, 22 and 23
Others 5.00% 95.00% Cancer Control Detection (without probe 7)
Cancer 85.34% 14.66% 1, 2, 3, 4, 10, 11, 15, 16, Control 2.33%
97.67% 23, 24, 27 and 28 Cancer Control Detection (only
commercially Cancer 81.20% 18.80% preferred probes) 8, 10, 11,
Control 1.16% 98.84% 19, 23 and 28
[0409] The panel performance for stepwise logistic regression
analysis is shown below, in Table 31:
TABLE-US-00035 TABLE 31 Panel Performance - Stepwise LR Cancer
Control Detection Panel Cancer 97.49% 2.63% 6, 7, 12, 23 and 24
Control 2.51% 97.49% Adeno Others Pair-wise Discrimination Adeno
96.39% 3.61% 14, 19, 20, 25 and 27 Others 12.29% 87.71% Squamous
Others Pair-wise Discrimination Squamous 94.93% 5.07% 3 and 10
Others 35.86% 64.14% Large Cell Others Pair-wise Discrimination
Large Cell 95.11% 4.89% 1, 4, 6, 16 and 21 Others 61.00% 39.00%
Mesothelioma Others Pair-wise Discrimination Mesothelioma 95.07%
4.93% 3, 7, 12 and 16 Others 10.89% 89.11% Small Cell Others
Pair-wise Discrimination Small Cell 98.90% 1.10% 12, 13 and 23
Others 4.00% 96.00% Cancer Control Detection (without probe 7)
Cancer 94.00% 6.00% 1, 10, 19, 23 and 28 Control 5.80% 94.20%
Cancer Control Detection (only commercially Cancer 93.88% 6.12%
preferred probes) 10, 19, Control 6.39% 93.61% 20, 23 and 28
[0410] iii. Neural Networks and Alternative Methods
[0411] Artificial neural networks ANN's are candidate pattern
recognition techniques which could readily be applied to select
features and design classifiers in association with this invention.
However such techniques give little insight to the structure of the
data and the influence of particular probes in the way that LDF
gives. For this reason this class of algorithm was not used in this
study. LDF stands for linear discriminant function, a linear
combination of features whose result is thresholded to determine
the classification.
[0412] This class of techniques includes algorithms such as
Multi-Layer Perceptron MLP, Back-Prop, Kohonen's Self-Organizing
Maps, Learning Vector Quantization, K-nearest neighbors and Genetic
Algorithms.
[0413] iv. Special topics
[0414] (1) Assumptions
[0415] Linear discriminant analysis [0416] Assumes the covariance
matrices for the two classes are equal. [0417] Minimizes the cost
of misclassification only when the two classes are multivariate
normal. [0418] Assumes that the explanatory variables are
continuous rather than categorical (in this study, the H-scores are
categorical while in practice (i.e., in an automated system)
intensity can be measured on a continuous scale).
[0419] Logistic regression (binomial generalized linear models)
[0420] See Venerables and Ripley, chapter 7 ("Modern Applied
Statistics with S-PLUS" (W. N. Venables and B. D. Ripley,
Springer-Verlag, New York, 1999)).
[0421] (2) Marker Rejection (De-Selection)
[0422] Computerized implementations of discriminant analysis and
regression procedures include stepwise variable selection
procedures; e.g., stepAIC in R. These procedures are designed to
select the best subset of variables for use as explanatory
variables. In reality, because of the step-by-step nature of these
procedures, there is no guarantee that the best variables are
selected for prediction (Johnson and Wichern, p. 299). Nevertheless
such procedures do provide the basis for marker selection and
de-selection.
[0423] (3) Pairwise Tests
[0424] Inherent problems in designing multiclass classifiers is
discussed in "Applied Mulitvariate Statistical Analysis", R. A.
Johnson and D. W. Wichern, 2nd Ed, 1988, Prentice-Hall, N.J. This
is motivation for developing several separate two-class classifiers
(discrimination panel).
[0425] (4) Redundancy Consideration in Panel Composition
[0426] "Linear models form the core of classical statistics and are
still the basis of much of statistical practice" "Modern Applied
Statistics with S-PLUS" (W. N. Venables and B. D. Ripley,
Springer-Verlag, New York, 1999. Linear models are the foundation
for the t-test, analysis of variance (ANOVA), regression analysis,
as well as a variety of multivariate methods including discriminant
analysis. Explanatory variables may or may not enter the model as
first-order terms. This is true also of (non-linear) logistic
regression. The logistic regression model is simply a non-linear
transformation of the linear regression model: the dependent
variable is replaced by a log odds ratio (logit). In summary these
statistical methods are based on linear relationships between the
explanatory variables. Consequently, one avenue for seeking
redundancy in panels is to identify highly correlated variables
(markers). It may be possible to replace one marker with the other
in a panel to achieve similar performance.
[0427] Another avenue for seeking redundancy in panels is to
undertake a "best subsets" regression analysis. Given a starting
model with all of the explanatory variables of interest, the aim is
to find the best single-variable regression models, the best
two-variable regression, etc. This methodology is implemented in
the SAS statistical package.
[0428] (5) Use of Weighting Scores
[0429] (a) Commercial and Clinical Considerations
[0430] For many reasons, including strategic and commercial
factors; cost; availability; ease of use, it may be preferred to
encourage the selection of certain probes in a panel and penalize
the selection of others, at the same time trading this off against
panel size or performance.
[0431] (b) Attribute Costing
[0432] Methods for such attribute weighting (in decision trees)
have been proposed in the machine learning literature in other
contexts such as the incorporation of background knowledge (M.
Nunez, "The Use of Background Knowledge", Machine Learning 6:
231-250, 1991.), and the differential cost of obtaining information
from robotic sensors (M. Tan, "Cost-sensitive Learning of
Classification Knowledge and its Applications in Robotics", Machine
Learning. 13: 7-33, 1993.)
[0433] Both of these cost-sensitive algorithms have been
implemented in the literature by minor changes to the standard
machine learning software package known as "C4.5 (J. Ross Quinlan,
"C4.5: programs for machine learning", Morgan Kaufmann, Calif.
1993.) For convenience, this approach was followed to implement the
"EG2" algorithm of Nunez.
[0434] In the C4.5 decision tree construction phase, the algorithm
compares each available attribute to split on and chooses the
single one which maximizes the information gain, G1. In the EG2
algorithm, (2.sup.Gi-1)/(Ci+1) is maximized which incorporates the
cost of information for attribute i, Ci. The vector of weights need
to be set a priori by the user.
