U.S. patent application number 13/760827 was filed with the patent office on 2017-06-08 for system and method for determining modified states of health of living tissue.
This patent application is currently assigned to PURDUE RESEARCH FOUNDATION. The applicant listed for this patent is PURDUE RESEARCH FOUNDATION. Invention is credited to David D. Nolte, John J. Turek.
Application Number | 20170156598 13/760827 |
Document ID | / |
Family ID | 48524491 |
Filed Date | 2017-06-08 |
United States Patent
Application |
20170156598 |
Kind Code |
A9 |
Nolte; David D. ; et
al. |
June 8, 2017 |
System and Method For Determining Modified States of Health of
Living Tissue
Abstract
A phenotypic profiling method for drug/dose physiological
response of living bodies utilizes feature recognition to segment
the information in time-frequency tissue-response spectrograms to
construct N-dimensional feature vectors. The feature vectors are
used to generate a correlation matrix among a large number of
different stimuli in the form of drugs, doses and conditions.
Multi-dimensional scaling is applied to the correlation matrix to
form a two-dimensional map of response relationships that retains
rank distances from the higher-dimensionality feature matrix. The
two-dimensional phenotypic profile space displays compact regions
indicative of particular physiological responses, such as regions
of enhanced active transport, membrane undulations and
blebbing.
Inventors: |
Nolte; David D.; (Lafayette,
IN) ; Turek; John J.; (West Lafayette, IN) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
PURDUE RESEARCH FOUNDATION |
West Lafayette |
IN |
US |
|
|
Assignee: |
PURDUE RESEARCH FOUNDATION
West Lafayette
IN
|
Prior
Publication: |
|
Document Identifier |
Publication Date |
|
US 20130144151 A1 |
June 6, 2013 |
|
|
Family ID: |
48524491 |
Appl. No.: |
13/760827 |
Filed: |
February 6, 2013 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
12874855 |
Sep 2, 2010 |
8886295 |
|
|
13760827 |
|
|
|
|
PCT/US09/36124 |
Mar 5, 2009 |
|
|
|
12874855 |
|
|
|
|
13704464 |
Dec 14, 2012 |
|
|
|
PCT/US11/40954 |
Jun 17, 2011 |
|
|
|
PCT/US09/36124 |
|
|
|
|
13704438 |
Dec 14, 2012 |
9514271 |
|
|
PCT/US11/40959 |
Jun 17, 2011 |
|
|
|
13704464 |
|
|
|
|
61034028 |
Mar 5, 2008 |
|
|
|
61397885 |
Jun 17, 2010 |
|
|
|
61397885 |
Jun 17, 2010 |
|
|
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
A61B 5/0075 20130101;
G03H 2210/62 20130101; G03H 1/08 20130101; G03H 1/0443
20130101 |
International
Class: |
A61B 5/00 20060101
A61B005/00 |
Goverment Interests
STATEMENT OF GOVERNMENT INTEREST
[0002] This invention was made with government support under grant
CBET 0756005 awarded by the National Science Foundation. The
government has certain rights in the invention.
Claims
1. A method for creating a phenotypic profile useful for drug
screening, comprising: generating a spectrogram of a differential
dynamic physiological response of a living body to a stimuli,
including a new drug and condition; and generating a feature vector
whose elements are obtained by an inner product between a feature
mask and the spectrogram, in which the feature mask corresponds to
a known response at an isolated region of a frequency-time
spectrogram.
2. The method of claim 1, wherein the frequencies in the feature
mask are obtained through an ensemble average over many
spectrograms.
3. The method of claim 2, wherein the frequencies in the feature
mask are obtained through a level-set approach that is averaged
over many spectrograms.
4. The method of claim 3, wherein the level-set is set to zero.
5. The method of claim 1, wherein the feature masks are selected
morphometrically using the spectrogram to define its own masked
regions.
6. The method of claim 1, wherein quantitative index values are
constructed from combinations of individual features.
7. The method of claim 6, wherein the quantitative index value is
an apoptotic index representative of apoptotic activity in the
living sample.
8. The method of claim 7, wherein the apoptotic index is defined by
the joint presence of low-frequency and high-frequency enhancements
in a spectrogram.
9. The method of claim 1, wherein Shannon entropy is used to
optimize the selection of one or more feature masks used to
generate a feature vector.
10. The method of claim 9, wherein the entropy of many sets of
feature masks are calculated and compared, and the feature set
leading to the largest information content is applied for
phenotypic profiling.
11. A method for creating a phenotypic profile useful for drug
screening, comprising: generating a spectrogram of a differential
dynamic physiological response of a living body to a stimuli,
including a new drug and condition; generating a feature vector
whose elements are obtained by an inner product between a feature
mask and the spectrogram, in which the feature mask corresponds to
a known response at an isolated region of a frequency-time
spectrogram; using the feature vectors to generate a phenotypic
profile for grouped drugs, including the new drug.
12. The method of claim 11, wherein the phenotypic profile is
generated through multidimensional scaling (MDS).
13. The method of claim 12, wherein a Venn diagram is overlaid on
the MDS graph.
14. The method of claim 13, wherein the Venn diagram has
physiological interpretations based on the physiological responses
isolated in the feature vector quantities.
15. The method of claim 12, wherein the MDS space is defined in
terms of trajectories of a continuously variable condition.
16. The method of claim 15, wherein the continuously variable
condition is pH.
17. The method of claim 15, wherein the continuously variable
condition is osmolarity.
18. A method for creating a phenotypic profile useful for drug
screening, comprising: generating a spectrogram of a differential
dynamic physiological response of a living body to a stimuli,
including a new drug and condition; generating a similarity matrix
of the spectrogram for the new drug and a plurality of spectrograms
for known drugs producing known physiological responses;
re-ordering the similarity matrix to group common drug responses
together; and generating a phenotypic profile for grouped drugs,
including the new drug.
19. The method of claim 1, wherein the step of re-ordering the
similarity matrix includes applying hierarchical clustering to the
similarity matrix to produce a nearly block-diagonal matrix with
groups of spectrograms that share common physiological
responses.
20. The method of claim 1, wherein the step of re-ordering the
similarity matrix includes applying multi-dimensional scaling (MDS)
to the similarity matrix that maintains the spatial near-far
relationships among drugs in each grouping.
21. The method of claim 3, wherein the multi-dimensional scaling
includes simulated annealing of a MDS cost function.
22. The method of claim 3, wherein the result of the MDS is a map
of nodes corresponding to each drug and condition in which nodes
corresponding to similar physiological responses are grouped
closely and far from unrelated drug responses.
23. The method of claim 5, further comprising generating a Venn
diagram from the map based on common physiological features.
24. The method of claim 1, wherein generating a similarity matrix
includes: generating feature masks corresponding to predetermined
physiological responses of the living body; applying the feature
masks to the spectrograms to generate a feature vector for each
spectrogram; and combining the feature vectors to produce the
similarity matrix.
25. The method of claim 7, wherein the feature masks correspond to
recognizable physiological responses that occur in characteristic
frequency ranges at characteristic times.
26. The method of claim 8, wherein the characteristic frequency
ranges include at least three ranges centered on 0.01 Hz (low), 0.1
Hz (mid) and 1 Hz (high).
27. The method of claim 8, wherein the characteristic times include
short times (0 to 50 minutes) for fast response, mid times (50 to
150 minutes) for slower response and long times (150 to 350
minutes) for long-term response.
Description
REFERENCE TO RELATED APPLICATIONS
[0001] This application is a continuation-in-part of and claims
priority to co-pending application Ser. No. 12/874,855, filed on
Sep. 2, 2010, and entitled "Method and Apparatus for Motility
Contrast Imaging," which claims priority to PCT/US2009/036124 filed
on Mar. 5, 2009, which further claims priority to U.S. provisional
application No. 61/034,028, filed on Mar. 5, 2008. This application
is also a continuation-in-part of and claims priority to co-pending
application Ser. No. 13/704,464 and to co-pending application Ser.
No. 13/704,438, both filed on Dec. 14, 2012, and entitled "Digital
Holographic Method of Measuring Cellular Activity and of Using
Results to Screen Compounds," and both claiming priority to
PCT/US2011/040954, filed on Jun. 17, 2011, which claims priority to
U.S. provisional application No. 61/397,885, filed on Jun. 17,
2010. The entire disclosure of the co-pending application Ser. Nos.
12/874,855; 13/704,438; and 13/704,464 are incorporated herein by
reference.
BACKGROUND
[0003] The present disclosure relates generally to the assessment
of the health of living tissue and the profiling of the phenotypic
response of tissue to perturbations, both environmental and
pharmaceutical.
[0004] Cellular systems are highly complex, with high redundancy
and dense cross-talk among signaling pathways. [1] (Note: Bracketed
numbers refer to general reference publications listed at the end
of the disclosure). Biochemical target-based high-content screening
(HCS) can isolate single mechanisms in important pathways, but
often fails to capture integrated system-wide responses. Phenotypic
profiling, on the other hand, presents a systems-biology approach
that has more biological relevance by capturing multimodal
influence of therapeutics. [2] Although phenotypic profiling
predates genomics that provided isolated targets, it remains today
one of the most successful approaches for the discovery of new
drugs. [3]
[0005] At present most phenotypic profiling is performed on
two-dimensional culture, even though two-dimensional monolayer
culture on flat hard surfaces does not respond to applied drugs in
the same way as cells in their natural three-dimensional
environment. This is in part because genomic profiles are not
preserved in primary monolayer cultures. [4-6] There have been
several comparative transcriptomic studies that have tracked the
expression of genes associated with cell survival, proliferation,
differentiation and resistance to therapy that are expressed
differently in 2D cultures relative to three-dimensional culture.
