Methods and systems for predicting drug responses and combinations

Abstract

Provided herein are methods of using a machine learning model, e.g., optimal transport, in analyzing responses of biological samples to drugs and drug combinations, as determined by infrared and / or Raman microscopy.

Description

[0001] METHODS AND SYSTEMS FOR PREDICTING DRUG RESPONSES AND COMBINATIONS

[0002] Field of the Invention

[0003] The invention relates to methods and systems for predicting responses to drug and drug combinations using vibrational microscopy.

[0004] Cross-Reference to Related Applications

[0005] This application claims benefit of U.S. Provisional Application No. 63 / 566,422 filed March 18, 2024, the content of which is incorporated by reference.

[0006] Background of the Invention

[0007] Understanding the effects of genetic or chemical perturbations on cellular state lies at the heart across every aspect of biology ranging from understanding molecular functions to the drug discovery pipeline. For instance, knowledge of how a cell's state and molecular programs are altered in response to specific genetic or chemical perturbations can be used to pinpoint key programs or genes as effective drug targets. Furthermore, by understanding the influence of drug dosages and combinations, these perturbations can be controlled to fine-tune, enhance, or personalize their effects. A key challenge in the study of perturbations, however, lies in the heterogeneous nature of their responses, which arise from complicating factors such as cell state, cell-to-cell signaling, and the cell’s local environment, in addition to the cell type.

[0008] Cell type-based approaches, which make the simplifying assumption of discrete cellular states, fail to model the influence of biological phenomena, and new analytical strategies are needed to compensate.

[0009] Summary of the Invention

[0010] In one aspect, the invention features a method of analyzing a biological sample comprising: (a) obtaining one or more infrared images of the biological sample using infrared microscopy and / or one or more Raman images of the biological sample using Raman microscopy; and (b) applying, by a computer, a machine learning model to analyze the one or more infrared images and / or the one or more Raman images.

[0011] In one aspect, the invention features a method of identifying a biological state in a biological sample comprising: (a) obtaining one or more infrared images of the biological sample using infrared microscopy and / or one or more Raman images of the biological sample using Raman microscopy; and (b) applying, by a computer, a machine learning model to analyze the one or more infrared images and / or the one or more Raman images, wherein an output of the analysis indicates the biological state.

[0012] In one aspect, the invention features a method of determining the efficacy of an agent in effecting a change in a biological sample comprising: (a) obtaining one or more infrared images and / or one or more Raman images of the biological sample after administration of the agent, wherein the one or more infrared images are obtained using infrared microscopy and / or the one or more Raman images are obtained using Raman microscopy; and (b) applying, by a computer, a machine learning model to analyze the one or more infrared images and / or the one or more Raman images, wherein an output of the analysis indicates that the agent is effective in effecting a change in the biological sample.

[0013] In some embodiments, prior to obtaining the one infrared images and / or the one or more Raman images, one or more infrared probes and / or one or more Raman probes are added to the biological sample.

[0014] In some embodiments, the one or more infrared probes and / or the one or more Raman probes comprise a probe for detecting unsaturated fatty acid uptake. In some embodiments, the probe for detecting unsaturated fatty acid uptake comprises ds4 oleic acid.

[0015] In some embodiments, the one or more infrared probes and / or the one or more Raman probes comprise a probe for detecting saturated fatty acid uptake. In some embodiments, the probe for detecting saturated fatty acid uptake comprises azido palmitic acid.

[0016] In some embodiments, the one or more infrared probes and / or the one or more Raman probes comprise a probe for detecting protein synthesis. In some embodiments, the probe for detecting protein synthesis comprises13C labeled amino acids.

[0017] In a preferred embodiment, the infrared microscopy and / or Raman microscopy are performed without infrared probes or Raman probes. In a preferred embodiment, the one or more infrared images and / or the one or more Raman images are obtained without addition of any infrared probes or Raman probes to the biological sample.

[0018] In some embodiments, each of the one or more infrared images comprises a collection of pixels, wherein each pixel comprises an infrared spectrum, and / or wherein each of the one or more Raman images comprises a collection of pixels, wherein each pixel comprises a Raman spectrum.

[0019] In some embodiments, the method further comprises determining a chemical composition of the biological sample at each pixel.

[0020] In some embodiments, the infrared microscopy produces a spectrum from 400 cm1to 4000 cm1and / or the Raman microscopy produces a spectrum from 500 cm1to 3500 cm1. In some embodiments, the spectrum is from 900 cm1to 1800 cm1.

[0021] In some embodiments, the one or more infrared images and / or the one or more Raman images comprises one or more segmentation images at a spectrum from 2830 cm1to 3000 cm1.

[0022] In some embodiments, the computer applies a cell segmentation method to the one or more segmentation images to generate an output of a segmented cell.

[0023] In some embodiments, each of the one or more infrared images comprises a collection of pixels, wherein each pixel comprises an infrared spectrum, and / or wherein each of the one or more Raman images comprises a collection of pixels, wherein each pixel comprises a Raman spectrum, and wherein: (a) the infrared spectra of a subset of the collection of pixels of the one or more infrared images comprising the segmented cell are averaged to generate an average infrared spectrum of the segmented cell; and / or (b) the Raman spectra of a subset of the collection of pixels of the one or more Raman images comprising the segmented cells are averaged to generate an average Raman spectrum of the segmented cell.

[0024] In some embodiments, the cell segmentation method is an Otsu thresholding method. In some embodiments, (a) the average infrared spectrum of the segmented cell is normalized by scaling the area of the average infrared spectrum to 1 ; and / or (b) the average Raman spectrum of the segmented cell is normalized by scaling the area of a subspectrum of from 2815 cm1to 3015 cm1to 1 .

[0025] In some embodiments, the machine learning model was developed using a combination of one or more reference infrared images and / or one or more reference Raman images. In some embodiments, the machine learning model is an optimal transport model.

[0026] In some embodiments, the one or more reference infrared images and / or the one or more reference Raman images were obtained using infrared microscopy and / or Raman microscopy, respectively, from a reference sample comprising reference cells having reference cell states.

[0027] In some embodiments, the optimal transport model is applied to the one or more infrared images and / or the one or more Raman images by minimizing the cost function for transforming the infrared spectra and / or the Raman spectra of the biological sample to an infrared spectra and / or a Raman spectra of the reference cells.

[0028] In some embodiments, the reference cells comprising cells of the biological sample prior to addition of an agent for effecting a change.

[0029] In some embodiments, the agent is a small molecule, a nucleic acid, a protein, a gene editing agent, or a CAR-T cell. In some embodiments, the nucleic acid is an RNAi, an shRNA, an mRNA, or an antisense oligonucleotide. In some embodiments, the protein is a transcription factor, an antibody, a peptide, a cytokine, a hormone, an enzyme, an antibody drug conjugate, or a fusion protein. In some embodiments, the gene editing agent is clustered regularly interspaced short palindromic repeats (CRISPR), a zinc finger nuclease, a base editor, a prime editor, or transcription activator-like effectors (TALEN). In some embodiments, the cell is a chimeric antigen receptor T (CAR-T) cell.

[0030] In some embodiments, the biological sample comprises a cell, a tissue, and / or a cellular component. In some embodiments, the biological sample was obtained from a human. In some embodiments, the biological sample is obtained from a subject with a disorder. In some embodiments, the biological sample comprises human cancer cells. In some embodiments, the biological sample comprises human breast cancer carcinoma cells.

[0031] In some embodiments, the effected change comprises treating a disorder, altering a gene or protein expression profile, inducing cell differentiation, inducing a cell state change, inducing cell death, or a combination thereof.

[0032] In one aspect, the invention features a computer whose input data is images obtained using infrared microscopy and / or images obtained using Raman microscopy, wherein the computer is programmed with a machine learning model developed using a combination of images obtained using infrared microscopy and / or Raman microscopy.

[0033] In some embodiments, the machine learning model is an optimal transport model.

[0034] In one aspect, the invention features a system comprising: (a) an infrared imaging component configured for infrared microscopy and / or a Raman imaging component configured for Raman microscopy; and (b) a computer configured to receive images from the infrared imaging component and / or images from the Raman imaging component, wherein the infrared imaging component and / or the Raman imaging component are each operatively coupled to the computer, wherein the computer is programed with a machine learning model whose input data is the images from the infrared imaging component and / or the Raman imaging component. In one aspect, the invention features a method of predicting the response of a subject to a drug comprising: (a) obtaining one or more infrared images and / or one or more Raman images of a biological sample after administration of the drug, wherein the biological sample was obtained from the subject, and wherein the one or more infrared images are obtained using infrared microscopy and / or the one or more Raman images are obtained using Raman microscopy; and (b) applying, by a computer, a machine learning model to analyze the one or more infrared images and / or the one or more Raman images, wherein an output of the analysis is predictive of the response of the biological sample to the drug.

[0035] In one aspect, the invention features a method of predicting the response of a subject to a drug combination comprising: (a) obtaining one or more infrared images and / or one or more Raman images of a biological sample after administration of the drug combination, wherein the biological sample was obtained from the subject, and wherein the one or more infrared images are obtained using infrared microscopy and / or the one or more Raman images are obtained using Raman microscopy; and (b) applying, by a computer, a machine learning model to analyze the one or more infrared images and / or the one or more Raman images, wherein an output of the analysis is predictive of the response of the biological sample to the drug combination.

[0036] In one aspect, the invention features a method of determining the mechanism of action, dosage, or target of a drug comprising: (a) obtaining one or more infrared images and / or one or more Raman images of a biological sample after administration of the drug, and wherein the one or more infrared images are obtained using infrared microscopy and / or the one or more Raman images are obtained using Raman microscopy; and (b) applying, by a computer, a machine learning model to analyze the one or more infrared images and / or the one or more Raman images, wherein an output of the analysis indicates the mechanism of action, dosage, or target of the drug.

[0037] In one aspect, the invention features a method of determining the mechanism of action, dosage, or target of a drug combination comprising: (a) obtaining one or more infrared images and / or one or more Raman images of a biological sample after administration of the drug combination, and wherein the one or more infrared images are obtained using infrared microscopy and / or the one or more Raman images are obtained using Raman microscopy; and (b) applying, by a computer, a machine learning model to analyze the one or more infrared images and / or the one or more Raman images, wherein an output of the analysis indicates the mechanism of action, dosage, or target of the drug combination.

[0038] Other features and advantages of the invention will be apparent from the following detailed description and figures, and from the claims.