[0435] (i) Code Modifications
[0436] The C4.5 source code was modified to implement the economic
generalizer "EG2" algorithm proposed by M. Nunez (The Use of
Background Knowledge, Machine Learning 6: 231-250, 1991.)
[0437] The exact modifications to the C4.5 package are as
follows.
After the following lines in file "R8/Src/contin.c". (J. Ross
Quinlan, "C4.5: programs for machine learning", Morgan Kaufmann,
Calif. 1993)
TABLE-US-00036 ForEach(i, Xp, Lp - 1) { if ( (Val = SplitGain[i] -
ThreshCost) > BestVal ) { BestI = i; BestVal = Val; } }
The new line:
[0438] Bestval =(powf(2.0,
B.estVal)-1.0)/(AttributeCosts[Att]+1.0);
is inserted. Where the vector of attribute costs has been
previously read in from a text file maintained by the user.
[0439] (ii) Experimental Methodology.
[0440] The commercially preferred probes are: 2, 4, 5, 6, 8, 10,
11, 12, 16, 19, 20, 22, 23, 28.
[0441] For the sake of example, suppose the above probes are
commercially preferred due to cost and it is desired to reselect
the detection panel taking this cost into account.
[0442] The modified C4.5 decision tree software was used to give
the commercially preferred probes a penalty of zero and
non-commercially preferred probes a penalty of two. The 10-fold
cross validated panel selection methodology (as described
elsewhere) was run using the modified C4.5 algorithm
[0443] (iii) Results
[0444] The standard decision tree detection panel consists of
probes 3, 7, 19, 25, 28.
[0445] Resulting Panel Members: are 2, 6, 7, 10, 19, 25, 28 which
used only 2 commercially preferred probes, P7 and P25. Note these
probes have been selected by the method in spite of their increased
cost due to their superior performance on this data. The panel is
now larger: 7 probes versus 5 originally. There is no
demonstratable drop in panel performance on this data although the
performance will now be sub-optimal as a trade off against the
reduced cost of probes.
[0446] (iv) Conclusion
[0447] A straightforward way has been established for incorporating
costs of using probes into the panel selection methodology.
[0448] (c) Misclassification costing
[0449] (i) Background
[0450] For many reasons it may be desired to select an optimal
panel bearing in mind that the costs of the different kinds of
classification errors may vary. For example, it may be desired to
select a panel which has an increased sensitivity to one disease
(say Large Cell Carcinoma) and be willing to trade this off against
reduced specificity and sensitivity elsewhere in the confusion
matrix.
[0451] In theory a matrix of misclassification costs (of the same
dimensions as the confusion matrix) to incorporate all the possible
combinations of costs may be needed. In practice, only those costs
which are non unity (the default) are entered.
[0452] The commercial decision tree software See 5. (RuleQuest
Research Pty Ltd, 30 Athena Avenue, St Ives Pathologist 3SW 2075,
Australia. (http://www.rulequest.com)) incorporates this capability
and was used in the following demonstration.
[0453] (ii) Aim
[0454] The standard joint discrimination panel (described
elsewhere) consists of the members: P2, 3, 4, 5, 12, 14, 16, 19,
22, 23, 28. And gives the following estimated confusion matrix:
TABLE-US-00037 classified as (a) (b) (c) (d) (e) 24 4 2 5 2 (a):
class Adenocarcinoma 8 7 3 5 4 (b): class Large Cell Carcinoma 1 1
33 1 4 (c): class Mesothelioma 6 2 1 23 (d): class Small Cell Lung
Cancer 4 4 3 2 24 (e): class Squamous Cell Carcinoma
[0455] The sensitivity of Large Cell Carcinoma is low at 26
percent. If one wished to increase this sensitivity in a newly
designed panel, the following method may be employed.
[0456] (iii) Methodology
[0457] The following costs file was generated:
TABLE-US-00038 costs file for ZF21Discrim Increase sensitivity for
"Large Cell Carcinoma" Mesothelioma, Large Cell Carcinoma: 10
Adenocarcinoma, Large Cell Carcinoma: 10 Mesothelioma, Large Cell
Carcinoma: 10 Small Cell Lung Cancer, Large Cell Carcinoma: 10
Squamous Cell Carcinoma, Large Cell Carcinoma: 10
[0458] This file upweights the misclassification of Large Cell
Carcinoma as any of the other cancers by a factor of 10. This will
tend to increase the sensitivity of detection in this class (with
reduced performance elsewhere) but no weighting can ensure perfect
classification.
[0459] The standard decision tree panel selection methodology was
applied (using See5 instead of C4.5).
[0460] (iv) Results
[0461] The new panel members are: P2, 3, 4, 5, 6, 9, 12, 14, 16,
17, 25, 28. With an estimated performance of:
TABLE-US-00039 classified as (a) (b) (c) (d) (e) Cancer 20 13 1 1 2
(a): class Adenocarcinoma Carcinoma 3 13 3 2 6 (b): class Large
Cell Carcinoma 1 9 27 2 1 (c): class Mesothelioma 2 9 21 (d): class
Small Cell Lung 1 15 2 1 18 (e): class Squamous Cell
[0462] The above demonstrates that the estimated sensitivity of
Large Cell Carcinoma has now increased to 48%.
[0463] (v) Conclusion
[0464] A straightforward way has been demonstrated for
incorporating the differential costs of misclassification into the
panel selection methodology.
[0465] d. Performance Metrics
[0466] Outputs provided by the analysis indicating the estimated
performance of each method include:
[0467] i. ROC Analyses
[0468] Receiver Operating Characteristic (ROC) curves show the
estimated percentage (or per unit probability) of false positive
and false negative scores for different threshold levels in the
classifier. An indifferent classifier, unable to discriminate
better than random choice, would present a ROC curve with equal
true and false readings. The area under this curve would be 50%
(0.5 probability).
[0469] Area Under the Curve (AUC) is often used as an overall
estimate of classifier performance and most commercial discriminant
function packages compute this figure. A perfect classifier would
have 100% Area Under the Curve, a useless classifier would have an
AUC near 50% (0.5).
[0470] ii. Confusion Matrices: Counts and Percentages
[0471] Confusion matrices show how data from the test set was
classified. For pair wise tests these are counts of true positive,
false positive, true negative or false negative scores. These may
be shown as actual counts or as percentages. For the multi-way
Panel, which attempts to give a unique diagnosis with one panel
only, the confusion matrix would show counts for each correct
classification. For instance, each time Small Cell carcinoma is
detected as such it would be entered in one diagonal of the matrix.