For example, three-dimensional culture from cell lines of
epithelial ovarian cancer [7, 8], hepatocellular carcinoma [9-11]
or colon cancer [12] display expression profiles more like those
from tumor tissues than when grown in 2D. In addition, the
three-dimensional environment of 3D culture presents different
pharmacokinetics than 2D monolayer culture and produce differences
in cancer drug sensitivities. [13-16]
[0006] An alternative approach to imaging form is to image
function, and in particular functional motions. Motion is
ubiquitous in all living things and occurs across broad spatial and
temporal scales. At one extreme, motions of molecules during
Brownian diffusion occur across nanometers at microsecond scales,
while at the other extreme motions of metastatic crawling cells
occur across millimeters taking many hours. As one spans these
scales, many different functional processes are taking place:
molecular diffusion, molecular polymerization or depolymerization
of the cytoskeleton, segregation of enzymes into vesicles,
exocytosis and endocytosis, shepherding of vesicles by molecular
motors, active transport of mitochondria, cytoskeletal forces
pushing and pulling on the nucleus, undulations of the cell
membrane, cell-to-cell adhesions, deformation of the cell, cell
division and ultimately to movement of individual cells through
tissue. All of these very different types of cellular dynamics can
be active and useful indicators of the functioning behavior of
cells. The functional response of target cells to applied drugs is
of particular relevance in drug screening.
[0007] Imaging motion in three-dimensional tissue is simpler than
imaging structure or performing molecular imaging, because motion
modulates coherent light through phase modulation. When light
scatters from an object that is displacing, the phase of the light
is modified. If the light has coherence, then the motion-induced
phase shifts of one light path interfere with the phase shifts of
other light paths in constructive and destructive interference.
Even light that is multiply scattered in tissue carries a record of
the different types of motions that the light encountered. By
measuring the fluctuating phase of light scattered from living
tissue, the different types of motion across the different space
and time scales can be measured. The trade-off for greater depth of
penetration into three-dimensional tissue is reduced spatial
imaging resolution. But volumetric imaging of intracellular motions
in tissue is still possible, within limits, by using low-coherence
interferometry that can select light from specified depths by using
coherence-gating approaches [17].
[0008] The 2D imaging techniques have been applied to drug
screening and in particular to evaluating efficacy of a drug in
disease treatment, such as anticancer drugs. Traditional 2D
techniques can lead to false positives when a drug is more
effective in 2D than in 3D, resulting in promising early drug leads
that fail in animal models because the drug is more effective at
killing tumor cells grown as monolayer cultures than as cells
within multicellular tissues. An even greater impact is the false
negative in drug screening in which a drug that would otherwise be
effective in 3D yields a negative result at the 2D screening stage,
thereby leading to the elimination of the drug. The end result of
false positives and false negatives is that new drug discovery
halves every nine years (in contrast to Moore's Law of the
electronics industry in which chip capacity doubles every 18
months).
[0009] Phenotypic profiling can provide a significant advance in
drug screening when based on 3D cultures. What is needed is a
system and method for extracting high-content phenotypic responses
from inside heterogeneous tissue and for accurately and
meaningfully analyzing the resulting information to build a
phenotype database to improve the efficiency of future drug
screening.
SUMMARY
[0010] The present system and method relies upon feature
recognition to segment the information in time-frequency
tissue-response spectrograms to construct N-dimensional feature
vectors. The feature vectors are used to generate a correlation
matrix among a large number of different stimuli in the form of
drugs, doses and conditions. Multi-dimensional scaling is applied
to the correlation matrix to form a two-dimensional map of response
relationships that retains rank distances from the
higher-dimensionality feature matrix. The two-dimensional
phenotypic profile space displays compact regions indicative of
particular physiological responses, such as regions of enhanced
active transport, membrane undulations and blebbing.
[0011] In a further attribute, the feature vectors can be used to
generate feature masks against which new drugs and drug responses
can be compared. A library of feature masks can be maintained and
applied to the spectrogram of a new drug to find a corollary drug
in the library that produces a similar physiological response.
DESCRIPTION OF THE FIGURES
[0012] FIG. 1 includes motility contrast images of three nominal
tumors, namely a fully proliferating tumor and tumors with a
hypoxic and a necrotic core.
[0013] FIG. 2 includes the spectrograms for 28 different stimuli,
namely drugs, doses and conditions.
[0014] FIGS. 3a, 3b and 3c are similarity matrices for,
respectively, shell-shell, core-core and shell-core correlations of
the spectrograms of FIG. 2
[0015] FIG. 4 is a clustered similarity matrix for the combination
core and shell responses to 56 stimuli.
[0016] FIG. 5 is a map of spectral frequency responses
corresponding to certain physiological responses of the tumor.
[0017] FIG. 6 is a graph of characteristic frequency (temporal
scale) in relation to size (spatial scale) for certain
physiological responses obtained from prior studies.
[0018] FIG. 7 includes feature masks generated from the feature
vector of stimuli pH 9 obtained from the spectrogram of FIG. 2.
[0019] FIG. 8a is a spectrogram for a particular feature
illustrating a frequency knee characteristic.
[0020] FIG. 8b is a plot of zero-crossings isolated from the
spectrogram of FIG. 8a showing he knee frequency.
[0021] FIGS. 9a and 9b are graphs corresponding to a morphometric
analysis in which FIG. 9a is the raw differential spectrogram and
FIG. 9b is the raw spectrogram modified according to a morphometric
algorithm.
[0022] FIG. 10 shows a feature vector obtained from a spectrogram
for the stimuli pH 9.
[0023] FIG. 11a is a spectrogram of feature vectors for 28
stimuli.
[0024] FIG. 11b is a similarity matrix for the feature vector
spectrogram of FIG. 11a.
[0025] FIG. 12a is a spectrogram of feature vectors for 28 stimuli
after unsupervised hierarchical clustering.
[0026] FIG. 12b is a similarity matrix for the feature vector
spectrogram of FIG. 12a.
[0027] FIG. 13 is a network diagram of illustrating strong
similarities among multiple stimuli.
[0028] FIG. 14 is a Venn diagram generated by multi-dimensional
scaling of the spectrogram of FIG. 12.
[0029] FIG. 15 is a diagram showing trajectories of pH and
osmolarity in a MDS space of several Raf inhibitors acting on
adenocarcinomas DLD-1 and HT-29.
[0030] FIG. 16 includes spectrograms of a proliferating shell and
hypoxic core of a tumor responding to particular stimuli, namely
cytochalasin D at 50 .mu.g/ml.
[0031] FIG. 17a is a graph of apoptotic indexes for the shell and
core of a tumor responding to eight stimuli.
[0032] FIG. 17b is graph comparing data for three stimuli obtained
from the tissue dynamic spectroscopy described herein and prior art
data.
[0033] FIG. 18 is a graph of entropy data for fourteen feature
vectors.
[0034] FIGS. 19a and 19b are maps of two-variable joint and mutual
entropies for the fourteen feature vectors shown in FIG. 18.
[0035] FIG. 20 shows a feature vector and hierarchical clustering
similarity matrix for 144 different stimuli.
[0036] FIG. 21 is a graph of joint entropy for 170
drugs/doses/conditions including separate shell and core
responses.
DETAILED DESCRIPTION
[0037] For the purposes of promoting an understanding of the
principles of the invention, reference will now be made to the
embodiments illustrated in the drawings and described in the
following written specification. It is understood that no
limitation to the scope of the invention is thereby intended. It is
further understood that the present invention includes any
alterations and modifications to the illustrated embodiments and
includes further applications of the principles of the invention as
would normally occur to one skilled in the art to which this
invention pertains.
[0038] The present invention determines the internal health of
living tissue by defining and monitoring internal states of health
based on motility spectrograms that are used to classify phenotypic
response of tissue to perturbations. In one aspect, the systems and
methods of present disclosure utilize motility contrast imaging to
map and evaluate cellular activity. In one embodiment, the imaging
may be performed using the optical configuration disclosed in
co-pending application Ser. No. 13/704,438, which is based on
published PCT application WO 2011/160064 (the '064 Publication), or
as disclosed in co-pending application Ser. No. 12/874,855,
published as US2020/0331672, in which a depth-resolved holographic
technique, namely holographic optical coherence imaging (OCI), is
utilized to extract information concerning cellular and subcellular
motion. The disclosure of the imaging systems and techniques in the
published application Nos. US2020/0331672 and WO 2011/160064 are
incorporated herein by reference.
[0039] The '064 Publication further describes the use of motility
contrast imaging (MCI) to produce spectrogram responses of cell
compounds. In particular, the '064 Publication describes extracting
a spectrogram fingerprint of a differential response of a tissue
sample in which a drug has been administered. The motility contrast
images of several nominal tumors are shown in
[0040] FIG. 1. The core can be strong proliferating (on the left)
or strongly hypoxic (on the right), depending on the pre-condition
of the tumor spheroids. When the core is hypoxic, it is ATP
depleted and responds differently to stimuli, such as drugs or
changing environmental conditions, than a healthy shell that has
normal ATP levels.