[0039] Brief Description of the Drawings

[0040] The patent or application file contains at least one drawing executed in color. Copies of this patent or patent application with color drawings will be provided by the Office upon request and payment of the necessary fee.

[0041] FIG. 1 is a scheme demonstrating the concept of VIB-OT, combining vibrational imaging with optimal transport (OT).

[0042] FIG. 2A - FIG. 2C are graphs showing responses in chemical activity to sixteen drugs across nine method of actions and analyzed with VIB-OT. FIG. 2A: Responses are plotted as line plots, grouped by the drug’s mechanism of action. Shaded regions cover 95% of the amplitude distribution at each wavenumber. Vibrational modes are associated to various chemical activities of interest, such as lipid activity, and are visualized as shaded backgrounds. For instance, C-0 bonds are mainly prevalent in glycogen and vibrate at a frequence of 1050 cm1, marked by the blue shaded region. FIG. 2B: Responses can be analyzed and interpreted by computing drug signatures as the average responses over these regions of interest. FIG. 2C: Protein synthesis rates can be computed as the ratio between13C-AA and intrinsic amide peak, i.e. 1615 cnr1 / (1615 cm-1+1650 cm'1). The difference between the rates of the coupled treated and control cells are plotted as a box plot, grouped by different methods of action, and sorted by the median value of each method of action group.

[0043] FIG. 3A - FIG. 3C are graphs showing relationship between animyosin-treated metabolic activity and cell states. FIG. 3A: Lineplot of observed treated states of anisomycin-treated cells. Lineplot shading reflects the 5-95% inclusion thresholds at each wavenumber. Shaded areas correspond to regions of interest. FIG. 3B: Jointplot of coupled control and treated metabolic activity. A coupling is computed using the entire spectra. Then metabolic activity is computed for control and target cells and the density of this coupling is visualized as kde plot (left). Marginal distributions over the metabolic activity are shown on x and y axes. KMeans (k=2) clustering is applied to the treated cells, labels are propagated to the control cells (right). FIG. 3C: Box plot of lipid ester activity computed in control cells. High / low [Metabolic activity] cluster labels are propagated from treated states via our induced coupling onto the control cells. The distribution was plotted for the high and low clusters and the marginal over all cells.

[0044] FIG. 4A - FIG. 4E are graphs demonstrating correction of batch effects. FIG. 4A: Baseline corrected control cells, without normalization or batch effect correction applied. FIG. 4B: Silhouette scores computed using two labels, batch labels and effect labels, for baseline-corrected spectra and three normalization strategies with and without batch correction. The silhouette score measures how well a labeling scheme separates data. A value close to zero implies that labels are well mixed, while larger positive values imply that the data is well separated with respect to the labels, i.e. for batch labels, large / small scores imply strong / weak batch effects. Dots represent treatments and are scattered as a function of the silhouette score computed over condition labels (x-axis) and batch labels (y-axis). Batch scores are computed within each treatment across all three batches and condition scores are computed between the treated and control cells. Scores are computed on data not involved in the computation of statistics used to perform batch effect correction. FIG. 4C: Batch-corrected signal-normalized spectra for each condition including control. FIG. 4D: KNN enrichment computed on coupled responses for k=25 (top) and k=50 (bottom). Scores are computed within each condition and then grouped by the batch. The dashed line indicates the ideal 1 / 3 enrichment. FIG. 4E UMAP projections computed from signal- normalized batch-corrected data.

[0045] FIG. 5 is a graph showing response to iniparib treatment.

[0046] FIG. 6A - FIG. 6F are graphs demonstrating determination of dosage effects and trajectories. FIG. 6A: Lineplot of the responses yielded by the induced coupling between each step of the ani trajectory including the full effect from control to the highest dosage. Regions of interest in the spectra are highlighted in color. FIG. 6B: Joint KDE plot visualizing the induced coupling for sample regions of interest at each step in the trajectory. The density represents the frequency of (source, target) pairs for activity in each region. The dashed line represents the diagonal, i.e. no response to the treatment is observed. FIG. 6C and FIG. 6D are similar to FIG. 6A and FIG. 6B, but for the tvb trajectory. FIG. 6E and FIG. 6F: Box plots showing the change in activity within each region of interest at different dosages for the anisomycin (FIG. 6E) and tvb-3166 (FIG. 6F) treatments.

[0047] FIG. 7 A - FIG. 7C are graphs demonstrating the use of VIB-OT with Raman microscopy. FIG. 7A: Control cells profiled with spontaneous Raman technology. Due to the difference in intensity, wavenumbers 900 cm1- 2300 cm1(left) are plotted separately from the 2700 cm1- 3100 cm1regions (right). Regions of interest are highlighted. FIG. 7B: Lineplots depicting response of cells to increasing dosage of ani. Cells are treated to DMSO control and three different dosages of ani. A dose trajectory is induced by computing couplings between each increasing dosage and the difference between coupled cells are plotted in each axis. FIG. 7C: Heatmap depicting drug signatures, computed as the average difference in absorption at each region of interest, for each step in the trajectory.

[0048] FIG. 8A - FIG. 8C are graphs demonstrating the use of VIB-OT with Raman microscopy. FIG. 8A: Control cells profiled with spontaneous Raman technology. Due to the difference in intensity, wavenumbers 900 cm1- 2300 cm1(left) are plotted separately from the 2700 cm1- 3100 cm1regions (right). Regions of interest are highlighted. FIG. 8B: Lineplots depicting response of cells to increasing dosage of tvb. Cells are treated to DMSO control and three different dosages of tvb. A dose trajectory is induced by computing couplings between each increasing dosage and the difference between coupled cells are plotted in each axis. FIG. 8C: Heatmap depicting drug signatures, computed as the average difference in absorption at each region of interest, for each step in the trajectory.

[0049] FIG. 9A - FIG. 9C are graphs showing dosage effects on combination of Gef and Bor treatments. FIG. 9A: changes in IR profiles under different combinations of Gef and Bor. Top left shows the untreated state. Orange (blue) lines show the changes in IR profile as Gef (Bor) concentration is increased. FIG. 9B and FIG. 9C: changes from control state for each combination dose. Boxplots show the change in activity under each area of interest in the IR spectra. Amide I activity is particularly sensitive to the dosage.

[0050] FIG. 10A - FIG. 10C are graphs demonstrating determination of dosage effects and trajectories of combination of Gef and Bor treatments. FIG. 10A: Lineplot of the responses yielded by the different concentrations of Gef or Bor treatment. Regions of interest in the spectra are highlighted in color. FIG. 10B: Joint KDE plot visualizing the induced coupling for sample regions of interest at each step in the trajectory. The density represents the frequency of (source, target) pairs for activity in each region. The dashed line represents the diagonal, i.e. no response to the treatment is observed. FIG. 10C: Lineplot of the responses yielded by a combination of the different concentrations of Gef and Bor treatment. Regions of interest in the spectra are highlighted in color.

[0051] Definitions

[0052] It is to be understood that aspects and embodiments of the invention described herein include “comprising,” “consisting,” and “consisting essentially of” aspects and embodiments.

[0053] As used herein, the singular form “a,” “an,” and “the” includes plural references unless indicated otherwise.

[0054] The term “about” as used herein refers to the usual error range for the respective value readily known to the skilled person in this technical field. Reference to “about” a value or parameter herein includes (and describes) embodiments that are directed to that value or parameter per se. In some instances, reference to “about” a value or parameter herein indicates the value or parameter ± 10%. The term “computer” as used herein may refer to device embodied in any of a number of forms, such as a rack-mounted computer, a desktop computer, a laptop computer, or a tablet computer, as nonlimiting examples. Additionally, a computer may be embedded in a device not generally regarded as a computer but with suitable processing capabilities, including a Personal Digital Assistant (PDA), a smartphone or any other suitable portable or fixed electronic device. A computer may have one or more communication devices, which may be used to interconnect the computer to one or more other devices and / or systems, such as, for example, one or more networks in any suitable form, including a local area network or a wide area network, such as an enterprise network, and intelligent network (IN) or the Internet. Such networks may be based on any suitable technology and may operate according to any suitable protocol and may include wireless networks or wired networks. A computer may have one or more input devices and / or one or more output devices. These devices can be used, among other things, to present a user interface. Examples of output devices that may be used to provide a user interface include printers or display screens for visual presentation of output and speakers or other sound generating devices for audible presentation of output. Examples of input devices that may be used for a user interface include keyboards, and pointing devices, such as mice, touch pads, and digitizing tablets. As another example, a computer may receive input information through speech recognition or in other audible formats. Computer-executable instructions may be in many forms, such as program modules, executed by one or more computers or other devices. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The functionality of the program modules may be combined or distributed as desired in various embodiments. Databases, if employed in the methods or devices or systems herein, may include computer readable memory (also referred to as “memory”). For example, data storage space 3memlN may be and / or include computer readable memory, used to store data as described in the disclosure. Memory may be embodied by suitable hardware, including but not limited to the following: hard disk drives, serial advanced technology attachment (SATA) hard drives, SATA solid state drives (SSDs), non-volatile memory express (NVMe) SSDs, tape drives. Computers described herein may implement an optimal transport model.

[0055] The terms “program,” “app,” and “software” are used herein in a generic sense to refer to any type of computer code or set of computer-executable instructions that may be employed to program a computer or other processor to implement various embodiments described herein. Additionally, it should be appreciated that, according to one aspect, one or more computer programs that when executed perform methods of this application need not reside on a single computer or processor but may be distributed in a modular fashion among a number of different computers or processors to implement various embodiments of this application. An example of a program that may be implemented on a computer described herein is a program that implements an optimal transport model.

[0056] As used herein, the term “biological sample” refers to a subset of its tissues, cells or component parts (e.g. body fluids, including but not limited to peripheral blood, serum, plasma, ascites, urine, cerebrospinal fluid (CSF), sputum, saliva, bone marrow, synovial fluid, aqueous humor, amniotic fluid, cerumen, breast milk, broncheoalveolar lavage fluid, semen, prostatic fluid, cowper's fluid or pre- ejaculatory fluid, sweat, fecal matter, hair, tears, cyst fluid, pleural and peritoneal fluid, pericardial fluid, lymph, chyme, chyle, bile, interstitial fluid, menses, pus, sebum, vomit, vaginal secretions, mucosal secretion, stool water, pancreatic juice, lavage fluids from sinus cavities, bronchopulmonary aspirates, blastocyl cavity fluid, and umbilical cord blood). A biological sample further may include a homogenate, lysate or extract prepared from a whole organism or a subset of its tissues, cells or component parts, or a fraction or portion thereof, including but not limited to, for example, plasma, serum, spinal fluid, lymph fluid, the external sections of the skin, respiratory, intestinal, and genitourinary tracts, tears, saliva, milk, blood cells, tumors, organs. A biological sample further refers to a medium, such as a nutrient broth or gel, which may contain cellular components, such as proteins or nucleic acid molecule. A biological sample may also refer to intact cells (e.g., imaged in 2D or 3D). A biological sample may refer to an in vitro cell or in vitro cells (e.g., cultured cells or cultured cell lines).