Incorrect scores; for instance how often a small cell carcinoma is
incorrectly identified as squamous cell cancer would be entered in
the appropriate off-diagonal element of the matrix. Error Rates are
used to summarize data in the confusion matrix as the sum of all
false classifications divided by the total number of
classifications made, expressed as a percentage.
[0472] iii. Sensitivity and Specificity
[0473] Specificity refers to the extent to which any definition
excludes invalid cases. If a definition has poor specificity, it is
high in false positives. This means that it labels individuals as
having a disorder when there is really no disorder present.
Sensitivity refers to the extent to which any definition includes
all valid cases. If a definition has poor sensitivity, it is high
in false negatives (individuals who have a disorder present are
falsely being diagnosed as not having one).
[0474] 3. Data Analysis and Results
[0475] a. Sample Size and Variability [0476] Of the 354 cases in
the combined Pathologist 1 and Pathologist 2 data set, only 202
cases possessed an H score for every marker (variable or feature).
[0477] The small number of complete observations and the large
number of variables leads to estimation problems (curse of
dimensionality). Hence it is necessary to prune severely back the
number of variables used to build a classifier. [0478] Due to the
small number of observations it is not prudent to divide the data
into separate training and testing sets (necessary for the robust
estimation of classifier performance). For this reason, it was
necessary to use resampling methods (such as cross-validation and
multiple random trials). [0479] The design of a multiclass
classifier for cancer discrimination is difficult because there are
so few observations for each type of cancer.
[0480] b. De-Selected Markers
[0481] Markers were de-selected using the methodology described
above. Markers that were de-selected are represented by
non-selection in the panels.
[0482] c. Detection Panel(s) Composition
[0483] i. Selected Marker Probes
[0484] The selected marker probes for all three methods are
summarized in FIG. 5.
[0485] ii. Minimum Selected Marker Set
[0486] For the detection panel it is clear that probe 7 delivered
the best detection performance for a single marker. Combinations of
probes were analyzed to see if a reliable panel could be obtained
with more probes.
[0487] (1) Method
[0488] The Logistic Regression method allows best subsets to be
ranked in terms of a performance measure (Fisher'score). This
analysis was used to select the combinations from 1 through 5
probes. Fishers linear discriminant function and logit models
(logistic regression) were used to illustrate the performance of
these combinations. Data shown above.
[0489] (2) Conclusions
[0490] Probe 7 performs well on its own as a classifier; however, a
drawback to using probe 7 alone is that probe 7 has a high false
negative score. The best performance using Fishers linear
discriminant function as a classifier was with probes 7 and 16. The
variability of results amongst panels using other combinations
suggests the noise added by more features is outweighing any
potential to improve classification scores. The small number of
incorrectly scored samples gives a poor representation of the
statistics of these rarer events. A classifier designed with a
larger number of cases may allow a better classifier to be
designed. Techniques to select best combinations of probes using
different classifiers may produce a different best panel, depending
on the structure of the data.
[0491] iii. Supplemental Markers
[0492] It is shown that panels can be designed to suit the
availability of different probes. Different methodologies can be
used for selecting these subsets: Decision Trees, Logistic
Regression, and Linear Discriminant Functions. Data are shown
above.
Method
[0493] Using SPSS a Fisher's Linear Discriminant function was
applied to the scores obtained from the panel in which constrains
were applied due to access constraints. For example, all of the
probes come from one vendor. Again, the stepwise option was
selected to find the best combination of features. Performance was
estimated using the Leave-One-Out cross validation test.
[0494] iv. Alternative Markers: Biological Mechanisms of Action
(Functionally Equivalent Markers)
[0495] A person of ordinary skill in the art is able to determine
functionally equivalent markers. The functional behaviors of the
markers used in the panel are described throughout this
document.
[0496] v. Marker Localization
[0497] The localizations of the various markers used in this study
are described elsewhere in this document.
[0498] vi. Panel Performance
[0499] The performance of the three methods is shown above.
[0500] vii. Limitations on Interpretation of Panel Performance
[0501] Due to small data set and the need to employ resampling
methods, there is the danger that the classifiers have been
over-trained (made to fit the data too closely). [0502] The panel
performance using cytology specimens is difficult to forecast
accurately since it is not clear whether sputum cytology samples
will contain adequate numbers of cells that are representative of
the cells analyzed in the histological validation studies.
Nevertheless, given an adequate cellular sample size, one would
expect the optimized panel to behave similarly with cytological
specimens.
[0503] d. Discriminant Panel Composition
[0504] i. A Single 5-Way Panel for All Cancers
[0505] Of the three analysis techniques, only a decision tree is
amenable to a single 5-way panel. A single decision tree was
therefore constructed to simultaneously classify all types of lung
cancer. The panel members are shown FIG. 5. The panel performance
is shown above in the panel performance tables.
[0506] ii. Panels for Discriminating a Single Type of Lung Cancer
Against All Others
[0507] Linear discriminant functions are not well suited to
performing simultaneous multi-class discrimination. The performance
of five separate classifiers, each designed separately to
discriminate one of the cancers from a pooled set of all the
cancers, was analyzed. Such combinations have the potential to
classify none of the cases as having one of the candidate cancers,
or classify a single case as having two or more of the candidate
cancers. This has a potential advantage in identifying inconsistent
cases for further review.
[0508] It has been seen that the overall error rate of a single
discriminant panel for all cancer types has a fairly high error
rate (a five way classifier). In the panel performance data shown
above, the performance of five pair-wise classifiers, each designed
to identify one cancer from the four other possible cancers is
shown. This approach is amenable to analysis using Decision Trees,
and Linear Discriminant functions. The technique has the potential
to deliver an ambiguous finding when applied, giving two or more
diagnoses for a single patient, suggesting further clinical
investigation. The technique has the potential to deliver no
finding, again suggesting further investigation (perhaps a re-test
with the detection panel).
[0509] iii. Panels to Account for Possibility of False Positive
Cases from Detection Panels
[0510] A further panel can be trained to discriminate among the
false positive cases (from the detection panel) and the five cancer
types. This involves selecting those individual cases from the
detection panel that were incorrectly classified as abnormal. This
trains a dedicated classifier on the `harder` problem of detecting
these `special` cases. However, while this is a theoretically sound
task, the data set only yielded four of these cases and the
population was deemed to be under-represented for analysis.
[0511] iv. Selected Markers
[0512] The selected marker probes for all three methods are
summarized in FIG. 5.