[0041] The '064 Publication further discloses that a power spectrum
in the frequency domain can be determined and more particularly
changes in the power spectral density as a function of time through
the shift in characteristic parameters are evaluated. The result is
a differential spectrogram, such as the spectrograms shown in FIG.
2, which plots the differential relative spectral density for a
given frequency at a time t relative to the initial time t.sub.0.
This differential spectrogram thus captures the changes in spectral
content as a function of time after a dose, condition or other
stimuli (or combination of stimuli) are applied to the target 3D
body. The differential spectrogram is sensitive to shifts in the
magnitudes of the spectral densities (dependent upon the number of
moving constituents) and to shifts in the characteristic
frequencies (i.e., the speed of the moving constituents).
Spectrogram Correlations
[0042] In one procedure, 28 different drugs, concentrations and
conditions were applied to 28 different tumor spheroids, and
spectrograms were obtained in each case using the imaging system
and techniques described in the above-identified published
applications. Table I below lists the different compounds, doses or
conditions (hereinafter referred to as "stimuli") applied to the
tumor spheroids, together with the expected physiological
response.
TABLE-US-00001 TABLE I Reference Compounds, Doses and Conditions
Compound or Condition Action Temperature 24 to 37.degree. C.
Increased motility pH 6 Weak acidic KCN 20 .mu.g/ml Inhibit
electron transport Iodoacetate 1 .mu.g/ml Inhibits glycolysis Osm
508 Hypertonic cell desiccation Osm 428 Hypertonic cell desiccation
Iodoacetate 10 .mu.g/ml Inhibit glycolysis KCN 200 .mu.g/ml Inhibit
electron transport TNF 5 .mu.g/ml Cytokine apoptosis induction
Cytochalasin 10 .mu.g/ml Anti-actin, anti-mitotic Cytochalasin 1
.mu.g/ml Anti-actin, anti-mitotic Iodoacetate 40 .mu.g/ml Inhibit
glycolysis pH 9 Strong basic Nocodazole 0.01 .mu.g/ml Anti-tubulin,
anti-mitotic Osm 154 Strong Hypotonic cell swelling Osm 77 Strong
Hypotonic cell swelling Cytochalasin 50 .mu.g/ml Anti-actin,
anti-mitotic, apoptosis pH 5 Strong acidic Colchicine 10 .mu.g/ml
Anti-tubulin, anti-mitotic Iodoacetate 20 .mu.g/ml Inhibit
glycolysis Nocodazole 1 .mu.g/ml Anti-tubulin, anti-mitotic Taxol
10 .mu.g/ml Tubulin stabilization, anti-mitotic Cycloheximide
Apoptosis induction Nocodazole 0.1 .mu.g/ml Anti-tubulin,
anti-mitotic Colchicine 1 .mu.g/ml Anti-tubulin, anti-mitotic
Nocodazole 10 .mu.g/ml Anti-tubulin, anti-mitotic Taxol 1 .mu.g/ml
Tubulin stabilization, anti-mitotic pH 8 Weak basic
[0043] The spectrograms for these drugs/stimuli are complex,
showing many different types of features in response to the applied
drugs. Similar drugs produced similar spectrograms, while widely
differing spectrograms could be elicited from drugs with very
different mechanisms of action. The spectrograms of the 28
different drugs/doses/conditions, separated by shell (top) and core
(bottom), are shown in FIG. 2. For each spectrogram, frequency is
along the vertical axis from 0.005 Hz to 5 Hz, while time is
plotted along the horizontal axis corresponding to a duration of 6
hours after the dose or stimuli is applied. In the spectrograms of
FIG. 2, increases in spectral density over the initial condition
are represented by increasing shades of red while decreasing
spectral density are represented by increasing shades of blue. For
instance, for the drug/dose Nocodazole 0.01 .mu.g/ml the spectral
density shows a moderate increase over time at the low frequencies,
while at the dose 0.1 .mu.g/ml the increase is much more
pronounced.
[0044] To initially capture the similarity or dissimilarity of the
drug response spectrograms, a cross-correlation coefficient between
spectrograms is calculated, as specified by:
C ( x , y ) = v , t S ( x ) ( v , t ) S ( y ) ( v , t ) v , t ( S (
x ) ( v , t ) ) 2 v , t ( S ( y ) ( v , t ) ) 2 ##EQU00001##
where S.sup.(x)(v,t) is the spectrogram of drug/condition x,
S.sup.(y)(v,t) is the spectrogram of drug/condition y and the sum
is over both frequency and time. The correlation coefficient was
calculated among the 28 conditions, treating the proliferating
shell and the core of the tumor spheroid separately.
[0045] A similarity matrix provides the basis for a hierarchical
clustering algorithm that can group different stimuli by their
similar drug-response spectrograms. Each row of the N.times.N
similarity matrix is an N-dimensional vector (where N corresponds
to the number of compounds, doses and conditions applied to the
tumor spheroid). The inner product of each row defines a measure of
distance:
Dist ( A , B ) = 1 N y = 1 N C ( A , y ) C ( B , y )
##EQU00002##
where C is the cross-correlation coefficient calculated above.
[0046] Inner products near unity correspond to "close" associations
between stimuli, and inner products near negative unity correspond
to "far" associations.
[0047] The stimuli given in Table I can be grouped according to the
hierarchical clustering of the drug responses based on the
spectrograms of the proliferating shell. In a hierarchical
clustering algorithm, at each stage the two closest vectors are
identified and grouped into an average vector, and then the number
of vectors decreases by one. The sequence of the vectors that are
grouped in this way are retained until all vectors have been
grouped. The sequential grouping of vectors produces a clustering
of similar drug responses. The similarity matrix can be rearranged
according to the sequence of grouping according to the shell-shell
correlations with the result shown in upper left matrix of FIG. 3a.
The similarity matrix shows structure with an approximately "block
diagonal" appearance. If the matrix were truly block diagonal, the
clustering would consist of groups of unique behaviors with no
correlation between stimuli. However, in the case of the drug
responses, there is considerable off-diagonal structure, which
indicates that different groups of drug responses do share some
aspects in common. Nevertheless, the drug response groupings
exhibit "themes" of similar behavior. For instance, many of the
anti-mitotic drugs, such as Nocodazole, Colchicine and Taxol,
cluster in the bottom right of the figure, the cytochalasin and
hypotonic osmolarity results are approximately in the middle, and
the metabolic and hypertonic osmolarity responses are at the upper
left.
[0048] The compounds were applied at multiple different doses, some
below and some above EC50. Therefore, the clustering does not
automatically group similar drugs together, but instead clusters
responses. If a higher dose induces apoptosis, but a lower dose
does not, then the higher dose will be grouped with other drugs or
conditions that induce apoptosis. In an alternative analysis, each
drug response can be referenced to its own EC50, which would remove
the dose dependence in the clustering.
[0049] Multicellular tumor spheroids have different conditions for
the proliferating shell relative to the core, which tends to be
hypoxic and ATP depleted. Furthermore, large tumors (larger than
500 micron diameter) have increasingly necrotic cores. Tissue
dynamics spectroscopy is volumetric (depth-gated) so the response
of the (deeper) core can be compared to the different response of
the (shallower) shell. The core-core correlations are shown in the
lower right FIG. 3b for the same ordering as for the clustered data
using the shell-shell correlations. There are many similarities in
the core-core correlations, but with notable differences, and with
several of the shell-shell blocks missing, such as the blocks in
the upper right of the shell-shell matrix of FIG. 3a. The
differences are because the shell and the core respond differently
to the applied conditions. This is seen more clearly in upper right
FIG. 3c which shows the correlation of the shell spectrograms with
the core spectrograms. The diagonal elements of this similarity
matrix are not all near unity, showing significant differences in
the response of the core relative to the shell, even responding in
the same tumor to the same conditions.
[0050] Because the core is under a different condition than the
shell, the 28 conditions can be doubled to 56 conditions, which can
be cross-correlated into a similarity matrix and clustered
according to all 56 conditions. The resulting clustered similarity
matrix is shown in FIG. 4. (Note that the stimuli applied to the
core are so designated and the stimuli applied to the shell carry
no separate identifier). Many of the shell and core spectrograms
are clustered together (highly similar), but there are many other
pairs that are separated in the clustering process because their
respective spectrograms are not highly similar. This is likely a
consequence of the ATP-depleted conditions of the core relative to
the proliferating shell. For instance, an apoptotic response
requires normal ATP concentrations and can occur in the shell,
while the cells in the core would be forced to follow a necrotic
path. This difference between apoptosis and necrosis based on ATP
concentrations may be one of the differentiators between the
spectrograms of the shell relative to the core. A more systematic
study of these differences between the shell and the core may
uncover specific spectrogram signatures for processes like
apoptosis.
Motility Spectrograms and Cellular Dynamics
[0051] The information content of each spectrogram can be
represented as feature maps that capture the signatures of the
different drugs. These features can be related to physiological
processes that occur inside the cells and generally in the tissue.