[0057] The term “peak area” refers to the area under the curve of a peak of a spectrum (e.g., an IR or Raman spectrum). Peak area may be calculated by determining the area under the curve between an upper limit and lower limit of the spectrum (e.g., between an upper vibrational frequency limit and a lower vibrational frequency limit).

[0058] Detailed Description

[0059] Described herein are methods of analyzing a biological sample. The methods include obtaining one or more infrared images of the biological sample using infrared microscopy and / or one or more Raman images of the biological sample using Raman microscopy; and applying, by a computer, a machine learning model to analyze the one or more infrared images and / or the one or more Raman images. In particular instances, the methods described herein are useful for predicting drug responses in in vitro cells (e.g., cultured cells, cultured cell lines). For example, the methods described herein are useful for predicting the response of in vitro cells to a drug or a drug combination. The methods described herein may also be useful for determining the mechanism of action, the dosage, and / or target of a drug or a drug combination when administered to a biological sample (e.g., in vitro cells).

[0060] Also described herein are methods of identifying a biological state in a biological sample. The methods include obtaining one or more infrared images of the biological sample using infrared microscopy and / or one or more Raman images of the biological sample using Raman microscopy; and applying, by a computer, a machine learning model to analyze the one or more infrared images and / or the one or more Raman images, wherein an output of the analysis indicates the biological state.

[0061] Also described herein are methods of determining the efficacy of an agent in effecting a change in a biological sample. The methods include obtaining one or more infrared images and / or one or more Raman images of the biological sample after administration of the agent, wherein the one or more infrared images are obtained using infrared microscopy and / or the one or more Raman images are obtained using Raman microscopy; and applying, by a computer, a machine learning model to analyze the one or more infrared images and / or the one or more Raman images, wherein an output of the analysis indicates that the agent is effective in effecting a change in the biological sample.

[0062] Also described herein are systems for implementing the methods described herein. The systems include an infrared imaging component configured for infrared microscopy and / or a Raman imaging component configured for Raman microscopy; and a computer configured to receive images from the infrared imaging component and / or images from the Raman imaging component, wherein the infrared imaging component and / or the Raman imaging component are each operatively coupled to the computer, wherein the computer is programed with a machine learning model whose input data is the images from the infrared imaging component and / or the Raman imaging component.

[0063] Also described herein are methods of predicting the response of a subject to a drug. The methods include obtaining one or more infrared images and / or one or more Raman images of a biological sample after administration of the drug, wherein the biological sample was obtained from the subject, and wherein the one or more infrared images are obtained using infrared microscopy and / or the one or more Raman images are obtained using Raman microscopy; and applying, by a computer, a machine learning model to analyze the one or more infrared images and / or the one or more Raman images, wherein an output of the analysis is predictive of the response of the biological sample to the drug.

[0064] Also described herein are methods of predicting the response of a subject to a drug combination. The methods include obtaining one or more infrared images and / or one or more Raman images of a biological sample after administration of the drug combination, wherein the biological sample was obtained from the subject, and wherein the one or more infrared images are obtained using infrared microscopy and / or the one or more Raman images are obtained using Raman microscopy; and (b) applying, by a computer, a machine learning model to analyze the one or more infrared images and / or the one or more Raman images, wherein an output of the analysis is predictive of the response of the biological sample to the drug combination.

[0065] The methods may include adding one or more infrared probes and / or one or more Raman probes prior to obtaining the one infrared images and / or the one or more Raman images. The one or more infrared probes and / or the one or more Raman probes may include a probe for detecting unsaturated fatty acid uptake. An exemplary probe for detecting unsaturated fatty acid uptake may be ds4 oleic acid. The one or more infrared probes and / or the one or more Raman probes may include a probe for detecting saturated fatty acid uptake. An exemplary probe for detecting saturated fatty acid uptake may be azido palmitic acid. The one or more infrared probes and / or the one or more Raman probes may include a probe for detecting protein synthesis. An exemplary probe for detecting protein synthesis may be13C labeled amino acids. In a preferred example, the infrared microscopy and / or Raman microscopy are performed without infrared probes or Raman probes. In a preferred example, the one or more infrared images and / or the one or more Raman images are obtained without addition of any infrared probes or Raman probes to the biological sample.

[0066] In any of the methods described herein, each of the one or more infrared images may include a collection of pixels, wherein each pixel includes an infrared spectrum. In any of the methods described herein, each of the one or more Raman images may include a collection of pixels, wherein each pixel comprises a Raman spectrum.

[0067] A method described herein may further include determining a chemical composition of the biological sample at each pixel.

[0068] In any of the methods described herein, the one or more images may be 3D images. In any of the methods described herein, the 3D images may include optical sections, and wherein the optical sections comprise a thickness of from about 5 pm to about 500 pm (e.g., about 5 pm to about 10 pm, about 5 pm to about 20 pm, about 5 pm to about 50 pm, about 10 pm to about 50 pm, about 10 pm to about 25 pm, about 50 pm to about 100 pm, about 50 pm to about 200 pm, about 100 pm to about 200 pm, about 100 pm to about 300 pm, about 200 pm to about 300 pm, about 300 pm to about 400 pm, about 400 pm to about 500 pm, about 5 irn to about 250 pm, about 250 pm to about 500 pm, about 5 pm to about 400 pm, about 5 pm to about 300 pm, about 5 pm to about 200 pm, about 5 pm to about 100 pm, about 100 pm to about 500 pm, about 200 pm to about 500 pm, about 300 pm to 500 pm, or about 400 pm to about 500 urn; e.g., about 5, 10, 15, 20, 25, 30, 40, 50, 60, 70, 80, 90, 100, 150, 200, 250, 300, 400, or 500 |jm).

[0069] In any of the methods described herein, the infrared microscopy may produce a spectrum from about 400 cm'1to about 4000 cm'1(e.g., about 500 cm'1to about 4000 cm'1, about 800 cm'1to about 4000 cm'1, about 1000 cm'1to about 4000 cm'1, about 1200 cm'1to about 4000 cm'1, about 1500 cm'1to about 4000 cm1, about 1800 cm1to about 4000 cm1, about 2000 cm1to about 4000 cm1, about 2200 cm'1to about 4000 cm1, about 2500 cm1to about 4000 cm1, about 2800 cm1to about 4000 cm1, about 3200 cm1to about 4000 cm1, about 3500 cm1to about 4000 cm1, about 400 cm1to about 3500 cm'1, about 400 cm1to about 3200 cm1, about 400 cm1to about 3000 cm1, about 400 cm1to about 2800 cm'1, about 400 cm'1to about 2500 cm'1, about 400 cm'1to about 2200 cm'1, about 400 cm'1to about 2000 cm'1, about 400 cm'1to about 1800 cm'1, about 400 cm'1to about 1500 cm'1, about 400 cm'1to about 1200 cm'1, about 400 cm'1to about 1000 cm'1, about 400 cm'1to about 800 cm'1, about 1000 cm'1to about 2000 cm'1, about 1000 cm'1to about 3500 cm'1, about 2500 cm'1to about 3500 cm'1, about 500 cm'1to about 1500 cm1, about 500 cm1to about 3000 cm1, about 1000 cm1to about 2000 cm1, about 2000 cm1to about 3000 cm1, about 1500 cm1to about 3000 cm1, about 800 cm1to about 1000 cm'1, about 1000 cm1to about 1300 cm1, about 1100 cm1to about 1400 cm1, about 1300 cm'1to about 1500 cm'1, about 1400 cm'1to about 1600 cm'1, about 1500 cm'1to about 1800 cm'1, about 1800 cm'1to about 2000 cm'1, about 2000 cm'1to about 2200 cm'1, about 2000 cm'1to about 2300 cm'1, about 1000 cm'1to about 2300 cm'1, or about 1000 cm'1to about 3100 cm'1) and / or the Raman microscopy may produce a spectrum from about 500 cm'1to about 3500 cm'1(e.g., about 800 cm'1to about 3500 cm1, about 1000 cm1to about 3500 cm1, about 1200 cm1to about 3500 cm1, about 1500 cm'1to about 3500 cm1, about 1800 cm1to about 3500 cm1, about 2000 cm1to about 3500 cm1, about 2200 cm1to about 3500 cm1, about 2500 cm1to about 3500 cm1, about 2800 cm1to about 3500 cm1, about 3200 cm1to about 3500 cm1, about 500 cm1to about 3200 cm1, about 500 cm1to about 3000 cm'1, about 500 cm'1to about 2800 cm'1, about 500 cm'1to about 2500 cm'1, about 500 cm'1to about 2200 cm'1, about 500 cm'1to about 2000 cm'1, about 500 cm'1to about 1800 cm'1, about 500 cm'1to about 1500 cm'1, about 500 cm'1to about 1200 cm'1, about 500 cm'1to about 1000 cm'1, about 500 cm'1to about 800 cm'1, about 1000 cm'1to about 2000 cm'1, about 2000 cm'1to about 3000 cm'1, about 1500 cm1to about 3000 cm1, about 800 cm1to about 1000 cm1, about 1000 cm1to about 1300 cm'1, about 1100 cm1to about 1400 cm1, about 1300 cm1to about 1500 cm1, about 1400 cm'1to about 1600 cm1, about 1500 cm1to about 1800 cm1, about 1800 cm1to about 2000 cm1, about 2000 cm1to about 2200 cm1, about 2000 cm1to about 2300 cm1, about 1000 cm1to about 2300 cm'1, or about 1000 cm'1to about 3100 cm'1). For example, the spectrum may be from about 900 cm'1to about 1800 cm'1.