[0513] v. Minimum Selected Marker Set
[0514] This topic is addressed below under "Robustness of Approach
Demonstrated by Similar Results Using Different Methods."
[0515] vi. Supplemental Markers
[0516] This topic is addressed below under "Robustness of Approach
Demonstrated by Similar Results Using Different Methods."
[0517] vii. Alternative Markers: Biological Mechanisms of
Action
[0518] A person of ordinary skill in the art is able to determine
functionally equivalent markers. The functional behaviors of the
markers used in the panel are described throughout this
document.
[0519] viii. Marker Localization
[0520] The localization of the various markers used in this study
are described throughout this document.
[0521] ix. Panel Performance
[0522] The performance of the three methods is summarized in FIG.
5.
[0523] e. Effect of Weighting Parameters
[0524] In addition to user-supplied weighting criteria for markers
and also for disease states (classes) as discussed earlier, one can
also use a binary weighting scheme. For example, if all non-DAKO
supplied probes are weighted "0" and all DAKO-supplied probes are
weighted "1", then the optimized panel will contain only
DAKO-supplied probes. This is an important product design
capability for any vendor who intends to develop and market
molecular diagnostic panel kits using only their supplies.
[0525] f. Effect of Using Other (Non H-Score) Objective Scoring
Parameters
[0526] i. Background
[0527] The Pathology Review sheet contains a set of boxes as
follows, in Table 32:
TABLE-US-00040 TABLE 32 Intensity None Weak Moderate Intense 0-5% 0
0 0 0 6-25% 1 1 1 1 26-50% 2 2 2 2 51-75% 3 3 3 3 >75% 4 4 4
4
[0528] The standard scoring system uses the "H score" which is
obtained by grading the intensity as: none=0, weak=1, moderate=2,
intense=3, and the percentage cells as: 0-5%=0, 6-25%=1, 26-50%=2,
51-75%=3, >75%=4, and then multiplying the two grades together.
For example, 50% weakly stained plus 50% moderate stained would
score 10=3.
[0529] ii. Method
[0530] An alternative scoring method was analyzed in which the
response was divided into low, medium and high as follows:
(a) if more than 50% of cells had moderate or above stain HIGH (b)
if more than 50% of cells had no stain LOW (c) otherwise MEDIUM
[0531] The decision tree detection panel selection methodology was
repeated using this 3-level factor instead of H-score. This caused
the tree to split into 3 branches at each node, if required.
[0532] iii. Results
[0533] The panel selected was: Probes 3, 7, 10, 11, 16, 19, 20,
28
[0534] With an estimated performance of:
TABLE-US-00041 Classified as (a) (b) Control (a) 79 22 Specificity
= 78% Cancer (b) 24 149 Sensitivity = 86%
This should be compared to the reference performance with H-scores
of:
TABLE-US-00042 Classified as (a) (b) Control (a) 85 6 Specificity =
93% Cancer (b) 5 120 Sensitivity = 96%
[0535] iv. Conclusions [0536] There is a substantial loss of
performance (larger panels, lower sensitivity and lower specificity
when the proposed alternative scoring system is used. [0537]
Treating the H-score as a continuous variable (in the range 0 to
12) seems to be near optimal for panel selection on the data
examined. [0538] The many other possible scoring systems have not
been examined, but may be feasible and applicable to the
experimentally tested panel design and development methodology.
[0539] 4. Lung Cancer Detection and Discrimination Panels
[0540] Listed below are exemplary lung cancer detection and
discrimination panels determined by the above illustrative example.
It is noted that although the panels listed below recite specific
probes, each specific probe may be substituted by a correlate probe
or a functionally related probe.
Detection (No Constraints)
[0541] anti-Cyclin A combined with one or more additional probes
[0542] anti-Cyclin A, anti-human epithelial related antigen
(MOC-31) [0543] anti-Cyclin A, anti-ER-related P29 [0544]
anti-Cyclin A, anti-mature surfactant apoprotein B [0545]
anti-Cyclin A, anti-human epithelial related antigen (MOC-31),
anti-VEGF [0546] anti-Cyclin A, anti-human epithelial related
antigen (MOC-31), anti-mature surfactant apoprotein B [0547]
anti-Cyclin A, anti-mature surfactant apoprotein B, anti-human
epithelial related antigen (MOC-31), anti-VEGF [0548] anti-Cyclin
A, anti-mature surfactant apoprotein B, anti-human epithelial
related antigen (MOC-31), anti-surfactant apoprotein A [0549]
anti-Cyclin A, anti-mature surfactant apoprotein B, anti-human
epithelial related antigen (MOC-31), anti-VEGF, anti-surfactant
apoprotein A [0550] anti-Cyclin A, anti-mature surfactant
apoprotein B, anti-human epithelial related antigen (MOC-31),
anti-VEGF, anti-Cyclin D1 [0551] anti-Cyclin A, anti-human
epithelial related antigen (MOC-31) combined with one or more
additional probes [0552] anti-Cyclin A, anti-ER-related P29
combined with one or more additional probes [0553] anti-Cyclin A,
anti-mature surfactant apoprotein B combined with one or more
additional probes [0554] anti-Cyclin A, anti-human epithelial
related antigen (MOC-31), anti-VEGF combined with one or more
additional probes [0555] anti-Cyclin A, anti-human epithelial
related antigen (MOC-31), anti-mature surfactant apoprotein B
combined with one or more additional probes [0556] anti-Cyclin A,
anti-mature surfactant apoprotein B, anti-human epithelial related
antigen (MOC-31), anti-VEGF combined with one or more additional
probes [0557] anti-Cyclin A, anti-mature surfactant apoprotein B,
anti-human epithelial related antigen (MOC-31), anti-surfactant
apoprotein A combined with one or more additional probes [0558]
anti-Cyclin A, anti-mature surfactant apoprotein B, anti-human
epithelial related antigen (MOC-31), anti-VEGF, anti-surfactant
apoprotein A combined with one or more additional probes [0559]
anti-Cyclin A, anti-mature surfactant apoprotein B, anti-human
epithelial related antigen (MOC-31), anti-VEGF, anti-Cyclin D1
combined with one or more additional probes
Detection (W/O Anti-Cyclin A)
[0559] [0560] anti-Ki-67 combined with one or more additional
probes. [0561] anti-Ki-67 combined with any one probe selected from
the group consisting of anti-VEGF, anti-human epithelial related
antigen (MOC-31), anti-TTF-1, anti-EGFR, anti-proliferating cell
nuclear antigen and anti-mature surfactant apoprotein B. [0562]
anti-Ki-67 combined with any two probes selected from the group
consisting of anti-VEGF, anti-human epithelial related antigen
(MOC-31), anti-TTF-1, anti-EGFR, anti-proliferating cell nuclear
antigen and anti-mature surfactant apoprotein B. [0563] anti-Ki-67
combined with any three probes selected from the group consisting
of anti-VEGF, anti-human epithelial related antigen (MOC-31),
anti-TTF-1, anti-EGFR, anti-proliferating cell nuclear antigen and
anti-mature surfactant apoprotein B. [0564] anti-Ki-67 combined
with any four probes selected from the group consisting of
anti-VEGF, anti-human epithelial related antigen (MOC-31),
anti-TTF-1, anti-EGFR, anti-proliferating cell nuclear antigen and
anti-mature surfactant apoprotein B. [0565] anti-Ki-67 combined
with any five probes selected from the group consisting of
anti-VEGF, anti-human epithelial related antigen (MOC-31),
anti-TTF-1, anti-EGFR, [0566] anti-proliferating cell nuclear
antigen and anti-mature surfactant apoprotein B. [0567] anti-Ki-67,
anti-VEGF, anti-human epithelial related antigen (MOC-31),
anti-TTF-1, anti-EGFR, anti-proliferating cell nuclear antigen and
anti-mature surfactant apoprotein B [0568] anti-Ki-67 combined with
any one probe selected from the group consisting of anti-VEGF,
anti-human epithelial related antigen (MOC-31), anti-TTF-1,
anti-EGFR, anti-proliferating cell nuclear antigen and anti-mature
surfactant apoprotein B, and with one or more additional probes.