For instance, the trend in the TDS (tissue dynamics spectroscopy)
spectral frequencies are shown in FIG. 5 for motions of the
organelles, membrane and general shape changes. The spectral
frequency is along the y-axis and spans three decades of dynamic
range, while time is along the x-axis. The baseline is set prior to
t=0, and is used as the quantity in the denominator of spectral
density equation for normalization. The change in the spectral
content is plotted in false color, with Region R in deep red equal
to 70% enhancement, and the Region B in deep blue equal to 70%
inhibition. The intermediate Region G in green represents little or
no change in spectral content. The response occurs in approximately
three frequency bands that show distinctly different behavior:
low-frequency (0.005 Hz to 0.1 Hz), mid-frequency (0.1 Hz to 1 Hz)
and high-frequency (1 Hz to 5 Hz). The proliferating shell shows
enhancements in the third frequency band after about an hour, while
the core shows nearly an opposite response. At low frequencies, the
shell is mostly unchanged until a strong onset at a much later
time, while the core shows enhanced low frequencies for most of the
duration of the experiment.
[0052] To interpret spectrograms, it is necessary to establish a
correspondence of the frequencies observed in DLS (dynamic light
scattering) with frequencies (and velocities and diffusion
coefficients) obtained from the literature that are connected with
specific biological targets and mechanisms. The lowest frequency in
the experimental spectrograms described above is 0.005 Hz, and the
highest frequency is 5 Hz. The general relationships for single
backscattering under heterodyne (holographic) detection is given
by: q.sup.2D=.omega..sub.D for diffusion, and qv=.omega..sub.d for
directed transport, where D is the diffusion coefficient and v is a
directed speed. The smallest and largest frequencies that can be
captured in the experiments define the physical ranges for directed
transport and diffusion, respectively, which are 0.006<v<2
.mu.m/sec, and 4.times.10.sup.-4<D<0.1 .mu.m.sup.2/sec.
[0053] The velocity range is well within the range of intracellular
motion in which molecular motors move organelles at speeds of
microns per second [18-22]. Diffusion of very small organelles, as
well as molecular diffusion, are too fast to be resolved by an
expected maximum frame rate of 10 fps. Membrane undulations are a
common feature of cellular motions, leading to the phenomenon of
flicker [23-27]. The characteristic frequency for membrane
undulations tends to be in the range around 0.01 to 0.1 Hz [21, 24,
28, 29]. Results from the literature are summarized in FIG. 6. The
graph captures the general connection of spatial scale with
temporal scale. Experiments on vesicles and the cytoplasm give the
highest backscatter frequencies generally above 1 Hz and extending
to tens of Hz. Larger mitochondria and organelles have slightly
lower backscattering frequencies, but these are still in the range
of the upper band frequencies of TDS shown in FIG. 5. Membrane
motions are much slower, coinciding with the frequencies of the
lower band in TDS. This spatial-temporal trend is only
semi-quantitative, but it provides a general principle that may
help disentangle the mixtures of frequencies that arise from
multiple dynamic light scattering mechanisms.
Feature Masks
[0054] Therefore, there is a general trend of higher frequencies
for smaller objects, and also trends in time with higher doses
producing faster responses. These trends in frequency and time are
captured by using feature masks that look at isolated regions of
the frequency-time spectrogram. One embodiment of these feature
masks is illustrated in FIG. 7.
[0055] The features used in the feature mask relate to a position
in time and frequency on the spectrogram. For instance, a membrane
undulation response is a mid-frequency response, while an organelle
response is a high frequency response. Cell shape changes are low
frequency. A hormetic response is a combined high/low frequency
shift that occurs immediately upon application of the drug, but
then decays quickly. All drug/condition spectrograms can be
analyzed in terms of these features.
[0056] The drug-response spectrograms exhibit recognizable features
that occur in characteristic frequency ranges at characteristic
times after a dose is applied. Two approaches are applied to
feature recognition and quantification of the drug-response
spectrograms. One approach is based on projections of the
spectrograms onto feature masks. There are many possible choices
for feature masks, such as binary masks versus continuous-valued
masks, local masks versus global masks, and orthonormal feature
masks versus non-orthonormal. The time axis on the spectrograms
primarily captures relaxation which is typically exponential. The
frequency axis is the Fourier transform of the autocorrelation
function, which also is typically exponentially correlated.
Therefore, both the time and frequency axes are characterized by
Laplace transforms for which there is no orthonormal basis.
Therefore, an approach of matched filtering [30] can be used in
which characteristic regions of the spectrograms can be interpreted
mechanistically. For instance, it has been shown that low
frequencies correspond to large-scale motion like blebbing or the
formation of apoptotic bodies, while high frequencies correspond to
membrane vesicles or internal organelle motions. [31] These
processes and frequencies present natural frequency ranges within
the data that are used to generate feature masks to extract these
specific features.
[0057] Along the time axis, an increasing sampling for the feature
masks is applied, sampling at short times (T1=0 to 50 minutes) for
fast response, mid times (T2=50 to 150 minutes) for slower response
and long times (T3=150 to 350 minutes) for long-term response.
Along the frequency axis there are several characteristic
frequencies that divide the response into natural frequency bands.
These occur at 0.01 Hz (low), 0.1 Hz (mid) and 1 Hz (high). The
resulting feature masks are shown in FIG. 7. The three time frames
and five frequency signatures generate 15 feature masks. Masks
M1-M3 capture the average spectral power change, masks M4-M6
capture low-frequency, masks M7-M9 measure the shift of the
spectral weight to higher frequencies, masks M10-M12 capture high
frequency motions, and masks M13-M15 capture enhanced mid
frequencies.
[0058] In one rudimentary approach, overlaying a particular
drug/dose spectrogram onto each of feature masks can reveal the
expected response for the drug based on the feature mask or masks
that generally correspond to the spectrogram. For instance
overlaying the spectrogram of FIG. 5 onto the feature masks of FIG.
7 shows a general correspondence with mask M6 indicative of large
scale motion of the cell and with M8 indicative of organelle
transport that starts at a later time.
[0059] One important aspect of the time-frequency feature mask
approach is the choice of frequency cutoffs for the different
features. For instance, it is clear by assessing many of the
spectrograms that there is a typical frequency knee, or edge,
around 0.1 Hz. This frequency would be a natural frequency to
define the boundary of a feature mask. To precisely establish this
frequency, a level-set algorithm can be applied to an ensemble of
many spectrograms from the same cell line responding to many
different types of drugs. The level is set to the zero-crossings of
the spectrogram. These zero-crossings are isolated in each of the
spectrograms and then averaged. The result is shown in FIGS.
8(a)-8(b). There is a strong knee feature at a frequency around 0.1
Hz that is shared by most of the spectrograms. This becomes the
bounding frequency of mid- or low-frequency masks. There is also a
knee frequency around 1 Hz that defines the lower bound of the
high-frequency mask that is associated with organelle
transport.
[0060] The second approach to feature extraction of spectrogram
features is based on morphometric analysis. This analysis is based
on the overall shape and position of certain features. An example
of a morphometric analysis is shown in FIGS. 9a-9b for the
mitochondrial toxin FCCP applied to a DLD-1 cell-line tumor
spheroid. Time is the vertical axis, increasing downwards on a
linear scale. The horizontal axis is frequency increasing to the
right on a log scale. The drug is applied at t=10 on the vertical
axis. The spectrogram in FIG. 9a is the raw differential
spectrogram. This is used as the starting point of a morphometric
algorithm that seeks broad areas of common response. The algorithm
is based on percolation cluster labeling combined with level sets.
The resulting identified regions are shown in FIG. 9b. There are 7
regions I-VII that were found in this case to be distinct. These
regions become the masks that are applied to the spectrogram in
FIG. 9a to extract the average values inside that time-frequency
region. It is important to note that different spectrograms may
differ quantitatively, but share common features. This morphometric
approach can lead to feature vectors that are similar, while the
more rigid approach to feature recognition described above might
create feature vectors that are not highly similar because of
frequency shifts between the two spectrograms. Therefore, the
morphometric approach is considered to provide more robust
comparisons of qualitatively similar features and hence more stable
clustering results based on these feature vectors.
Feature Vectors
[0061] In one aspect of the present disclosure, each feature in the
feature masks can be assigned a numerical value related to the
strength and sign of the response for that feature in a particular
spectrogram. The collection of feature values constitutes a feature
vector. Therefore, each spectrogram has an associated feature
vector.
[0062] The value of a feature is obtained by the inner product of
the k-th feature mask M.sub.k(v,t) with the j-th differential
spectrogram D.sup.j(v,t)
V.sub.k.sup.j=D.sup.j(v,t),M.sub.k(v,t)
that produces a k-dimensional vector component V.sub.k.sup.j where
j is the index for a condition (drug or dose or perturbation), and
k is the index for a feature (time-frequency signature). The
brackets denote the inner product through:
D j ( v , t ) , M k ( v , t ) = 1 N norm p , q D pq j M k , pq
##EQU00003##
where the indexes p, q are along the frequency and time axes,
respectively. The normalization N.sub.norm is
N norm = p , q M pq M pq i , j D pq D pq ##EQU00004##
where D.sub.pq is taken only in the associated time region of the
mask. This normalization captures the shape of features, but
de-emphasizes the magnitude. This procedure is well suited to
recognizing signature features and for clustering drug responses
according to their features rather than the strength of their
response.