[0070] In any of the methods described herein, the one or more infrared images and / or the one or more Raman images may include one or more segmentation images at a spectrum from about 2830 cm'1to about3000 cm1. In any of the methods described herein, computer may apply a cell segmentation method to the one or more segmentation images to generate an output of a segmented cell. The cell segmentation method may be an Otsu thresholding method. In an exemplary implementation of the Otsu thresholding method, an image is divided into areas of foreground and background based on a moving threshold value. For every possible threshold value, the variance of the foreground and the background is calculated. Otsu thresholding finds the value of the threshold that minimizes the weighted sum of the variances of the foreground and background. For example, in some segmentation methods, the average infrared spectrum of the segmented cell may be normalized by scaling the area of the average infrared spectrum to 1 ; and / or the average Raman spectrum of the segmented cell may be normalized by scaling the area of a subspectrum of from about 2815 cm1to about 3015 cm1to 1 .

[0071] In any of the methods described herein, each of the one or more infrared images may include a collection of pixels, wherein each pixel includes an infrared spectrum, and / or each of the one or more Raman images may include a collection of pixels, wherein each pixel includes a Raman spectrum. Furthermore, the infrared spectra of a subset of the collection of pixels of the one or more infrared images including the segmented cell may be averaged to generate an average infrared spectrum of the segmented cell; and / or the Raman spectra of a subset of the collection of pixels of the one or more Raman images including the segmented cells may be averaged to generate an average Raman spectrum of the segmented cell.

[0072] In any of the methods described herein, the machine learning model may be developed using a combination of one or more reference infrared images and / or one or more reference Raman images. In any of the methods described herein, the machine learning model may be an optimal transport model.

[0073] In any of the methods described herein, the one or more reference infrared images and / or the one or more reference Raman images may have been obtained using infrared microscopy and / or Raman microscopy, respectively, from a reference sample comprising reference cells having reference cell states.

[0074] In any of the methods described herein, the optimal transport model may be applied to the one or more infrared images and / or the one or more Raman images by minimizing the cost function for transforming the infrared spectra and / or the Raman spectra of the biological sample to an infrared spectra and / or a Raman spectra of the reference cells. As an example, the reference cells may include cells of the biological sample prior to addition of an agent for effecting a change. As an example, an optimal transport model described herein may be applied to one or more infrared images and / or one or more Raman images obtained from a biological sample prior to the addition of an agent to the biological sample and to one or more infrared images and / or the one or more Raman images obtained from the biological sample after the addition of the agent to the biological sample, e.g., to determine any changes in a cell state (e.g., cell profile) of any cell within the biological sample. Cell states (e.g., cell profiles) determined by any of the methods described herein may include expression states, epigenetic states, metabolic states, or a combination thereof. A particular example of cell states is metabolic states (e.g., the production of carbohydrates, lipids (e.g., saturated or unsaturated fatty acids), and / or proteins).

[0075] In any of the methods described herein including an agent, the agent may be a small molecule, a nucleic acid, a protein, a gene editing agent, or a cell. Examples of agents which are nucleic acids include an RNAi, an shRNA, an mRNA, or an antisense oligonucleotide. Examples of agents which are proteins include a transcription factor, an antibody, a peptide, a cytokine, a hormone, an enzyme, an antibody drug conjugate, or a fusion protein. Examples of agents which are gene editing agents include clustered regularly interspaced short palindromic repeats (CRISPR), a zinc finger nuclease, a base editor, a prime editor, or transcription activator-like effectors (TALEN). Examples of agents which are cells include a chimeric antigen receptor T (CAR-T) cell.

[0076] In any of the methods described herein, the biological sample may include a cell, a tissue, and / or a cellular component. The biological sample may have been obtained from a human. The biological sample may have been obtained from a subject with a disorder (e.g., a cancer; e.g., a breast cancer). The disorder may also be an infectious disease (e.g., resulting from a bacteria, virus, fungus, or parasite), a cardiovascular disease, a respiratory disease, a metabolic disorder, an endocrine disorder, a neurological disorder, a mental disorder, an autoimmune disease, a gastrointestinal disease, a rare disease. The biological sample may include human cancer cells. The biological sample may include human breast cancer carcinoma cells. In some instances, the biological sample includes an in vitro cell sample. The cell sample may include one or more cell lines (e.g., one or more cell lines suitable for modeling a disorder).

[0077] In any of the methods described herein including an effected change, the effected change may include treating a disorder, altering a gene or protein expression profile, inducing cell differentiation, inducing a cell state change, inducing cell death, or a combination thereof.

[0078] Any of the methods or systems described herein may include a computer whose input data may include images obtained using infrared microscopy and / or images obtained using Raman microscopy. The computer may have been programmed with a machine learning model developed using a combination of images obtained using infrared microscopy and / or Raman microscopy. The computer may be programmed to implement an optimal transport model.

[0079] A system described herein may include a Fourier transformed infrared (FTIR) spectrometer. A system described herein may include a Raman spectrometer. A system described herein may also include an irradiation source, a beam splitter, a moving mirror, and a fixed mirror. A system may be suitable for performing infrared microscopy, including any one or more of optical photothermal infrared (O- PTIR) microscopy, mid-infrared photothermal (MIP) microscopy, nano-FTIR spectroscopy, dual-comb photothermal microscopy, shortwave infrared photothermal (SWIP) microscopy, quantum cascade lasers (QCLS) based IR microscopy. A system described herein may be suitable for performing Raman microscopy, including stimulated Raman scattering (SRS) microscopy.

[0080] Cell lines and materials that may be used in any of the methods described herein include: MDA- MB-231 (ATCC HTB-26) which may be purchased from ATC). For reagents, azido-palmitic acid (1346) which may be purchased from Click chemistry tools; algal amino acid mixture (U-13C, 97-99%, CLM- 1548) which may be purchased from Cambridge; deuterated oleic acid (683582) which may be purchased from Sigma-Aldrich. For drugs, anisomycin (A9789), bortezomib (179324-69-7), cycloheximide (01810), emetine (SMB01061 ), everolimus (94687), epirubicin hydrochloride (E9406), gefitinib (184475-35-2), lapatinib (231277-92-2), TVB-3166 (SML1694), taxol (PHL89806), vincristine sulfate (V8388) and apicidin (A8851 ) may be purchased from Sigma-Aldrich. Daunorubicin hydrochloride (AAJ60224MA), iniparib (AC469161000), triacsin C (24-721 -00U), doxorubicin hydrochloride (BP25165), MG 132 (AAJ63250LB0), dactolisib (NC0298104) may be purchased from Fisher Scientific. For cell culture agents, DMEM medium (11965), FBS (10082), penicillin / streptomycin (1514), may be purchased from ThermoFisher Scientific. CaF2 substrates (CAFP13-1 ) may be purchased from Crystran. Probes that may be used in any of the methods described herein may be prepared according to the following methods: Azido palmitic acid-bovine serum albumin (BSA) solution: For the solution, couple azido-palmitic acid with BSA to prepare a 2-mM stock solution. Prepare 20 mM sodium palmitic acid solution by dissolving palmitic acid in NaOH solution with the following recipe: azido-PA (5.5 mg) + 1 .0 ml dd-H2O + 35 pl 1 M NaOH. Mix and incubate the solution in 70 °C water baths until no oil droplets are visible. Then slowly add the sodium palmitic acid solution into 2.7 ml 20% BSA under room temperature water baths. Quickly add 6.3 ml DMEM culture medium and filter the solution with a 0.22-pm sterile filter.13C-amino acids DMEM: 4 mg ml-1algae13C-amino acids mix was dissolved in dd-H2O with 10% FBS and 1 % penicillin, which matched the concentrations of regular amino acids in DMEM. Deuterated oleic acid-bovine serum albumin (BSA) solution: For the solution, couple d34-oleic acid with BSA to prepare a 2- mM stock solution. Prepare 20 mM oleic acid solution by dissolving oleic acid in NaOH solution with the following recipe: ds4 oleic acid (6.3 mg) + 1 .0 ml dd-H2O + 24 pl 1 M NaOH. Mix and incubate the solution in 70 °C water baths until no oil droplets are visible. Then slowly add the ds4 oleic acid solution into 2.7 ml 20% BSA under room temperature water baths. Quickly add 6.3 ml DMEM culture medium and filter the solution with a 0.22-pm sterile filter.

[0081] MDA-MB-231 cells can be cultured in DMEM media supplemented with 10% FBS and 1 % penicillin. Cells are grown in a humidified atmosphere containing 5% CO2 at 37 °C in the incubator. At -80% confluence, cells are dissociated with trypsin and passaged.

[0082] For cell viability assay and drug IC50 calculus, the IC50 values of the drugs can be determined by Alamar blue assay. Cells are seeded at 10,000 per well in 96-well plates. After 24 h, the cells are washed twice with phosphate buffered saline (PBS) and are treated with drugs at different concentrations in cell culture media for 48 hrs. Each drug concentration has 6-8 replicates. After the drug treatments, cells are washed with PBS twice and the cell viability is determined by Alamar blue assay following the manufacturer’s protocol (Invitrogen) using plate reader. The IC50 values are determined by fitting the data using a dose response model with variable Hill slope built in Prism.

[0083] For drug combination and synergy score calculus, cell viability assay is performed on cells treated by drug combinations to evaluate synergy scores. Everolimus-Lapatinib and Everolimus-Doxorubicin can be chosen as model systems to study drug combinations. Each drug concentration combination has 5 replicates. To calculate synergy scores, an open-source package SynergyFinder is applied. The synergy scores are calculated based on the ZIP model available in the package.

[0084] In the methods described herein, cells may be prepped using the following method: MDA-MB-231 cells are seeded on clean CaF2 substrates with 5*104cells per well in cell culture media (DMEM, 10% FBS, 1 % penicillin) overnight for control and other drug treatment conditions. Then the culture media is replaced by13C-amino acids DMEM with 50 M azido-palmitic acid, 50 M ds4 oleic acid, either single drug at its IC50 concentration or drug combinations at chosen concentrations. Drugs are prepared in 100% DMSO and diluted to 0.1 % DMSO in labeling media. For the control group, only cell labeling media with 0.1 % DMSO is added (without any drugs). Cells are treated for 48hrs. The culturing time is selected based on the trade-off between signal and experimental time. After that, cells are fixed by 4% PFA at room temperature for 15 min and washed three times with PBS buffer and five times with dd-H2O. The samples are then air-dried before FTIR imaging. To perform FTIR imaging, Agilent Cary 620 Imaging FTIR equipped with an Agilent 670-IR spectrometer and 128 x 128-pixels FPA mercury cadmium telluride (MCT) detector is used in the transmission mode. A background spectrum is collected on a clean CaF2 substrate using 128 scans at 8 cm-1spectral resolution, suggesting that the IR absorbance is measured every 4 cm1. Cell spectra are recorded using 64-128 scans at 8 cnr1spectral resolution. A x25 IR objective (pixel size, 3.3 pm, 0.81 numerical aperture (NA)) is used for cell imaging.