[0569] anti-Ki-67 combined with any two probes selected from the
group consisting of anti-VEGF, anti-human epithelial related
antigen (MOC-31), anti-TTF-1, anti-EGFR, anti-proliferating cell
nuclear antigen and anti-mature surfactant apoprotein B, and with
one or more additional probes. [0570] anti-Ki-67 combined with any
three probes selected from the group consisting of anti-VEGF,
anti-human epithelial related antigen (MOC-31), anti-TTF-1,
anti-EGFR, anti-proliferating cell nuclear antigen and anti-mature
surfactant apoprotein B, and with one or more additional probes.
[0571] anti-Ki-67 combined with any four probes selected from the
group consisting of anti-VEGF, anti-human epithelial related
antigen (MOC-31), anti-TTF-1, anti-EGFR, anti-proliferating cell
nuclear antigen and anti-mature surfactant apoprotein B, and with
one or more additional probes.
[0572] anti-Ki-67 combined with any five probes selected from the
group consisting of anti-VEGF, anti-human epithelial related
antigen (MOC-31), anti-TTF-1, anti-EGFR, anti-proliferating cell
nuclear antigen and anti-mature surfactant apoprotein B, and with
one or more additional probes.
[0573] anti-Ki-67, anti-VEGF, anti-human epithelial related antigen
(MOC-31), anti-TTF-1, anti-EGFR, anti-proliferating cell nuclear
antigen, anti-mature surfactant apoprotein B and one or more
additional probes.
Detection with Commercially Preferred Probes [0574] anti-Ki-67
combined with one or more additional probes. [0575] anti-TTF-1
combined with one or more additional probes. [0576] anti-EGFR
combined with one or more additional probes. [0577]
anti-proliferating cell nuclear antigen combined with one or more
additional probes. [0578] two probes selected from the group
consisting of anti-Ki-67, anti-TTF-1, anti-EGFR and
anti-proliferating cell nuclear antigen. [0579] three probes
selected from the group consisting of anti-Ki-67, anti-TTF-1,
anti-EGFR and anti-proliferating cell nuclear antigen. [0580]
anti-Ki-67, anti-TTF-1, anti-EGFR and anti-proliferating cell
nuclear antigen [0581] two probes selected from the group
consisting of anti-Ki-67, anti-TTF-1, anti-EGFR and
anti-proliferating cell nuclear antigen, and one or more additional
probes. [0582] three probes selected from the group consisting of
anti-Ki-67, anti-TTF-1, anti-EGFR and anti-proliferating cell
nuclear antigen, and one or more additional probes. [0583]
anti-Ki-67, anti-TTF-1, anti-EGFR, anti-proliferating cell nuclear
antigen, and one or more additional probes.
Discrimination Between Adenocarcinoma and Other Lung Cancers
[0583] [0584] anti-mucin 1 and anti-TTF-1 [0585] anti-mucin 1 and
anti-TTF-1 combined with any one probe selected from the group
consisting of anti-VEGF, anti-surfactant apoprotein A, anti-BCL2,
anti-ER-related P29 and anti-Glut 3 [0586] anti-mucin 1 and
anti-TTF-1 combined with and two probes selected from the group
consisting of anti-VEGF, anti-surfactant apoprotein A, anti-BCL2,
anti-ER-related P29 and anti-Glut 3 [0587] anti-mucin 1 and
anti-TTF-1 combined with any three probes selected from the group
consisting of anti-VEGF, anti-surfactant apoprotein A, anti-BCL2,
anti-ER-related P29 and anti-Glut 3 [0588] anti-mucin 1 and
anti-TTF-1 combined with any four probes selected from the group
consisting of anti-VEGF, anti-surfactant apoprotein A, anti-BCL2,
anti-ER-related P29 and anti-Glut 3 [0589] anti-VEGF,
anti-surfactant apoprotein A, anti-mucin 1, anti-TTF-1, anti-BCL2,
anti-ER-related P29 and anti-Glut 3 [0590] anti-mucin 1, anti-TTF-1
and one or more additional probes [0591] anti-mucin 1 and
anti-TTF-1 combined with any one probe selected from the group
consisting of anti-VEGF, anti-surfactant apoprotein A, anti-BCL2,
anti-ER-related P29 and anti-Glut 3, and with one or more
additional probes [0592] anti-mucin 1 and anti-TTF-1 combined with
and two probes selected from the group consisting of anti-VEGF,
anti-surfactant apoprotein A, anti-BCL2, anti-ER-related P29 and
anti-Glut 3, and with one or more additional probes [0593]
anti-mucin 1 and anti-TTF-1 combined with any three probes selected
from the group consisting of anti-VEGF, anti-surfactant apoprotein
A, anti-BCL2, anti-ER-related P29 and anti-Glut 3, and with one or
more additional probes anti-mucin 1 and anti-TTF-1 combined with
any four probes selected from the group consisting of anti-VEGF,
anti-surfactant apoprotein A, anti-BCL2, anti-ER-related P29 and
anti-Glut 3, and with one or more additional probes [0594]
anti-VEGF, anti-surfactant apoprotein A, anti-mucin 1, anti-TTF-1,
anti-BCL2, anti-ER-related P29, anti-Glut 3 and one or more
additional probes
Discrimination Between Squamous Cell Carcinoma and Other Lung
Cancers
[0594] [0595] anti-CD44v6 combined with one or more additional
probes [0596] anti-CD44v6 combined with any one probe selected from
the group consisting of anti-VEGF, anti-thrombomodulin, anti-Glut
1, anti-ER-related P29 and anti-melanoma-associated antigen 3
[0597] anti-CD44v6 combined with any two probes selected from the
group consisting of anti-VEGF, anti-thrombomodulin, anti-Glut 1,
anti-ER-related P29 and anti-melanoma-associated antigen 3 [0598]
anti-CD44v6 combined with any three probes selected from the group