[0063] An example of a feature vector is shown in FIG. 10
associated with a spectrogram for pH 9. A collection of feature
vectors for each of the 28 stimuli/conditions (see FIGS. 3a-c, FIG.
4), and including shell and core response separately, is shown in
FIG. 11a. Each feature vector is like an N-dimensional vector. In
the example of FIG. 11a, each drug or condition has a
10-dimensional vector that captures ten features of the
spectrograms. There can be any number of dimensions, depending on
the image masks that are generated to capture specific features. In
this example, the key features include transparency,
membrane/contacts, undulations, expansion, organelle transport,
delayed membrane, delayed undulations, delayed expansion, delayed
organelle transport and hormetic response. The strength of the
responses are color-coded from red for a strong positive response
to blue for a strong negative response.
Clustering
[0064] The similarity between two feature vectors for two different
spectrograms is obtained as the inner, or dot, product between the
two vectors. After the feature vectors are generated as shown in
FIG. 11a, they are used to generate a similarity matrix among all
the different drugs and conditions. The similarity of the i-th and
j-th spectrogram is defined as the normalized correlation:
S ij = k V k i V k j k V k i V k i k V k j V k j ##EQU00005##
where the similarity is normalized to unity when i=j. The feature
vectors V.sub.k.sup.j and the similarity matrix S.sup.ij for the 28
doses and conditions of Table I above, plus two negative controls,
are shown in FIGS. 11a-b. In these figures, the proliferating shell
and the core are treated independently, leading to 60 feature
vectors and a 60.times.60 similarity matrix. The similarity matrix
in the present example is shown in FIG. 11b and is unity along the
main diagonal. The ordering in FIG. 11b is by drug and dose, with
the shell alternating with the core. While there is some clustering
of similarities, the similarity matrix and feature vectors do not
show strong clustering, except for the tubulin destabilizing
compounds that share many features in common.
[0065] Despite the grouping of increasing doses of some drugs (e.
g., Nocodazole, Iodoacetate, cytochalasin, etc.), there is little
structure to the matrix in FIG. 11b so the informational content is
of little value in drug screening. One of the goals of phenotypic
profiling is to re-order the similarity matrix to group common drug
responses together. There is no unique way to order similarity, and
there are several common algorithms that lead to similar, but not
identical, orderings. One known approach is unsupervised
hierarchical clustering. [32] (See another example in "Unsupervised
Hierarchical Clustering via a Genetic Algorithm", William A. Greene
(Proc. IEEE Cong. Evol. Comput., Canberra , Australia, 2003, pp.
998-1005), the entire disclosure of which is incorporated herein by
reference). This is an iterative process that sequentially groups
pairs of similar responses to produce tree structures. Unsupervised
ordering has the advantage that no predefined structures or
substructures are defined for the process, leading to an objective
grouping of similar responses.
[0066] The similarity matrix and feature vectors after unsupervised
hierarchical clustering according to "distance" are shown in FIG.
12a-12b. In comparing the "non-clustered" similarity matrix of FIG.
11b with the matrix of FIG. 12b it can be seen that the latter
similarity matrix is nearly block-diagonal in which similar
spectrograms are grouped together, with low cross-talk among the
groups. The feature vectors are also grouped in FIG. 12a, showing
groups of spectrograms that share features in common. The selected
features have putative physiological interpretations (such as
blebbing, undulations and organelle motions)[31], so the groupings
of features also group the physiological responses into a
phenotypic profile for these standard compounds and conditions.
[0067] It can thus be appreciated that the clustered similarity
matrix of FIG. 12b is higher in information content than the
similarity matrix of FIG. 11b, and therefore more useful in
screening a new drug. In one aspect, the spectrogram of a new
drug/dose/stimuli can be reduced to a feature vector, as explained
above, and then combined with feature vectors of known
drug/dose/stimuli, in the manner of the feature vector matrix of
FIG. 11a, and ultimately incorporated into a similarity matrix,
such as the matrix of FIG. 11b. After clustering a clustered
similarity matrix such as the matrix of FIG. 12b can be used to
identify the group in which the new drug/dose/stimuli resides. The
new drug can then be expected to produce the physiological
responses of that group.
[0068] There are several aspects to note in the ordered list of the
vector in FIG. 12a. In general, similar compound classes are
clustered together, such as the anti-tubulin drugs Nocodazole and
Colchicine. But these compounds cluster separately from Taxol which
is a tubulin stabilizing drug. In addition, similar environmental
conditions cluster together, such as hypotonic osmolarity and low
pH separately from hypertonic osmolarity and high pH. One aspect
that is striking in its absence is the low correlation between the
shell and the core response of the same tumor spheroid to the same
compound or condition. This phenotypic difference between the shell
and the core of a single spheroid is discussed in more detail
herein.
Multidimensional Scaling
[0069] The disadvantage of hierarchical clustering is that it
produces a linear arrangement that does not faithfully represent
distance between nodes (in which a node corresponds to a
drug/dose/stimuli). In addition, it is an ordering algorithm rather
than a clustering algorithm. [33] To test the results of
hierarchical clustering, a k-means clustering and level-set
clustering can be applied on the raw similarity matrix in FIG. 4.
The resulting ordering of the drugs and conditions change in
detail, but the overall groupings and block-diagonal structure of
the reordered similarity matrix remain nearly the same with respect
to the similarity matrix shown in FIG. 12b. However, none of these
ordering or clustering algorithms faithfully represent the distance
relationships that are contained within the original similarity
matrix. As subclusters are combined into larger clusters,
especially in hierarchical clustering, it is possible to have two
adjacent nodes that have very low similarity and hence should be
far apart. The final ordering is linear, and distance relationships
are not represented.
[0070] As one solution to this "distance" problem, the groups of
clustered cases can be converted to a network diagram, as shown in
FIG. 13. Each numbered node represents a particular stimuli (i.e.,
compound, drug or condition). Strong similarity is denoted by a
line connecting two nodes. Close proximity of nodes signifies
similar phenotypic response. The network diagram of FIG. 13
provides a clear visual indication of groupings according to
phenotypic response that can be useful in drug screening.
[0071] A more quantitative solution to this "distance" problem is
provided by multidimensional scaling (MDS). [32] In
multidimensional scaling, a low-dimensional representation is
sought in which similar response nodes are placed close to one
another and dissimilar response nodes are placed far apart. The
spatial near-far relationships in a low-dimensional graph preserve
the high-dimensional near-far relationships of the original feature
vectors. Multidimensional scaling is not a "projection" to a lower
dimensional plane, but is instead a one-to-one mapping that retains
ordering of distances. As with all data compression the actual
metric distances are lost in the process; however, the general
relationships of nearness or similarity are preserved.
[0072] To apply multidimensional scaling to a similarity matrix,
such as the matrix of FIG. 4, a distance matrix is defined as
d.sup.ij=1-S.sup.ij. The MDS algorithm assigns a cost function that
is to be minimized. The cost function is the difference between the
distance of the two-dimensional location of nodes and the distance
defined from the similarity matrix. The cost function is:
C = j > i ( r i - r j - d ij ) 2 ##EQU00006##
where r.sub.i is a vector in the low-dimensional target space.
Using a simple Euclidean distance norm this cost function is
minimized using simulated annealing [34], starting from a set of
initial positions in two dimensions determined by the first two
principal components from principal component analysis. [32]
(Examples are also shown in "Pairwise data clustering by
deterministic annealing", T. Hofmann, J. M. Buhmann, Pattern
Analysis and Machine Intelligence, IEEE Transactions, Vol. 19, no.
1, pp. 1-14, January 1997; and "Multidimensional Scaling by
Deterministic Annealing," H. Klock, J. Buhmann, Proc. Int'l
Workshop Energy Minimization Methods in Computer Vision and Pattern
Recognition, pp. 245-260, 197; both disclosures of which are
incorporated herein by reference). The simulated annealing proceeds
through 1000 iterations, with a 0.1% decrease in effective
temperature at each iteration. The final positions that emerge from
the MDS algorithm appear in the low-dimensional space in a pattern
that preserves the trend of distances d.sub.ij from the
high-dimensional space spanned by the feature vectors.
[0073] The results of multidimensional scaling of the data in FIG.
12 are shown in FIG. 14. This low-dimensional phenotypic space has
two dimensions. The axes do not have direct physical meaning, but
the distribution of nodes retains the ordering of distances that
occur in the similarity matrix of FIG. 12. Similar drug responses
are grouped closely, and far from other unrelated drug responses.
Because of this spatial grouping and separation, it is possible to
draw a Venn diagram based on the physiological features in the
feature vectors. The Venn diagram of FIG. 14 shows the drug
responses with enhanced mid-frequency (membrane undulations),
enhanced low frequency (shape changes and blebbing) and enhanced
high frequency/active (vesicle and organelle transport). The
network connections drawn on the figure are from k-means clustering
and show local clusters. The dashed line shows a separation between
core response (outside the line) and shell response (inside the
line). There are exceptions in this shell-core separation for
hypotonic osmolarity, high pH and high-dosage cytochalasin, but in
general the shell responses lie closer together near the center of
the phenotypic space than the outlying core responses.