[0085] To perform spontaneous Raman imaging, spontaneous Raman imaging is performed using an upright confocal Raman microscope (Xplora, HORIBA Jobin Yvon). Cell samples are illuminated by 532 nm laser (80 mW on sample) through a 50x objective (air, NA 0.75, MPIan N, Olympus). Raman images are acquired using the point-by-point mapping mode with an acquisition time of 5s and 1 x accumulation for each point measurement. The step size is set as 7 pm. The grating is set as 1200 gr / mm. Both the slit size and the hole size is set as 100 pm. 500-1000 cells are imaged for each condition.

[0086] To process Raman images, raw Raman spectra are first processed in the LabSpec 6 software (HORIBA). The “Despike” function is used to remove cosmic rays. The “Threshold” function is used to remove spectra with high background. The “Correction” function is used to subtract background from raw spectra by selecting a field of view without any cells. After that the spectral range from 2830 cm1to 3000 cm’1is integrated to generate a cell image, using custom-written MATLAB scripts. The cell image is used to segment single cells and generate a cell segmentation mask with the Otsu thresholding method using the CellProfiler software. Based on the cell segmentation mask, single-cell spectra are generated by averaging all the spectra within each cell, using custom-written MATLAB scripts.

[0087] For calculation of the protein-to-lipid ratio, lipid unsaturation ratio, protein synthesis rate, saturated lipid synthesis rate, and unsaturated lipid synthesis rate in any of the methods described herein: The protein-to-lipid ratio is defined as the intensity at 2930 cm’1divided by the intensity at 2850 cm’1. The lipid unsaturation ratio is defined as the intensity at 3050 cm1divided by the intensity at 2850 cnr1. The protein synthesis rate is defined as the peak area of 1600-1630 cm1divided by the sum of the peak area of 1600-1630 cm’1and the peak area of 1634-1700 cm1. The saturated lipid synthesis rate is defined as the peak area of 2080-2130 cm1divided by the peak area of 2825-2875 cnr1. The unsaturated lipid synthesis rate is defined as the peak area of 2130-2230 cm’1divided by the peak area of 2825-2875 cm’1.

[0088] Any of the methods described herein may include normalization of the vibrational spectra (e.g., the IR or Raman spectra). Chemical imaging spectra typically have two major sources of technical noise that are removed with different normalization strategies. The first is a scattering effect that introduces a drift in the spectra and is removed with a baseline correction method. The second is introduced by the variable depth of the measured tissue and can be thought of as similar to the library size in transcriptomic settings. These may be removed by computing cell-specific normalization constant as the average over the whole spectra, or specific peaks. Five possible normalization methods include: “Amide I” normalizes the area of the Amide I peak (1600-1800) to 1 ; “Amide II” normalizes the area of the Amide II peak (1470- 1570) to 1 ; “CH” normalizes the area of the (a)symmetric CH stretching (2815-3015) to 1 ; “Signal” normalizes the area of the whole spectra to 1 ; and “Min-Max” linearly scales and shifts the whole spectra to lie in [0, 1 ]

[0089] In any of the methods described herein, IR datasets may be analyzed with the signal normalization and Spontaneous Raman datasets may be analyzed with the CH normalization. An exemplary machine learning model implemented by any of the computers described herein and / or used in any of the methods described herein is optimal transport (OT), a mathematical framework that aligns probability distributions. With this approach, observations of cells were viewed from a profiled population as samples from an underlying probability distribution. Given a control and perturbed population, the perturbation effect is modeled as one that transforms the distribution over control states into the distribution over treated states. Alternative functions could fulfil this transformation, for instance a trivial random assignment, but optimal transport identifies the transformation that alters cell states as efficiently as possible, with respect to some cost function. This principle of least effort is well-aligned with our assumption that perturbations only alter some targeted sub-programs of a cell as it essentially means that the coupling can be induced by utilizing the unaltered cellular features. An exemplary implementation of OT is as follows: let / z and v be the probability distributions underlying the control and treated cellular states and T: v be a function that maps cells from the control state to treated in the treated state. The optimal transport problem can be formulated as: where the pushforward constraint, T# / J. = v, ensures that T transforms / z into v.

[0090] The optimal transport problem can be solved via an efficient solution to the Kantorovich formulation, as implemented in the PythonOT package. This solution may induce soft assignments between control and treated cells, when a discrete assignment is necessary, for example when labels are propagated, a hard assignment may be induced by a weighted sample over all assigned states.

[0091] The machine learning model may be useful for modeling the change in one or more cells from one state to another (e.g., altering a gene or protein expression profile, inducing cell differentiation, inducing a cell state change, inducing cell death, or a combination thereof). The machine learning model may also be useful in determining the mechanism of action (MoA), suitable dosage, and / or target of a drug or drug combination. For example, determining the change in one or more cells from one state to another as a result of the administration of a drug or drug combination may indicate whether the drug or drug combination targets a carbohydrate synthesis pathway, a lipid synthesis pathway, a protein synthesis pathway, or an expression pathway. In addition, determining the degree of change in one or more cells from one state to another as a result of the administration of different dosages of a drug or drug combination may indicate the dosage necessary to effect a particular change.

[0092] Example 1. Predicting Biochemical Landscape and Dynamics to Perturbations at Single Cell and Molecular Resolution with Vibrational Imaging and Optimal Transport

[0093] Background

[0094] Dissecting the heterogeneous response of individual cells towards genetic and chemical perturbations is central to understanding dynamic functions of cells. Currently, imaging-based approaches to study these responses typically rely on fluorescence-based techniques that make measurements by tagging cells with chemical or genetically encoded fluorophores. While proven to be useful, these approaches can themselves disturb the cell state, and are further limited by fewer than five biological features they can measure. Here, VIB-OT, the first approach that integrates emerging molecular vibrational imaging techniques including high-dimensional infrared (IR) imaging and Raman imaging with optimal transport theory based in silico modeling, is introduced to dissect heterogeneous responses to perturbations at single cell and molecular resolution. These vibrational imaging techniques measure hundreds to a thousand of chemical activity features and molecular vibrational patterns of a cell in a scalable, low-cost, and non-invasive manner, resulting in high-dimensional biochemical phenotyping of cell states and minimal influences on the original cellular states. More than 20,000 perturbation responses of 16 different chemical treatments are profiled at single-cell resolution in human breast carcinoma cells. Vibrational imaging techniques that have minimal batch effects and capture expressive biochemical profiles and cell states are shown. Single-cell perturbation responses by computing optimal transport couplings are studied, which shows that not only can the mechanism of actions of different chemical perturbations be identified, but also molecular interpretation can be provided. Finally, this approach is demonstrated for its ability to predict and optimize the effects of chemical perturbation dosages and combinations. VIB-OT opens a new direction to dissect heterogeneous perturbation responses at single cell resolution and serves as a foundation to build virtual simulators of cells under massive perturbations through the utility of high-throughput, high-dimensional, and low-cost vibrational imaging and interpretable in silica modeling.

[0095] Understanding the effects of genetic or chemical perturbations on cellular state lies at the heart across every aspect of biology, ranging from understanding molecular functions of multicellular systems to the drug discovery pipeline. For instance, knowledge of how a cell's state and molecular programs are altered in response to specific genetic or chemical perturbations can be used to dissect gene regulatory networks underlying various functions of cells and pinpoint key programs or genes as effective drug targets. Furthermore, by understanding the influence of perturbation dosages and combinations, these perturbations can be controlled to fine-tune, enhance, or personalize their effects regulating cellular states and functions. A key challenge in the study of perturbations, however, lies in the heterogeneous nature of their responses, which can arise from complicating factors, such as dynamic and continuum intrinsic cell state, extrinsic cell-to-cel I signaling, and the cell’s local microenvironment, in addition to cell type. Here, cell type-based approaches, which rely on discretized cell types, begin to fail to model the influence of these key phenomena, and new analytical strategies will need to be developed to compensate. Thus, to characterize, model, and predict heterogeneous cellular perturbation responses as a dynamic continuum at single cell resolution, required are: I) profiling technologies capable of measuring nuanced individual cellular states through high-dimensional molecular and cellular features within their environmental contexts and II) scalable computational methodology to analyze, track, and predict the responses of individual cells within heterogeneous populations.

[0096] Traditional fluorescence-based optical imaging methods capture morphological information about cells and require fluorophores, which have broad spectral width and thus limit the number of measured reporters to be typically no more than five, which is known as the fluorescence color barrier. Furthermore, these methods rely on fluorescence labeling, which themselves introduce strong disturbance on cellular state.

[0097] Molecular vibrational imaging techniques are emerging profiling technologies that measure highdimensional biochemical profiles of individual cells. These techniques measure, at each pixel, the interaction of light with matter to identify and quantify the molecules and chemical bonds present in a sample. Different molecules and chemicals absorb and scatter light at characteristic wavelengths, creating unique high-dimensional fingerprints for molecular compositions, resulting in a composite signal of biochemical features in the corresponding vibrational spectrum. The measured vibrational features / peaks are associated with different biochemicals including proteins, lipids, nucleic acids, and carbohydrates. A key feature of vibrational imaging techniques is that not only do they provide a direct observation on a cell’s biochemical state at molecular resolution, but also they can do so in a label-free and non-invasive manner. Unlike traditional florescence-based optical imaging, label-free molecular vibrational imaging methods, on the other hand, profile rich spectral features with diverse biochemical information, enabling high-dimensional data acquisition and analysis, without labeling cells and altering the cellular state. In addition to the label-free mode, the measured metabolic sensitivity and specificity can be furthermore enhanced with the aid of vibrational probes, which are typically metabolites modified with small-size vibrational tags. These probes, when compared to fluorescent labels, have 50 to 100-fold narrower spectral width, allowing for multiplexed vibrational imaging beyond the color barrier of fluorescence imaging. Moreover, the sizes of vibrational probes are much smaller than fluorophores with minimal modification to the original metabolites, leading to minimal disturbance on cellular state.