consisting of anti-VEGF, anti-thrombomodulin, anti-Glut 1,
anti-ER-related P29 and anti-melanoma-associated antigen 3 [0599]
anti-CD44v6 combined with any four probes selected from the group
consisting of anti-VEGF, anti-thrombomodulin, anti-Glut 1,
anti-ER-related P29 and anti-melanoma-associated antigen 3 [0600]
anti-CD44v6, anti-VEGF, anti-thrombomodulin, anti-Glut 1,
anti-ER-related P29 and anti-melanoma-associated antigen 3 [0601]
anti-CD44v6 combined with any one probe selected from the group
consisting of anti-VEGF, anti-thrombomodulin, anti-Glut 1,
anti-ER-related P29 and anti-melanoma-associated antigen 3, and
with one or more additional probes [0602] anti-CD44v6 combined with
any two probes selected from the group consisting of anti-VEGF,
anti-thrombomodulin, anti-Glut 1, anti-ER-related P29 and
anti-melanoma-associated antigen 3, and with one or more additional
probes [0603] anti-CD44v6 combined with any three probes selected
from the group consisting of anti-VEGF, anti-thrombomodulin,
anti-Glut 1, anti-ER-related P29 and anti-melanoma-associated
antigen 3, and with one or more additional probes [0604]
anti-CD44v6 combined with any four probes selected from the group
consisting of anti-VEGF, anti-thrombomodulin, anti-Glut 1,
anti-ER-related P29 and anti-melanoma-associated antigen 3, and
with one or more additional probes [0605] anti-CD44v6, anti-VEGF,
anti-thrombomodulin, anti-Glut 1, anti-ER-related P29,
anti-melanoma-associated antigen 3 and one or more additional
probes
Discrimination Between Large Cell Carcinoma and Other Lung
Cancers
[0605] [0606] anti-VEGF combined with one or more additional
probes. [0607] anti-VEGF and anti-p120 [0608] anti-VEGF and
anti-Glut 3 [0609] anti-VEGF, anti-p120 and anti-Cyclin A [0610]
anti-VEGF, anti-p120 and one or more additional probes [0611]
anti-VEGF, anti-Glut 3 and one or more additional probes [0612]
anti-VEGF, anti-p120, anti-Cyclin A and one or more additional
probes
Discrimination Between Mesothelioma and Other Lung Cancers
[0612] [0613] anti-CD44v6 combined with one or more additional
probes. [0614] anti-proliferating cell nuclear antigen combined
with one or more additional probes. [0615] anti-human epithelial
related antigen (MOC-31) combined with one or more additional
probes. [0616] two probes selected from the group consisting of
anti-CD44v6, anti-proliferating cell nuclear antigen and anti-human
epithelial related antigen (MOC-31), combined with one or more
additional probes [0617] anti-CD44v6, anti-proliferating cell
nuclear antigen, anti-human epithelial related antigen (MOC-31) and
one or more additional probes.
Discrimination Between Small Cell and Other Lung Cancers
[0617] [0618] anti-proliferating cell nuclear antigen combined with
one or more additional probes. [0619] anti-BCL2 combined with one
or more additional probes. [0620] anti-EGFR combined with one or
more additional probes. [0621] two probes selected from the group
consisting of anti-proliferating cell nuclear antigen, anti-BCL2
and anti-EGFR [0622] anti-proliferating cell nuclear antigen,
anti-BCL2, anti-EGFR [0623] two probes selected from the group
consisting of anti-proliferating cell nuclear antigen, anti-BCL2
and anti-EGFR, combined with one or more additional probes [0624]
anti-proliferating cell nuclear antigen, anti-BCL2, anti-EGFR and
one or more additional probes
Simultaneous Discrimination of Adenocarcinoma, Squamous Cell
Carcinoma, Large Cell Carcinoma, Mesothelioma and Small Cell
Carcinoma
[0624] [0625] two or more probes selected from anti-VEGF,
anti-thrombomodulin, anti-CD44v6, anti-surfactant apoprotein A,
anti-proliferating cell nuclear antigen, anti-mucin 1, anti-human
epithelial related antigen (MOC-31), anti-TTF-1, anti-N-cadherin,
anti-EGFR and anti-proliferating cell nuclear antigen [0626]
anti-VEGF, anti-thrombomodulin, anti-CD44v6, anti-surfactant
apoprotein A, anti-proliferating cell nuclear antigen, anti-mucin
1, anti-human epithelial related antigen (MOC-31), anti-TTF-1,
anti-N-cadherin, anti-EGFR and anti-proliferating cell nuclear
antigen [0627] two or more probes selected from anti-VEGF,
anti-thrombomodulin, anti-CD44v6, anti-surfactant apoprotein A,
anti-proliferating cell nuclear antigen, anti-mucin 1, anti-human
epithelial related antigen (MOC-31), anti-TTF-1, anti-N-cadherin,
anti-EGFR and anti-proliferating cell nuclear antigen, combined
with one or more additional probes [0628] anti-VEGF,
anti-thrombomodulin, anti-CD44v6, anti-surfactant apoprotein A,
anti-proliferating cell nuclear antigen, anti-mucin 1, anti-human
epithelial related antigen (MOC-31), anti-TTF-1, anti-N-cadherin,
anti-EGFR and anti-proliferating cell nuclear antigen, combined
with one or more additional probes
[0629] 5. Conclusions
[0630] a. Validity of Panel Approach to Molecular Diagnostics
[0631] i. Non-Intuitive Solutions
[0632] Histograms were plotted (PathologistData.xls, worksheet:
Histograms) showing the distribution of marker scores for each
probe for Control vs. Cancer. It is clear from these histograms
that an intuitive selection of probes for specific panels is
certainly not obvious and the invention described does allow
effective combinations to be found in the absence of an obvious
method.