[0074] An interesting region is the overlap between blebbing and
active organelle/vesicle transport. Taken together, these may be
expected if there is apoptosis during which apoptotic bodies are
separated from the main cell, and active organelle transport drives
the cellular decomposition. There are no drug responses in this
overlap region from the hypoxic core which is ATP depleted and
cannot support apoptosis.
[0075] It can be seen that the information content of the map of
FIG. 14 far exceeds the content of the original similarity matrix
of FIG. 4. The map resulting from the multidimensional scaling can
be easily used to identify similar physiological responses among
drugs and more importantly to identify expected physiological
responses of new drugs being screened. It can be appreciated that
the all of the steps leading from the original spectrogram for a
new drug to the map of FIG. 14 can be performed with software.
[0076] Former applications of multidimensional scaling have had the
drawback that the axes do not relate to physically meaningful
quantities. However, the application of specific conditions to
living tissues can be performed with conditions that have
continuously variable values that are continuously tuned through
normal physiological conditions. By including these data in the
multidimensional scaling, one obtains a so-called trajectory for
that condition through the MDS space. One example of such a
condition is the pH of the culture medium that can be tuned from
pH=5 to pH=9 passing through physiological normal pH=7.3. Another
example is the osmolarity that can be tuned from 150 mOsm through
physiological 300 mOsm to 450 mOsm. Examples of these trajectories
on the MDS phenotypic profile of a group of Raf inhibitors is shown
in FIG. 15. The pH and osmolarity data define trajectories through
the MDS space. These trajectories enable the axes of the MDS space
to acquire physiological meaning in relation to the drug
responses.
Proliferating Shell and Hypoxic Core Phenotypes
[0077] As an example of phenotypic profiling based on hypoxic
phenotypes, multicellular tumor spheroids provide a natural format
to study phenotypic differences in the drug response between
normoxic and hypoxic tissue. When tumor spheroids have a diameter
larger than approximately 400 microns, the transport of oxygen into
the core of the spheroid is impeded, resulting in hypoxic tissue
and a band of quiescent cells inside the outer shell of
proliferating cells. [35] There is evidence that the phenomenon of
multicellular resistance to anticancer drugs displayed by avascular
solid tumors is caused, in part, by the population of quiescent
cells. [36, 37] In addition, hypoxia is a factor in oncogenic
progression [38, 39] and may participate in the epithelial to
mesenchymal transition that ultimately leads to metastasis. [40-43]
For these reasons, comparing the effect of anticancer drugs on the
hypoxic core relative to the quiescent or proliferating shells may
shed light on multicellular resistance.
[0078] An understanding of the shell-core differences can be gained
by analyzing the shell and core spectrograms responding to
cytochalasin D at 50 .mu.g/ml shown in FIG. 16. The core
spectrogram at the bottom of FIG. 16 shows an initially strong
mid-frequency enhancement (shortly after time 0) caused by the
weakening of the actin cortex that reduces the stiffness of the
cell membrane. After two hours, there is a strong onset of
low-frequency enhancement for both the shell and the core. However,
the shell, shown in the top spectrogram in FIG. 16, shows a strong
enhancement of the high-frequencies (starting just before two
hours) associated with vesicle and organelle transport that is
missing in the core spectrogram response. This represents a
significant physiological difference in the role of the drug for
normoxic vs. hypoxic tissue. Based on the physiological processes
of apoptosis (active transport requiring ATP and hence normal
oxygen levels) relative to necrosis (passive degradation of the
cell without need for ATP), this phenotypic difference can be
tentatively ascribed to apoptosis in the shell as opposed to
necrosis in the core. This assignment of high-frequency
fluctuations associated with apoptosis is consistent with
decorrelation times measured using OCT in vitro [44], and is also
supported by confocal two-photon microscopy of these drugs on
spheroids of these cell types.
[0079] The active transport that is present in the shell but
missing in the core, combined with the presence of low-frequency
enhancement that is common to both, suggests a metric that is the
logical AND of both of the features to create an "apoptotic" index
that quantitatively captures this feature. This index is
constructed from the feature masks appropriate to the low- and
high-frequency features, which are masks M6 and M12 of FIG. 7. The
apoptotic index can be defined by:
M.sup.i=sqrt(V.sub.6.sup.i)sqrt(V.sub.12.sup.i)
A.sup.i=Rc{M.sup.i}-Im{M.sup.i}
in which V.sub.6.sup.i and V.sub.12.sup.i are the feature vector
values for masks M6 and M12 (determined as discussed above). The
difference between the real and the imaginary parts generates the
logical AND with positive apoptotic indexes only for features that
themselves are both positive.
[0080] The apoptotic index was calculated for the shell and the
core spectrograms and ranked in decreasing order by the shell. The
apoptotic index of the shell response for the top eight apoptotic
indexes is shown in FIG. 17a above the zero line. The strongest
apoptotic index is for the shell responding to cytochalasin at 50
.mu.g/ml. Also shown in FIG. 17a below the zero line are the
corresponding apoptotic indexes for the core responses in each
case. For this set of 8 drug doses, that produced the strongest
positive apoptotic indexes in the shell, the apoptotic index of the
cores are all negative. In the cases when the shell has strong
active transport in addition to strong low-frequency enhancements,
the core is missing one or the other of these features. Upon closer
inspection, the missing feature is the active transport band at
high frequency in all cases. The key physiological aspect of the
core is the lack of ATP to drive active processes, such as the
formation of sequestering vesicles and apoptotic bodies. The
negative apoptotic indexes of the core region for each of these
drugs reflects the quiescent/hypoxic resistance of solid tumors to
anti-cancer drugs. For instance, the anti-actin anti-mitotic drug
cytochalasin D shows the strongest apoptotic index for the shell,
but the strongest negative mitotic index for the core. Taxol and
Colchicine also show strong differences in the apoptotic index
between the shell and the core.
[0081] The independent validation for apoptotic processes to
confirm their participation in the drug response was obtained
through a series of multiphoton confocal microscopy experiments
carried out using the UMR-106 spheroids. Spheroids were cultured at
37.degree. C. in the presence of 10 .mu.g/mL cytochalasin D,
paclitaxel, and Colchicine for 4 hours and Iodoacetate for 3 hours.
The live, apoptotic, and dead cells in the spheroid outer shell
were visualized by optical section using a 20.times. water
immersion objective up to a depth of 100 .mu.m using Hoechst 33342
(live), Yo-Pro-1 (apoptotic), and propidium iodide (dead)
(Invitrogen, Grand Island, N.Y.) vital dyes on a Nikon AIR
multiphoton microscope with Mai Tai DeepSee tunable IR laser at 750
nm. The percentage of live, apoptotic, and dead cells in the
spheroid outer shell was determined for six different spheroids.
The results for Nocodazole, cytochalasin and Iodoacetate, all at 10
.mu.g/ml, correspond to the upper line in the graph of FIG. 17b.
The data are ranked by decreasing apoptotic ratio. These data show
that Nocodazole, cytochalasin and Iodoacetate all induce an
apoptotic response in the tumor spheroids. The strongest apoptotic
response is for 10 .mu.g/ml Iodoacetate, which has the strongest
apoptotic index in FIG. 17b. The bottom line in the graph of FIG.
17b shows the corresponding data for the tissue dynamic
spectroscopy shown in FIG. 17a.
Information Content of TDS
[0082] A key question about the drug-response spectrograms is how
much information is contained in these data structures. There are
many types of intracellular motions, including organelle transport,
endo- and exocytosis, membrane undulations, cytoplasmic streaming,
cytoskeletal rearrangements, force relaxation and shape changes,
among others. While general trends in the spectrograms are
understood in terms of these types of motion, it needs to be
established how much information can be obtained from tissue
dynamics spectroscopy.
[0083] Information can be measured in terms of Shannon entropy. For
a probability distribution p(x), the Shannon entropy is defined
as:
H ( x ) = - i p ( x i ) log p ( x i ) ##EQU00007##
[0084] The probability distribution for the feature vectors of
phenotypic profiling by tissue dynamics spectroscopy is obtained as
the histogram of the values among the individual features. The
entropy for the data in FIGS. 12a-12b is shown in FIG. 18 for the
fifteen features comparing the shell and core responses versus
shell alone. The horizontal axis corresponds to the 15 features
while the vertical axis represents the Shannon entropy value, which
if applied to the feature vectors identifies how much information
is obtained from the vectors. The entropy values for shell and core
feature vectors are all moderately higher than for only shell
feature vectors. The average entropy per channel is approximately
two bits.
[0085] An important aspect of information analysis of a data
structure like a drug-response spectrogram is the definition of
joint and mutual entropies for joint probability distributions.
These entropy measures arise when there is shared information among
numerous channels. For instance, for two channels, the joint
entropy is:
J ( x , y ) = - ij p ( x i , y j ) log p ( x i , y j )
##EQU00008##
and the mutual entropy is:
M ( x , y ) = ij p ( x i , y j ) log p ( x i , y j ) p ( x i ) p (
y j ) = H ( x ) + H ( y ) - J ( x , y ) ##EQU00009##
where p(x.sub.i,y.sub.j) is the joint probability of two variables.
The mutual entropy is equal to zero for independent variables,
while M(x,y)=H(x)=H(y) for linearly dependent variables. Similarly,
J(x,y)=H(x)+H(y) for independent variables, and J(x,y)=H(x)=H(y)
for linearly dependent variables. In other words, when there are
two channels that have perfect correlation, the total information
content is simply the information content of a single channel.