[0098] Compared to fluorescence-based imaging with fewer than five color features, molecular vibrational imaging can profile 100 to 200 times more molecular features. While vibrational imaging techniques provide a unique high-dimensional profiling of a cell’s biochemical activities, their application to study perturbation responses has not been reported till recently when we have developed VIBRANT, which is the first demonstration and application, to measure large-scale single-cell chemical perturbation responses based on one type of vibrational imaging, Fourier transformed IR (FTIR) imaging, with multiplexed vibrational probes. The cell spectral profile measured by VIBRANT was demonstrated to be highly sensitive to phenotypic changes under chemical perturbation, making it suitable for predicting mechanism of action (MoA) of chemicals and identifying novel perturbation candidates. This represents a completely new phenotyping modality in dissecting perturbation responses in multicellular systems and drug discovery pipelines. Despite its promises, relevant computational methods, especially for analyzing, tracking, and predicting the biochemical activity responses of the cells within complex populations have not been developed.

[0099] To analyze, track, and predict the biochemical activity responses of the cells within complex populations, one way to approach this problem is to consider a profiled population as a probability distribution over the high-dimensional space of its molecular activity, and reason about how this distribution changes upon perturbations. Within this view, optimal transport theory is a rich and powerful mathematical framework that describes how a pair of probability distributions are transformed from one into the other according to a principle of minimum action. Optimal transport was applied to single-cell biology contexts to reconstruct developmental trajectories and predict genetic and chemical perturbation responses. While a classical cell-typing approach would first cluster cells into cell types and then assume a homogeneous responses within each cell type, when equipped with an optimal transport approach, a perturbation’s effect can be determined not as piecewise linear approximations over discrete cellular types, but instead capture the full heterogeneity and continuum of a perturbation by considering the response over the distribution of individual cellular states given distinct cell intrinsic and extrinsic factors of individual cells. Further, these optimal transport couplings can be computed in a pairwise manner and chained across a series of conditions to describe, for example, how responses act as a function of a perturbation’s dosage, or to learn how they are modulated as the ratio of perturbation change in a combination experiment.

[0100] A first integrated experimental and computational framework, named, Vibration-Optimal Transport (VIB-OT), was developed to model and predict perturbation effects, dosages, and combinations through two vibrational imaging technologies, infrared (FTIR) and Raman imaging, and optimal transport theory based in silico modeling. More than 20,000 perturbation responses of 16 different chemical treatments were profiled at single-cell resolution in human breast carcinoma cells. Using optimal transport analysis, the full heterogeneity of individual cellular responses were recovered in IR imaging data and was demonstrated in an analytical context to dissect molecular activities and mechanisms of actions after perturbations. Furthermore, the effects of perturbation doses and combinations were determined by computing and chaining optimal transport couplings across a series of different dosages. Finally, the modularity of VIB-OT was determined by analyzing the response of cells profiled with another vibrational imaging technique spontaneous Raman imaging, demonstrating the full combability of VIB-OT in versatile vibrational imaging techniques. Using VIB-OT, subpopulations in chemical treated breast cancer cells were discovered, with specific heterogeneity information. In addition, multiple molecular responses were directly visualized and quantified at various chemical treatments, providing important insights of their mechanism of actions (MoAs), therapeutic effects and toxicity at single-cell level. VIB-OT opens a new venture to dissect cellular functions through various label-free vibrational imaging and in silico modelling, paving the way to build virtual simulators to understand biochemical activities and functions of physical cells.

[0101] Results

[0102] VIB-OT framework

[0103] VIB-OT (FIG. 1), a vibrational imaging-based approach, was developed to dissect the perturbation effects on individual cells and explore its application to chemical perturbation using in vitro cellular model. Besides the biochemical bonds and molecular activities measured in label-free manner, vibrational probes that are both IR and Raman active were designed to report more specific metabolic activities as introduced in the recent VIBRANT approach. Three distinct vibrational probes were used to measure three essential metabolic activities, including13C amino acids (13C-AA) labeling for protein synthesis, azido palmitic acid (azido-PA) for saturated fatty acid uptake and deuterated oleic acid (d34-OA) for unsaturated fatty acid uptake. Due to the narrow spectrum width of these vibrational probes, many more probes (on the scale of 10-100) could be introduced to measure specific biochemical features and activities.

[0104] Once two or more conditions imaged were imaged through either IR or Raman imaging, computational methods were used to reason about the individual cellular responses. Currently, it was assumed that the same cell can only be profiled once, e.g., in either a control or perturbed state. We then recover perturbation effects from the perturbed and control populations by learning to couple them with optimal transport. Ultimately, this approach identifies pairs of cells between control and perturbed cells. Individual cellular responses are computed by taking the difference of the coupled treated and control spectra. Perturbation effects can then be interpreted by the peaks in the resulting spectrum. Modeling chemical perturbation responses through vibrational imaging and optimal transport

[0105] The responses in chemical activity to sixteen drugs across nine methods of actions analyzed with VIB-OT. Cells are treated with each of the drugs and a DMSO control and then imaged by Fourier transform infrared (FTIR) imaging. Single-cell responses are recovered by computing a coupling between the distribution of control and treated cells, individually for each drug (FIG. 2A - FIG. 2C).

[0106] First, VIB-OT to was applied to IR imaging. VIB-OT enables the analysis of chemical perturbation effects at a single-cell level. The data reported in VIBRANT (Liu, X., et al. Nat Methods. 21 :501 — 511 , 2024.), which collects the single-cell data of MDA-MB-231 , a human breast cancer carcinoma cell line, treated with DMSO control and 16 drugs across nine distinct drug MoA classes at IC50 concentrations, was analyzed. These 16 drugs were separated into two main categories: those inhibiting general metabolism (e.g., protein synthesis inhibition, DNA inhibition) and those targeting specific molecular pathways (e.g., mTOR / PI3K inhibition, EGFR inhibition). The specific MoAs are shown in FIG. 2A. An optimal-transport coupling was calculated between the perturbed and control states individually for each treatment, yielding assignments from control cells to perturbed cells. Then, single-cell perturbation responses were computed as the difference between these coupled cells, i.e. , the perturbed state spectrum minus the control state spectrum, which represents the difference of biochemical compositions or metabolic activities between the perturbed cells and control cells. This analysis method better revealed the actual spectrum responses of cells upon chemical perturbations, compared with directly using the vibrational spectrum. Line plots over these individual cellular responses are shown in FIG. 2A, where upward peaks indicate elevated absorbance compared with the control group and backward peaks represent decreased absorbance.

[0107] By identifying several regions of known biological annotations, the measured response spectra became interpretable, reflecting the global compositions of biomolecules including glycogen (carbohydrates), amide II (proteins), phosphate vibrations (nucleic acids) and carbonyl bonds (lipids) in the fingerprint region. Furthermore, the use of vibrational probes allowed for detailed analysis of specific cellular metabolic activities, including protein synthesis rate and fatty acids uptake rates. These regions are highlighted by different colors for better differentiation in FIG. 2A, with specific spectrum ranges concluded in the SI. Through this identification, the responses of different macromolecules could be clearly grouped and visualized. For instance, the fatty acid uptakes of breast cancer cells were consistently reduced compared to the control under all chemical treatments, though to varying degrees, indicating altered metabolism under chemical perturbations. The averaged molecular responses of cells under chemical perturbations are shown in FIG. 2B, where red colors indicate elevated absorbance / concentrations compared with the control group while blue color represents decreased absorption. This visualization highlights the molecular composition shifts in cells subjected to various chemical perturbations. For instance, small molecules that inhibit protein synthesis (e.g., anisomycin, cycloheximide, and emetine) show a pronounced increase (0.001 -0.002 absorbance units (a.u.)) in the amide I region while exhibiting a decrease (-0.0007 to -0.001 a.u.) in the amide I probe region (signal derived from13C-AA). This results in a splitting of the original amide region, suggesting a marked decrease in protein synthesis during these drug treatments compared to the control state. This is because the signal from13C-AA reflects newly synthesized proteins, and the sum of these two amide peaks, representing total protein concentration, should remain consistent across cell states after spectrum normalization. Beyond protein composition, protein synthesis inhibition led to a slight increase (4.95e-5 to 1 .5e-4 a.u.) in glycogen concentration, while fatty acid uptake remained less affected compared to the control state. This biomolecular characterization using VIB-OT provides a direct visualization of molecular composition shifts in response to different chemical perturbations.

[0108] Considering the significant change of spectrum region related to protein compositions under various chemical perturbation, as shown in FIG. 2B, protein synthesis rate was further calculated and ranked across all drug treatments (FIG. 2C). Protein synthesis rates were quantified as the ratio of the13C-AA probe region to the “amide I” region. Small molecules that directly inhibit protein synthesis exhibited the lowest synthesis rates, consistent with their mechanisms of action (MoAs). For compounds that inhibit protein degradation (bortezomib, MG132), the observed reduction in protein synthesis rate was less pronounced, suggesting these agents exert a milder inhibition of protein synthesis. Additionally, compounds not directly targeting proteins also affected protein synthesis rates, highlighting VIB-OT's sensitivity to diverse chemical perturbations. For instance, DNA intercalators (doxorubicin, daunorubicin, epirubicin) and EGFR inhibitors (gefitinib) caused significant reductions in protein synthesis rates. These findings align with their known mechanisms: DNA intercalation disrupts macromolecule synthesis (Sci. Rep. 8:13672, 2018), while EGFR inhibition disrupts downstream signaling pathways critical for activating protein translation machinery (C / / n Cancer Res, 24(17_Supplement):B31 , 2018]. Other compounds, such as those targeting microtubules or the mTOR / PI3K pathway, showed more subtle effects on protein synthesis in MDA-MB-231 cells. Triacsin C, a drug inhibiting lipid metabolism, led to an elevated protein synthesis rate (~0.1 in ratio) compared to the control group, indicating a shift in anabolic processes towards protein synthesis under lipid metabolism inhibition. Furthermore, iniparib, a chemical with an unclear MoA, displayed strong inhibition of protein synthesis, suggesting its potential regulatory effect on protein synthesis in breast cancer cells.

[0109] Overall, the ranking of molecular responses obtained using VIB-OT offers a direct and quantitative method for comparing the effects of various drugs. This approach provides new insights into their MoAs and expands the understanding of drug-induced cellular responses.

[0110] Discovering heterogeneity of molecular responses under chemical perturbation at single cell resolution

[0111] VIB-OT identifies heterogeneity in the metabolic activity of anisomycin-treated cells. Via the induced control-treated coupling, this heterogeneity could be propagated into the control state, where markers could be identified for this differential response. Heterogeneity of the protein synthesis rate in response to anisomycin is linked to lipid activity in the control state (FIG. 3A - FIG. 3C).