[0633] ii. Optimization for Varied Product Applications
[0634] iii. Robustness of Approach Demonstrated by Similar Results
Using Different Methods
[0635] Detailed scrutiny of the results obtained by the various
analyses in the body of this report, and as summarized in the
tables and figures, shows the following findings.
[0636] 1. Careful scrutiny of the performance of individual probes
does not make apparent probe combinations that might perform better
than any one probe alone.
[0637] 2. All three classification methodologies evaluated hone in
on similar sets of features. The small differences can be
attributed to the data structure that may favor one classifier over
another.
[0638] 3. All the classifiers designed with one of these methods
were shown to give good performance when tested on data from an
independent pathologist, unseen during the design process. This
gives high confidence in the invention.
[0639] 4. A detection panel based on probe 7 alone gives a high
performance.
[0640] 5. If probe 7 is combined with probe 16 or 25 then a better
performance is obtained.
[0641] 6. While combinations of other probes with probe 7 appear to
improve performance further, the number of extra cases captured is
so low that they may be unrepresentative and the classifier so
designed may not generalize.
[0642] 7. The performance of panels selected from probes excluding
probe 7 provided some discrimination, good enough in comparison
with current practice using human screening, but perhaps not good
enough for an automated cytometer in tomorrow's clinical diagnostic
cytology world (see FIG. 6).
[0643] 8. Other combinations of probes can provide a useful, but
lesser, performance.
[0644] 9. If some probes become unavailable this invention allows
the selection of other combinations of probes. This was illustrated
by classifier designs based on a commercially preferred set of
probes only. See FIG. 7.
[0645] 10. The invention allows a weighting to be applied against
costly probes. Rather than totally excluding them from the analysis
this allows their inclusion in the panel if their contribution is
important.
[0646] 11. The invention allows the design of single lung cancer
type specific discrimination panels that can discriminate one type
of lung cancer from among all other cancers.
[0647] 12. Analysis of the performance of a single panel to
classify five cancers showed discrimination was possible but the
overall error rate was worse than a set of five panels each
designed to discriminate one of the cancers from the others.
[0648] 13. A very useful discrimination was obtained with the
combination of five two way classifiers.
[0649] 14. Common sets of probes were selected by the three
classification methodologies for the five discrimination panels,
again giving confidence in this result.
[0650] 15. Probes for isolating cases of Adenocarcinoma are 1, 14,
19, 20, 25, and 27.
[0651] 16. Probes for isolating cases of Squamous Cell cancer are
1, 2, 3, 24, 25, and 26.
[0652] 17. Probes for isolating cases of Large Cell cancer are 1
and 7 or 1, and 21.
[0653] 18. Probes for isolating cases of Mesothelioma are 3, 12,
and 16.
[0654] 19. Probes for isolating cases of Small Cell cancer are 12,
20, and 23.
[0655] 20. Probes for recognizing all cancers simultaneously are 1,
2, 3, 4, 12, 14, 19, 22, 23, and 28.
[0656] 21. An advantage of using the multiple pair-wise panels as
defined by this invention is that doubtful cases may not score on
any of the five panels, also confusing cases may show on two or
more panels. Such anomalous reports would alert the cytologist that
further analysis is indicated.
[0657] iv. Risk Management Study
[0658] All the tests applied in this study were statistical in
nature. There is a risk that probes selected on the basis of small
improvements in performance will have statistical variations when
tested on new data. To give confidence in the results, the best
classifier emerging from the Linear Discriminant analysis on the
Pathologist 1 and Pathologist 2 data was tested. It should be
remembered that the Pathologist 3 data was statistically different
from the Pathologist 1 and Pathologist 2 data, so if good
performances are obtained when tests using the Pathologist 3 data,
then this would be encouraging indeed.
[0659] (1) Report on Testing with Unseen Data--Detection Panel
[0660] (a) Method
[0661] In the Section titled "Detection Panel(s) Composition"
above, we showed that good classification is obtained with features
7 and 16. Using SPSS all the Pathologist 3 data that reported H
scores for both 7 and 16 was selected. Then, using Transform and
Compute, the canonical discrimination function was generated as a
new feature. The performance of this feature alone was then
tested.
[0662] (b) Results
[0663] These are the results of testing the classifier designed on
Pathologist 1 and Pathologist 2 data and testing on Pathologist 3
data. The classifier was designed using the linear discriminant
function on probes 7 and 16. The Canonical Pathologist 2 function
was =0.965*Probe7-0.298*Probe16.
TABLE-US-00043 Classification Results on Pathologist 3 data using
probes 7 and 16 Predicted Group Diagnosis Membership (UCLA) 0 1
Total Original Count 0 20 1 21 1 6 41 47 % 0 95.2 4.8 100.0 1 12.8
87.2 100.0 Cross-validated Count 0 20 1 21 1 6 41 47 % 0 95.2 4.8
100.0 1 12.8 87.2 100.0 a Cross validation is done only for those
cases in the analysis. In cross validation, each case is classified
by the functions derived from all cases other than that case. b
89.7% of original grouped cases correctly classified. c 89.7% of
cross-validated grouped cases correctly classified.
[0664] This is better than classifying the Pathologist 3 data on
probe 7 only show as follows
TABLE-US-00044 Classification Results on Pathologist 3 data using
probe 7 only Predicted Group Diagnosis Membership (UCLA) 0 1 Total
Original Count 0 20 1 21 1 8 39 47 % 0 95.2 4.8 100.0 1 17.0 83.0
100.0 Cross-validated Count 0 20 1 21 1 8 39 47 % 0 95.2 4.8 100.0
1 17.0 83.0 100.0 a Cross validation is done only for those cases
in the analysis. In cross validation, each case is classified by
the functions derived from all cases other than that case. b 86.8%
of original grouped cases correctly classified. c 86.8% of
cross-validated grouped cases correctly classified.
[0665] (c) Conclusion
[0666] This gives confidence that the two-probe classifier based on
7 and 16 is better than alone
[0667] (2) Report on Testing with Unseen Data--Discrimination
Panel
[0668] (a) Background
[0669] Reported below is the performance of the classifier designed
with Pathologist 1 and Pathologist 2 data using LDF and tested with
the unseen Pathologist 3 data. The numbers of cases at the design
stage was relatively small and the numbers in the test data are
also small, so a good degree of variability can be expected between
performance on the first and second set.