[0086] Examples of two-variable joint and mutual entropies are
plotted in FIGS. 19a and 19b for the fourteen features of the 60
drugs, doses and conditions of the spectrograms of FIGS. 12a-12b.
From the mutual entropy figure, there are cases of strong mutual
information content among some pairs of channels, for instance
channel 11 and channel 2 share almost 2 bits of information. (See,
FIG. 19b).
[0087] The joint entropy among three channels is:
J ( x , y , z ) = - ijk p ( x i , y j , z k ) log p ( x i , y j , z
k ) ##EQU00010##
and by extension for n channels is:
J ( x n ) = - 1 n p ( x n ) log p ( x n ) ##EQU00011##
[0088] Therefore, the total information content of the
drug-response spectrograms is obtained in the limit:
J .infin. = lim n -> .infin. J ( x n ) ##EQU00012##
[0089] To calculate the total information content of tissue
dynamics spectroscopy, an expanded data set of 170 different drugs,
doses and conditions (and hence 170 spectrograms) is shown in FIG.
20. The joint entropy for n=2, 3, 4 and 5 is shown in FIG. 21. The
asymptotic value in the limit of large n is 6.8 for the 170
spectrograms. Therefore, in the drug-response spectrograms, there
are 6.8 bits of information per spectrogram. It is important to
note that high-content screening (HCS) typically uses approximately
this many channels. Therefore, tissue dynamics spectroscopy
qualifies as a high-content screening approach for drug phenotypic
profiling.
Application to Drug Screening
[0090] The drug-response spectrograms described above exhibit
recognizable features that occur at characteristic frequency ranges
at characteristic times after a dose and/or other stimuli is
applied. Each spectrogram acts like a fingerprint, or more
appropriately a voice print, showing a distinct pattern in response
to the applied stimuli that is unique to the particular stimuli,
which may be a particular drug and cell line. An analysis of the
information content of the drug-response spectrograms based on
Shannon entropy (as described in more detail above) indicates that
more than 144 distinct spectrograms can be distinguished from one
another, which can thereby be used to establish 144 different
"classes" for compound/cell-line classifiers.
[0091] The present disclosure contemplates feature recognition and
quantification of the drug-response spectrograms from projections
of the spectrograms onto feature masks. The time axis of
spectrograms primarily captures relaxation which is typically
exponential. The frequency axis is the Fourier transform of the
autocorrelation function, which is typically exponentially
correlated. Thus, both the time and frequency axes are
characterized by Laplace transforms for which there is no
orthonormal basis. Consequently, in one aspect the systems and
methods disclosed herein apply matched filtering in which
characteristic regions of the spectrograms can be interpreted
mechanistically. For instance, as discussed above low frequencies
correspond to slow cellular shape changes like blebbing, mid
frequencies correspond to membrane undulations and high frequencies
correspond to membrane vesicles or internal organelle motions.
These processes and frequencies present natural frequency ranges
that are used to generate the feature masks, such as the masks
shown in FIG. 7.
[0092] The inner product of each mask with a drug response
spectrogram generates a feature vector that serves as the
fingerprint for that particular drug. For instance, the feature
vector for pH 9 is shown in FIG. 10. The feature vector illuminates
the sensitivity of some features of the associated spectrogram,
such as the organelle transport in the high frequency range after
one hour, mid-frequency suppression of delayed undulations after
four hours, delayed membrane at the low frequencies after 31/2
hours and an initial hormetic response. The feature masks capture
these changes in the feature vector for the particular stimuli
response.
[0093] As described above, 15 feature vectors can be used
corresponding to five frequency ranges across three time frames
(although other frequency and time frame divisions can be utilized)
and a similarity matrix is constructed from the correlation
coefficients among the many pairs of feature vectors for different
drugs and cell lines. A clustering algorithm (such as unsupervised
hierarchical clustering) can be applied to the similarity matrix to
order the tissue-response spectrograms of the different drugs and
cell lines into groups based on the similarity of their response.
The result is a clustered matrix that exhibits a block diagonal
structure which is indicative that groups with high similarity have
little overlap with other groups. By comparing a feature vector
with the groupings it is possible to assign physiological
attributes to the different groups. It is this quasi-orthogonality
among the groups that provides the basis for phenotypic
classification schemes in which a new lead compound of unknown
mechanism may be compared against a reference compound library of
dynamic tissue response spectrograms having known mechanisms of
action.
[0094] To improve on the results from the clustering approach, the
present system and method contemplates the use of multidimensional
scaling to more faithfully represent the closeness and farness of
compound/cell type groupings and to display the information in a
readily understandable format. The result is a two dimensional map
with nodes corresponding to stimuli distributed in a manner that
retains the order of distances found in the similarity matrix, such
as the Venn diagram shown in FIG. 14. Similar drug responses are
grouped closely and farther from other unrelated drug responses.
The map or Venn diagram thus provides a ready means to determine
the expected phenotypic response of a new drug based on its
proximity to other drugs in a grouping. For instance, if a new
condition is OSM 154, a quick review of the map of FIG. 14 would
reveal that the phenotypic response is blebbing and similar to the
response of pH6, among other drugs/doses.
[0095] The feature masks, feature vectors, similarity matrices,
clustering and multi-dimensional scaling all readily lend
themselves to implementation in software. Thus, the drug screening
process can be significantly automated from the generation of the
spectrogram for a new drug/stimuli to the presentation of a Venn
diagram similar to FIG. 14 from which the expected physiological
response of a new drug can be readily discerned. The systems and
methods disclosed herein can thus significantly reduce the amount
of time required to assess a new drug. More importantly the
multi-dimensional scaling approaches implemented herein more
accurately align drugs/conditions having similar physiological
responses, which can significantly reduce, if not eliminate, the
risk of false positives and false negatives.
[0096] A library of known drugs/conditions and associated
spectrograms can provide the foundation for the analysis of a new
drug/stimuli/condition. All of the known drugs/conditions can be
incorporated into a similarity matrix with the spectrogram of one
or more new drugs/conditions to produce the clustered matrices and
maps described above. In other words, as new drugs are categorized
based on their physiological response they can be added to the
library from which comparisons can be made for future
drugs/conditions. Since the methods described herein can be readily
implemented in software, the time required to validate or screen a
new drug can be significantly reduced.
[0097] While the invention has been illustrated and described in
detail in the drawings and foregoing description, the same should
be considered as illustrative and not restrictive in character. It
is understood that only the preferred embodiments have been
presented and that all changes, modifications and further
applications that come within the spirit of the invention are
desired to be protected.
REFERENCES
[0098] [1] C. Ainsworth, "NETWORKING FOR NEW DRUGS," Nature
Medicine, vol. 17, pp. 1166-1168, October 2011. [0099] [2] C. T.
Keith, A. A. Borisy, and B. R. Stockwell, "Multicomponent
therapeutics for networked systems," Nature Reviews Drug Discovery,
vol. 4, pp. 71-U10, January 2005. [0100] [3] D. C. Swinney and J.
Anthony, "How were new medicines discovered?," Nature Reviews Drug
Discovery, vol. 10, pp. 507-519, July 2011. [0101] [4] P. Hamer, S.
Leenstra, C. J. F. Van Noorden, and A. H. Zwinderman, "Organotypic
Glioma Spheroids for Screening of Experimental Therapies: How Many
Spheroids and Sections are Required?," Cytometry Part A, vol. 75A,
pp. 528-534, June 2009. [0102] [5] J. Poland, P. Sinha, A. Siegert,
M. Schnolzer, U. Korf, and S. Hauptmann, "Comparison of protein
expression profiles between monolayer and spheroid cell culture of
HT-29 cells revealed fragmentation of CK18 in three-dimensional
cell culture," Electrophoresis, vol. 23, pp. 1174-1184, April 2002.
[0103] [6] K. Dardousis, C. Voolstra, M. Roengvoraphoj, A.
Sekandarzad, S. Mesghenna, J. Winkler, Y. Ko, J. Hescheler, and A.
Sachinidis, "Identification of differentially expressed genes
involved in the formation of multicellular tumor spheroids by HT-29
colon carcinoma cells," Molecular Therapy, vol. 15, pp. 94-102,
January 2007. [0104] [7] N. A. L. Cody, M. Zietarska, A.
Filali-Mouhim, D. M. Provencher, A. M. Mes-Masson, and P. N. Tonin,
"Influence of monolayer, spheroid, and tumor growth conditions on
chromosome 3 gene expression in tumorigenic epithelial ovarian
cancer cell lines," Bmc Medical Genomics, vol. 1, Aug. 7, 2008.
[0105] [8] M. Zietarska, C. M. Maugard, A. Filali-Mouhim, M.
Alam-Fahmy, P. N. Tonin, D. M. Provencher, and A. M. Mes-Masson,
"Molecular description of a 3D in vitro model for the study of
epithelial ovarian cancer (EOC)," Molecular Carcinogenesis, vol.
46, pp. 872-885, October 2007. [0106] [.sup.9] T. T. Chang and M.