[0112] Perturbation responses are often heterogeneous at single cell level. VIB-OT was applied to investigate heterogeneous responses to each small molecules in breast cancer cells. Among 16 small molecules, anisomycin treated breast cancer cells show great heterogeneity.

[0113] Single-cell responses were recovered by computing optimal transport couplings between the control and anisomycin treated states. Given the responses at a single-cell resolution, we can investigate heterogeneity therein. For anisomycin, we found a heterogeneity in how the protein synthesis rate responded to the treatment. FIG. 3B, left shows a joint plot of the coupled control and treated states of the protein synthesis rate. While control states are uni-modal, the distribution of treated states was more complex, suggesting varied responses of MDA-MB-231 cells under the treatment of anisomycin. A clustering of the treated states was computed according to their protein synthesis rates and label coupled cells accordingly (orange or blue in FIG. 2B, right). Finally, we can investigate potential explanations for this heterogeneity in response by propagating the label that was computed on the treated states to the control state using the coupling. FIG. 3C shows the distribution of control state in several regions of interest, grouped by the cluster ID of the coupled treated state. Here cells with an initially low lipid activity ultimately were paired with treated states that have decreased protein synthesis rates. While there does not appear to be any specific studies revealing heterogeneous response of anisomycin in MDA-MB-231 human breast cancer carcinoma, this observed molecular heterogeneity may be attributed to anisomycin’s JNK activation mechanism in MDA-MB-231 cells, which is known to promote apoptosis. The intrinsic variability in JNK expression within the MDA-MB-231 cell population could underlie the metabolic heterogeneity observed in response to anisomycin treatment.

[0114] Overall, the use of optimal transport in VIB-OT provides a robust method to identify and quantify heterogeneous molecular response at single-cell level upon drug treatment, showing great promise in evaluating drug effects and assessing drug resistance.

[0115] VIB-OT mitigates batch effects

[0116] One of the major challenges in imaging-based phenotyping in in vitro cellular systems lies in the strong batch effects, which are often stronger than the perturbation effects. An important feature of VIB- OT is its ability to profile cells in a minimally invasive and perturbed manner. Traditional imaging methods rely on techniques like staining or destructive imaging techniques that introduce strong perturbations to cells when they are measured (Loo, L.-H. et al., Nat. Methods 4:445-453, 2007; Bendall, S.C. et al., Science. 332:687-696, 2011 ). These perturbations represent sources of strong technical noise that result in batch effects that can be challenging to correct. For instance, Cell Painting (Bray, M.-A. et al., Nat. Protoc. 11 :1757-1774, 2016) is known to have batch effects and plate-layout effects. Here, to overcome these challenges, a correction scheme was used that can mitigate batch effects from VIB-OT and was shown to be robust to the treatment applied and normalization strategy.

[0117] The batch effect correction strategy is similar to popular linear-mixed model approaches, however statistics over fixed variables are computed and removed explicitly. As datasets grow and experimental settings become more complex, linear mixed models may be substituted with likely similar behavior. First, from each treatment and batch pair we compute and remove the average control signal (FIG. 4A) computed in that batch to yield a signal composed of the treatment and batch effects. Next, the average treatment effect across all batches was computed and removed, which results in a signal composed of the batch effects, residual heterogenous treatment effects and / or observational noise. Finally, the batch effects were computed as the average within each of the resulting (treatment, batch) pairs. Once computed, this linear correction was subtracted from all cells in the experiment, given their (treatment, batch) information. To avoid introducing bias, and unless stated otherwise, all statistics were computed on one split of the dataset. The correction method was evaluated on heldout cells, i.e. those not involved in computing the correction.

[0118] The mean Silhouette coefficient was used to measure the success in the removal of batch effects. Given a dataset of labelled cells, the Silhouette coefficient is computed for each cell as the difference in means between the intra-label distance, i.e. average distance to all points with the same label, and the distance to the nearest label the sample is not a part of. This distance is normalized to (-1 , 1 ) with the maximum of these two statistics. Values near zero imply well-mixed, overlapping clusters, positive values imply more distinct clusters, and negative values imply that the sample is mis-labeled. To measure the quality of the batch correction, for each treatment, the Silhouette coefficient was computed using the batch labels, where values near zero indicate that the batches are well-mixed and the correction is successful. Since a trivial (over-)correction can be achieved by removing all signal from the data, the batch silhouette score was supplemented with an effect silhouette score computed, for each treatment, between control and treated cells. Positive values of the effect silhouette score indicate that the control and treated states are distinct. Note that this statistic is influenced by the strength of the treatment, as the treated states of subtle effects will naturally intermix more with control states. Since large portions of the spectra remain unchanged by treatment effect, before computing the metric, the spectra are centered with respect to the average control spectra of each batch.

[0119] The batch and effect Silhoutte scores are plotted in FIG. 4B, for each treatment and several normalization strategies. The metrics computed on the raw data, the normalized uncorrected data, and finally the normalized and corrected data were reported. The initial baseline-corrected data has relatively larger batch effects and smaller treatment effects. Normalization tended to increase the magnitude of the treatment, with respect to the control state, but batch effects were still retained. Finally, the correction approach mitigated batch effects while retaining the treatment effect sizes. The coupled single-cell response spectra computed on the signal-normalized batch-corrected data were plotted for all conditions in FIG. 4C, where consistent effects were recovered across batches. The Silhoutte score metrics were further supplemented with a neighbor enrichment analysis. For each treatment, cells were sampled from each batch to equal proportion and the nearest neighbors are computed using the held out batch corrected cells. The kNN enrichment was computed as the fraction of cells in each query cell’s top k=25 (top), k=50 (bottom) neighbors that come from the same batch as the query cell (FIG. 4D) using the k- nearest neighbors (k-NN) algorithm. All conditions performed close to the ideal enrichment (33%, dashed line) indicating that the correction scheme was robust. Finally, in FIG. 4E, Uniform Manifold Approximation and Projection (UMAP) projections computed on the response spectra were computed for each condition, colored by their batch label, demonstrating in general, good inter-mixing.

[0120] While there does not yet exist an agreed-upon method of action for iniparib, VIB-OT provides us with the tools to interpret its signature with respect to the cell’s chemical activity. Like protein inhibitors, we observe a decrease in the rate of protein synthesis, however we also observe a decrease in glycogen, DNA / RNA, and the uptake of saturated fatty acids, and in increase in amide I activity (FIG. 5).

[0121] VIB-OT uncovers dosage effects and trajectories

[0122] Understanding the impact of chemical concentration on cellular function perturbation is an important aspect to optimize therapeutic efficacy, provide the molecular mechanisms of small molecules and evaluate safety and toxicity. As the dosage of a chemical increases, its effect ranges from minimal to toxic. Nevertheless, the detection of dosage effects requires both high sensitivity and high signal-to-noise ratio of measuring methods. Demonstrated here is the ability of VIB-OT to detect changes in the effect of a drug as the dosages increase. MDA-MB-231 cells were treated with a control and increasing concentrations of anisomycin and tvb-3166. Then, couplings between all consecutive dose concentrations were calulated, e.g. from control state to anisomycin 50 nM, then anisomycin 50 nM to anisomycin 1 pm, and etc. VIB-OT was used to determine how cellular processes were disrupted as a function of the concentration of drugs, using anisomycin as an example (FIG. 6A). At first, in the low concentration of anisomycin (50 nM), negligible absorbance change was observed across the spectra, outside of small changes in the amide I region, suggesting that only protein synthesis rate was decreased at this concentration, which was consistent with the MoA of anisomycin of protein synthesis inhibition. Next, at the 1 pM concentration, in relation to the 50nM states, strong alteration were observed in the amide I region (133% increase of 12C amide I and 89% decrease of 13C amide I probe) and slightly decreased absorbance in the glycogen, nucleic acids, saturated & unsaturated fatty acid uptake regions were observed, indicating that other molecular responses were also weakly affected besides protein composition at this concentration. Finally, at the highest considered concentration, 5 pM, in relation to the 1 uM states, stronger reductions were observed in the glycogen (124% decrease), nucleic acids (165% decrease), and saturated (295% decrease) & unsaturated fatty acid uptake (260% decrease) regions, while the activity in amide I region seemed to no longer increase. This result suggested that under high concentration of anisomycin, various metabolic activities were inhibited besides protein synthesis. For the tvb-3166 (FASN inhibitor) trajectory (FIG. 6C), minimal change was observed in the amide I peak at the start, then, in the second concentration 10 uM, stronger alteration in amide I region as well as reduction in glycogen, nucleic acid, lipid, saturated & unsaturated fatty acid uptake regions. Finally at the highest concentration, clear trends were no longer observed, but a high degree of noise was instead seen, suggesting that 10 pM was already a toxic concentration for MDA-MB-231 cells under TVB-3166 treatment. The specific quantification of molecular response changes at different dosages are shown in FIG. 6E and FIG. 6F.

[0123] The heterogeneity of molecular responses at different dosages for both anisomycin and TVB- 3166 treatment were further analyzed in FIG. 6B and FIG. 6D. The altered distribution clearly demonstrated different molecular heterogeneity at varied dosages. For instance, for anisomycin, the glycogen composition distribution remained similar at low and middle concentrations, while it was significantly shifted with increasing heterogeneity under high concentration treatment (5uM). On the other hand, the saturated fatty acid uptake (azido-PA) had different distributions under three concentrations and the heterogeneity was decreased under high concentration, suggesting low and more unified fatty acid uptake at high concentration. For TVB-3166, clearly different distributions were observed for both lipid and saturated fatty acid uptakes at different concentrations, which aligned with its MoA of FASN inhibition.

[0124] In conclusion, this was the first time that the dosage effects of anisomycin and TVB-3166 were quantified with heterogeneity information across multiple molecular responses in MDA-MB-231 human breast cancer carcinoma cells. These results suggest that not only does VIB-OT have sufficient sensitivity to detect dosage effects and trajectories, but also it can provide additional molecular activity information, providing important insights into chemical MoAs and toxicity.

[0125] VIB-OT is generalizable and modular across different vibrational imaging technologies

[0126] To demonstrate the generality of VIB-OT, we next performed Raman microscopy of drug-treated cells and analyzed drug response. Compared to FTIR, Raman microscopy provides complementary vibrational information, higher spatial resolution, and the capability of live-cell imaging. FIG. 7A shows the Raman spectra of cells treated with anisomycin. Compared to control cells treated with DMSO, cells treated with 50 nM anisomycin show a Raman spectrum with evident changes in the CH stretching region, the cell-silent region, and the fingerprint region (FIG. 7B).