[0670] (b) Method
[0671] In SPSS, the canonical discrimination functions derived in
the section titled "Pattern recognition", were built and tested on
Pathologist 3 data for all five classes of cancer
[0672] (c) Results
[0673] Mesothelioma
LDF=probe3sc*0.385-probe12s*0.317+probe16s*1.006
TABLE-US-00045 Classification Results Predicted Group Meso = 1,
Membership others = 0 0 1 Total Original Count 0 38 2 40 1 1 7 8 %
0 95.0 5.0 100.0 1 12.5 87.5 100.0 Cross-validated Count 0 38 2 40
1 1 7 8 % 0 95.0 5.0 100.0 1 12.5 87.5 100.0 a Cross validation is
done only for those cases in the analysis. In cross validation,
each case is classified by the functions derived from all cases
other than that case. b 93.8% of original grouped cases correctly
classified. c 93.8% of cross-validated grouped cases correctly
classified.
[0674] Small cell cancer
LDF=probe12s*0.575-probe20s*0.408-probe22s*0.423+probe23s*0.344
TABLE-US-00046 Classification Results Predicted Group Small = 1,
Membership others = 0 0 1 Total Original Count 0 39 3 42 1 1 5 6 %
0 92.9 7.1 100.0 1 16.7 83.3 100.0 Cross-validated Count 0 39 3 42
1 1 5 6 % 0 92.9 7.1 100.0 1 16.7 83.3 100.0 a Cross validation is
done only for those cases in the analysis. In cross validation,
each case is classified by the functions derived from all cases
other than that case. b 91.7% of original grouped cases correctly
classified. c 91.7% of cross-validated grouped cases correctly
classified.
[0675] Squamous cell cancer
LDF=-probe1sc*0.328-probe2sc*0.295+probe3sc*0.741+probe24s*0.490+probe25s-
*0.393+probe26s*0.426
TABLE-US-00047 Classification Results Predicted Group Squamous = 1,
Membership others = 0 0 1 Total Original Count 0 31 4 35 1 2 9 11 %
0 88.6 11.4 100.0 1 18.2 81.8 100.0 Cross-validated Count 0 31 4 35
1 2 9 11 % 0 88.6 11.4 100.0 1 18.2 81.8 100.0 a Cross validation
is done only for those cases in the analysis. In cross validation,
each case is classified by the functions derived from all cases
other than that case. b 87.0% of original grouped cases correctly
classified. c 87.0% of cross-validated grouped cases correctly
classified.
[0676] Large cell cancer LDF=probelsc*0.847+probe7sc*0.452
TABLE-US-00048 Classification Results Predicted Group Large = 1,
Membership others = 0 0 1 Total Original Count 0 23 15 38 1 4 5 9 %
0 60.5 39.5 100.0 1 44.4 55.6 100.0 Cross-validated Count 0 23 15
38 1 4 5 9 % 0 60.5 39.5 100.0 1 44.4 55.6 100.0 a Cross validation
is done only for those cases in the analysis. In cross validation,
each case is classified by the functions derived from all cases
other than that case. b 59.6% of original grouped cases correctly
classified. c 59.6% of cross-validated grouped cases correctly
classified.
[0677] The lower, but useful, performance was on a classifier
designed and tested with a very small number of cases of large cell
cancer, so this result is still very encouraging.
[0678] Adenocarcinoma,
LDF=-probe4sc*0.515+probe5sc*0.299-probe14s*0.485-s*0.347+probe20s*0.723+-
probe25s*0.327+probe27s*0.327
TABLE-US-00049 Classification Results Predicted Group Adeno = 1,
Membership Others = 0 0 1 Total Original Count 0 29 5 34 1 0 14 14
% 0 85.3 14.7 100.0 1 .0 100.0 100.0 Cross-validated Count 0 29 5
34 1 0 14 14 % 0 85.3 14.7 100.0 1 .0 100.0 100.0 a Cross
validation is done only for those cases in the analysis. In cross
validation, each case is classified by the functions derived from
all cases other than that case. b 89.6% of original grouped cases
correctly classified. c 89.6% of cross-validated grouped cases
correctly classified.
[0679] (d) Conclusion
[0680] It is very encouraging to note the performance of these
classifiers stand up to the tests of applying unseen data. This
gives a very high confidence in the ability to detect the
individual cancers.
[0681] (3) Training and Testing on Data from Different Patients and
Pathologists
[0682] As a "final final" test of robustness a LDF was trained on
the data that was reviewed by both Pathologist 1 and Pathologist 2.
This removes data reviewed by Pathologist 3. Hence testing on data
reviewed by both Pathologist 3 plus Pathologist 1 data is not
biased. Previously the test process was biased through using data
from the same patient for test and train.
[0683] LDF produced the same set of features except for probe 4
which was not included. The LDF was
=probelsc*0.288+probe7sc*0.846-probe15s*0.249-probe16s*0.534
TABLE-US-00050 Classification Results Area under the Curve = .977
Predicted Group Diagnosis Membership (UCLA) 0 1 Total Original
Count 0 20 0 20 1 9 37 46 % 0 100.0 .0 100.0 1 19.6 80.4 100.0
Cross-validated Count 0 20 0 20 1 9 37 46 % 0 100.0 .0 100.0 1 19.6
80.4 100.0 a Cross validation is done only for those cases in the
analysis. In cross validation, each case is classified by the
functions derived from all cases other than that case. b 86.4% of
original grouped cases correctly classified. c 86.4% of
cross-validated grouped cases correctly classified.
[0684] Still a reasonable result, but a similar result, but with a
smaller area under the curve, was obtained with probe7 alone on
Pathologist 3 only data
TABLE-US-00051 Classification Results Area under the curve = .908
Predicted Group Diagnosis Membership (UCLA) 0 1 Total Original
Count 0 19 1 20 1 7 39 46 % 0 95.0 5.0 100.0 1 15.2 84.8 100.0
Cross-validated Count 0 19 1 20 1 7 39 46 % 0 95.0 5.0 100.0 1 15.2
84.8 100.0 a Cross validation is done only for those cases in the
analysis. In cross validation, each case is classified by the
functions derived from all cases other than that case. b 87.9% of
original grouped cases correctly classified. c 87.9% of
cross-validated grouped cases correctly classified.
* * * * *
References