Hughes-Fulford, "Monolayer and Spheroid Culture of Human Liver
Hepatocellular Carcinoma Cell Line Cells Demonstrate Distinct
Global Gene Expression Patterns and Functional Phenotypes," Tissue
Engineering Part A, vol. 15, pp. 559-567, March 2009. [0107] [10]
M. Shimada, Y. Yamashita, S. Tanaka, K. Shirabe, K. Nakazawa, H.
Ijima, R. Sakiyama, J. Fukuda, K. Funatsu, and K. Sugimachi,
"Characteristic gene expression induced by polyurethane
foam/spheroid culture of hepatoma cell line, Hep G2 as a promising
cell source for bioartificial liver," Hepato-Gastroenterology, vol.
54, pp. 814-820, April-May 2007. [0108] [11] Y. Yamashita, M.
Shimada, N. Harimoto, S. Tanaka, K. Shirabe, H. Ijima, K. Nakazawa,
J. Fukuda, K. Funatsu, and Y. Maehara, "cDNA microarray analysis in
hepatocyte differentiation in Huh 7 cells," Cell Transplantation,
vol. 13, pp. 793-799, 2004. [0109] [12] L. Gaedtke, L. Thoenes, C.
Culmsee, B. Mayer, and E. Wagner, "Proteomic analysis reveals
differences in protein expression in spheroid versus monolayer
cultures of low-passage colon carcinoma cells," Journal of Proteome
Research, vol. 6, pp. 4111-4118, November 2007. [0110] [13] A.
Frankel, R. Buckman, and R. S. Kerbel, "Abrogation of taxol-induced
G(2)-M arrest and apoptosis in human ovarian cancer cells grown as
multicellular tumor spheroids," Cancer Research, vol. 57, pp.
2388-2393, Jun. 15, 1997. [0111] [14] L. A. Hazlehurst, T. H.
Landowski, and W. S. Dalton, "Role of the tumor microenvironment in
mediating de novo resistance to drugs and physiological mediators
of cell death," Oncogene, vol. 22, pp. 7396-7402, Oct. 20, 2003.
[0112] [15] I. Serebriiskii, R. Castello-Cros, A. Lamb, E. A.
Golemis, and E. Cukierman, "Fibroblast-derived 3D matrix
differentially regulates the growth and drug-responsiveness of
human cancer cells," Matrix Biology, vol. 27, pp. 573-585, July
2008. [0113] [16] L. David, V. Dulong, D. Le Cerf, L. Cazin, M.
Lamacz, and J. P. Vannier, "Hyaluronan hydrogel: An appropriate
three-dimensional model for evaluation of anticancer drug
sensitivity," Acta Biomaterialia, vol. 4, pp. 256-263, March 2008.
[0114] [17] R. C. Youngquist, S. Carr, and D. E. N. Davies,
"Optical coherence-domain reflectometry: a new optical evaluation
technique," Opt. Lett., vol. 12, p. 158, 1987. [0115] [18] M.
Suissa, C. Place, E. Goillot, and E. Freyssingeas, "Internal
dynamics of a living cell nucleus investigated by dynamic light
scattering," European Physical Journal E, vol. 26, pp. 435-448,
August 2008. [0116] [19] X. L. Nan, P. A. Sims, and X. S. Xie,
"Organelle tracking in a living cell with microsecond time
resolution and nanometer spatial precision," Chemphyschem, vol. 9,
pp. 707-712, Apr. 4, 2008. [0117] [20] K. J. Karnaky, L. T.
Garretson, and R. G. Oneil, "Video-Enhanced Microscopy of Organelle
Movement in an Intact Epithelium," Journal of Morphology, vol. 213,
pp. 21-31, July 1992. [0118] [21] N. A. Brazhe, A. R. Brazhe, A. N.
Pavlov, L. A. Erokhova, A. I. Yusipovich, G. V. Maksimov, E.
Mosekilde, and O. V. Sosnovtseva, "Unraveling cell processes:
Interference imaging interwoven with data analysis," Journal of
Biological Physics, vol. 32, pp. 191-208, October 2006. [0119] [22]
B. Trinczek, A. Ebneth, and E. Mandelkow, "Tau regulates the
attachment/detachment but not the speed of motors in
microtubule-dependent transport of single vesicles and organelles,"
Journal of Cell Science, vol. 112, pp. 2355-2367, July 1999. [0120]
[23] F. Brochard and J. F. Lennon, "Frequency Spectrum of Flicker
Phenomenon in Erythrocytes," Journal De Physique, vol. 36, pp.
1035-1047, 1975. [0121] [24] H. Strey and M. Peterson, "Measurement
of Erythrocyte-Membrane Elasticity by Flicker Eigenmode
Decomposition," Biophysical Journal, vol. 69, pp. 478-488, August
1995. [0122] [25] A. Zilker, M. Ziegler, and E. Sackmann,
"Spectral-Analysis of Erythrocyte Flickering in the 0.3-4-Mu-M-1
Regime by Microinterferometry Combined with Fast Image-Processing,"
Physical Review A, vol. 46, pp. 7998-8002, Dec. 15, 1992. [0123]
[26] M. A. Peterson, H. Strey, and E. Sackmann, "Theoretical and
Phase-Contrast Microscopic Eigenmode Analysis of Erythrocyte
Flicker--Amplitudes," Journal De Physique Ii, vol. 2, pp.
1273-1285, May 1992. [0124] [27] Y. Z. Yoon, H. Hong, A. Brown, D.
C. Kim, D. J. Kang, V. L. Lew, and P. Cicuta, "Flickering Analysis
of Erythrocyte Mechanical Properties: Dependence on Oxygenation
Level, Cell Shape, and Hydration Level," Biophysical Journal, vol.
97, pp. 1606-1615, Sep. 16, 2009. [0125] [28] J. Evans, W. Gratzer,
N. Mohandas, K. Parker, and J. Sleep, "Fluctuations of the red
blood cell membrane: Relation to mechanical properties and lack of
ATP dependence," Biophysical Journal, vol. 94, pp. 4134-4144, May
15, 2008. [0126] [29] V. Racine, M. Sachse, J. Salamero, V.
Fraisier, A. Trubuil, and J. B. Sibarita, "Visualization and
quantification of vesicle trafficking on a three-dimensional
cytoskeleton network in living cells," Journal Of
Microscopy-Oxford, vol. 225, pp. 214-228, March 2007. [0127] [30]
Ifeachor, E. and B. Jervis (2001). Digital Signal Processing: A
Practical Approach, Prentice Hall. [0128] [31] Nolte, D. D., R. An,
et al. (2011). "Holographic tissue dynamics spectroscopy." Journal
of Biomedical Optics 16(8): 087004-087013. [0129] [32] Johnson, R.
A. (2001). Applied Multivariate Statistical Analysis, Prentice
Hall. [0130] [33] Andreopoulos, B., A. J. An, et al. (2009). "A
roadmap of clustering algorithms: finding a match for a biomedical
application." Briefings in Bioinformatics 10(3): 297-314. [0131]
[34] Lampert, T. A. and S. E. M. O'Keefe (2010). "A survey of
spectrogram track detection algorithms." Applied Acoustics 71(2):
87-100. [0132] [35] Venkatasubramanian, R., M. A. Henson, et al.
(2006). "Incorporating energy metabolism into a growth model of
multicellular tumor spheroids." Journal of Theoretical Biology
242(2): 440-453. [0133] [36] Mellor, H. R., D. J. P. Ferguson, et
al. (2005). "A model of quiescent tumour microregions for
evaluating multicellular resistance to chemotherapeutic drugs."
British Journal of Cancer 93(3): 302-309. [0134] [37] Fayad, W., L.
Rickardson, et al. (2011). "Identification of Agents that Induce
Apoptosis of Multicellular Tumour Spheroids: Enrichment for Mitotic
Inhibitors with Hydrophobic Properties." Chemical Biology &
Drug Design 78(4): 547-557. [0135] [38] Zhou, J., T. Schmid, et al.
(2006). "Tumor hypoxia and cancer progression." Cancer Letters
237(1): 10-21. [0136] [39] Gillies, R. J. and R. A. Gatenby (2007).
"Hypoxia and adaptive landscapes in the evolution of
carcinogenesis." Cancer and Metastasis Reviews 26(2): 311-317.
[0137] [40] Sullivan, R. and C. H. Graham (2007). "Hypoxia-driven
selection of the metastatic phenotype." Cancer and Metastasis
Reviews 26(2): 319-331. [0138] [41] Gort, E. H., A. J. Groot, et
al. (2008). "Hypoxic regulation of metastasis via hypoxia-inducible
factors." Current Molecular Medicine 8(1): 60-67. [0139] [42]
Klymkowsky, M. W. and P. Savagner (2009). "Epithelial-Mesenchymal
Transition A Cancer Researcher's Conceptual Friend and Foe."
American Journal of Pathology 174(5): 1588-1593. [0140] [43] Yan,
W., Y. Fu, et al. (2009). "PI3 kinase/Akt signaling mediates
epithelial-mesenchymal transition in hypoxic hepatocellular
carcinoma cells." Biochemical and Biophysical Research
Communications 382(3): 631-636. [0141] [44] Farhat, G., A.
Mariampillai, et al. (2011). "Detecting apoptosis using dynamic
light scattering with optical coherence tomography." Journal of
Biomedical Optics 16(7).
* * * * *