[0127] In the CH stretching region, cells treated with 50 nM anisomycin have lower CH2 stretching peaks and higher CH3 stretching peaks than control cells, indicating 50 nM anisomycin induces an increase of the protein-to-lipid ratio in cells (FIG. 7B). We then defined the protein-to-lipid ratio as the intensity at 2930 cm'1divided by the intensity at 2850 cm'1and calculated this ratio for individual cells (FIG. 7B). Cells treated with 50 nM anisomycin have -31% increase in the protein-to-lipid ratio compared to the control (FIG. 7C). The protein-to-lipid ratio is a crucial parameter for tumor heterogeneity (Wei et al. PNAS. 2019). Besides, a higher unsaturated CH stretching peak indicates an increase of unsaturated lipids upon drug treatment (FIG. 7B). We then defined the lipid unsaturation ratio as the intensity at 3050 cm1divided by the intensity at 2850 cm'1. We found cells treated with 50 nM anisomycin have -35% increase in the lipid unsaturation ratio compared to the control (FIG. 7B). High lipid unsaturation is associated with increased cancer sternness and tumor initiation capacity in ovarian cancer (Li et al. Cell Stem Cell. 2017], In the fingerprint region, cells treated with 50 nM anisomycin show a lower13C Amide I peak and a higher12C Amide I peak (FIG. 7B). The lower13C Amide I peak indicates a decreased rate of protein synthesis, consistent with the MoA of anisomycin as a protein-synthesis inhibitor. The higher12C Amide I peak suggests an extended lifetime of pre-existing proteins as compensation of protein-synthesis inhibition, which is corroborated by a higher phenylalanine peak at -1000 cm'1(FIG. 7B). We then defined protein synthesis rate, saturated lipid synthesis rate, and unsaturated lipid synthesis rate (See Methods, below) and calculated these parameters for individual cells. Compared to the control, cells treated with 50 nM anisomycin have -13%, -14%, and 9% decrease in the protein synthesis rate, saturated lipid synthesis rate, and unsaturated lipid synthesis rate, respectively (FIG. 7C).

[0128] Raman images of cells treated with increasing concentrations of anisomycin were collected. Compared to the response from the control to 50 nM treated, the response from 50 nM to 500 nM is reduced (FIG. 7C). The responses from 500 nM to 1 pM and from 1 pM to 2 pM are further diminished (FIG. 7C). These results indicate that major biochemical changes in cells may occur at a concentration of only 50 nM. The response from 2 pM to 5 pM show large noise across the entire spectral range, indicating 5 pM anisomycin exerts evident toxic effects, which might kill or fundamentally alter a proportion of cells. Quantitative analyses indicate that the increase in the protein-to-lipid ratio and lipid unsaturation ratio and the decrease in the saturated / unsatu rated lipid synthesis rate plateau at 50 nM, while the decrease in the protein synthesis rate plateau at 500 nM (FIG. 7C).

[0129] Next, the drug response of cells to TVB-3166 was studied using Raman microscopy. As shown in FIG. 8A, the response from 5 pM to 10 pM is most evident. Compared to cells treated with 5 pM TVB- 3166, cells treated with 10 pM TVB-3166 show lower peaks of newly synthesized lipids from azido- palmitic acid and deuterated oleic acid (FIG. 8B). Compared to the control, cells treated with 10 pM TVB- 3166 have 37% and 143% increase in the protein-to-lipid ratio and lipid unsaturation ratio, respectively; 54%, 39%, and 29% decrease in the saturated lipid synthesis rate, unsaturated lipid synthesis rate, and protein synthesis rate, respectively (FIG. 8C). These results suggest dysregulated lipid metabolism upon TVB-3166 treatment, consistent with its mechanism of action as a fatty acid synthase inhibitor. VIB-OT explores perturbation interactions

[0130] Combination therapy, which combines two or more therapeutic agents, has become a cornerstone of cancer therapy. It provides a complementary strategy to enhance chemical efficacy by offering synergistic or additive interaction effects and reducing drug resistance. VIB-OT was used to study chemical interactions at different dosages, using gefitinib (Gef) and bortezomib (Bor) as a demonstration. The corresponding coupling is computed between consecutives steps for both chemical treatments, as shown in FIG. 9A. For instance, for the combination of Gef IC-20 + Bor IC20, the coupling is computed separately from Gef IC20 to combination (orange) and Bor IC20 to combination (blue). In this way, the shifts of molecular responses from different states / trajectory can be clearly visualized (FIG. 10A - 10C). Interestingly, for the combination of Bor IC50 + Gef IC20, the orange curve suggests an evident decrease of13C amide I probe and increase of 12C amide I normal, suggesting the addition of bortezomib at higher concentration (IC50) decreases protein synthesis compared to the previous state (Bor IC20 + Gef IC20). The blue curve indicates a slight decrease of all molecular responses when gefitinib is added at IC20 concentration compared with the state where only Bor IC50 is added. This clearly shows how the molecular response of chemical combination is transitioned from single treatment or other combination treatment states.

[0131] To quantify the molecular shifts, we further computed the changes of molecular responses compared to control states, for both single chemical treatments and corresponding combinations (FIG. 9B and FIG. 9C). The "high” annotation here refers to IC50 concentration while the “low” annotation indicates the IC20 concentration. An interesting observation is that under both high and low concentrations, the fatty acid uptakes (both saturated and unsaturated) of the combination treatment are consistently the lowest compared with both single treatments. This indicates that this chemical combination effectively inhibits fatty acid uptakes compared with single drug treatment. Overall, these results reveal that VIB-OT can serve as an approach to visualize the transitions and molecular responses in chemical combinations, promising combination therapy evaluations.

[0132] Discussion

[0133] VIB-OT is the first integrated experimental and computational framework using vibrational imaging-based approach to the drug discovery pipeline. VIB-OT utilizes imaging technologies such as IR and spontaneous Raman that are able to measure cellular states via the vibrational modes of their biochemical activities. This information represents an interpretable and direct measure of fundamental cellular behavior, like protein compositions, protein synthesis rate, abundance of RNA / DNA and sugars, and other metabolic processes. Crucially, VIB-OT can take these measurements in a cheap, non- invasive, and minimally perturbative manner.

[0134] Using VIB-OT, single-cell drug responses can be characterized by computing couplings between observed control and treated cells using the mathematics of optimal transport. By associating regions of vibrational modes to known chemical activities, VIB-OT can measure nuanced perturbation responses in an interpretable manner. The methods of action can be differentiated and understood, and that batch effects can be addressed easily. Heterogeneity in responses can be linked to features in the control state, and VIB-OT can understand the effects of different dose concentrations and combinations. VIB-OT is rooted in imaging-based technologies and thus is well suited for complex biological models, such as (3-D) organoids and tissues. These biological models are becoming increasingly relevant in the drug discovery pipeline as they can more closely resemble the biological settings of their target applications. However, their increased complexity represents a challenging setting, where complicated cellular states can arise from the abundance of multiple cell types, cell-to-cell signaling, effects due to the local environments, etc. While currently unexplored in this work, VIB-OT is capable of spatially resolved measurements, which can be utilized to help study such complicated phenomena. Furthermore, computational techniques can be used to recover heterogeneity in response. Here, optimal transportbased techniques are used to induce a coupling between control and treated states.

[0135] Screens executed with VIB-OT are cheap and scalable on the order of 1000 perturbations. The above-described results were obtained using the following materials and methods.

[0136] Methods

[0137] Vibrational Imaging

[0138] Cell line and materials: MDA-MB-231 (ATCC HTB-26) was purchased from ATCC. For reagents, azido-palmitic acid (1346) was purchased from Click chemistry tools; algal amino acid mixture (U-13C, 97-99%, CLM-1548) was purchased from Cambridge; deuterated oleic acid (683582) was purchased from Sigma-Aldrich. For drugs, anisomycin (A9789), bortezomib (179324-69-7), cycloheximide (01810), emetine (SMB01061 ), everolimus (94687), epirubicin hydrochloride (E9406), gefitinib (184475-35-2), lapatinib (231277-92-2), TVB-3166 (SML1694), taxol (PHL89806), vincristine sulfate (V8388) and apicidin (A8851 ) were purchased from Sigma-Aldrich. Daunorubicin hydrochloride (AAJ60224MA), iniparib (AC469161000), triacsin C (24-721 -00U), doxorubicin hydrochloride (BP25165), MG 132 (AAJ63250LB0), dactolisib (NC0298104) were purchased from Fisher Scientific. For cell culture agents, DMEM medium (11965), FBS (10082), penicillin / streptomycin (1514), were purchased from ThermoFisher Scientific. CaF2 substrates (CAFP13-1 ) were purchased from Crystran.

[0139] Probe preparation and media recipe for cell labeling: Azido palmitic acid-bovine serum albumin (BSA) solution. For the solution, couple azido-palmitic acid with BSA to prepare a 2-mM stock solution. Prepare 20 mM sodium palmitic acid solution by dissolving palmitic acid in NaOH solution with the following recipe: azido-PA (5.5 mg) + 1 .0 ml dd-H2O + 35 pl 1 M NaOH. Mix and incubate the solution in 70 °C water baths until no oil droplets are visible. Then slowly add the sodium palmitic acid solution into 2.7 ml 20% BSA under room temperature water baths. Quickly add 6.3 ml DMEM culture medium and filter the solution with a 0.22-pm sterile filter.13C-amino acids DMEM. 4 mg mF algae13C-amino acids mix was dissolved in dd-H2O with 10% FBS and 1 % penicillin, which matched the concentrations of regular amino acids in DMEM. deuterated oleic acid-bovine serum albumin (BSA) solution. For the solution, couple d34-oleic acid with BSA to prepare a 2-mM stock solution. Prepare 20 mM oleic acid solution by dissolving oleic acid in NaOH solution with the following recipe: ds4 oleic acid (6.3 mg) + 1 .0 ml dd-H2O + 24 pl 1 M NaOH. Mix and incubate the solution in 70 °C water baths until no oil droplets are visible. Then slowly add the ds4 oleic acid solution into 2.7 ml 20% BSA under room temperature water baths. Quickly add 6.3 ml DMEM culture medium and filter the solution with a 0.22-pm sterile filter. Cell culture. MDA-MB-231 cells were cultured in DMEM media supplemented with 10% FBS and 1% penicillin. Cells were grown in a humidified atmosphere containing 5% CO2 at 37 °C in the incubator. At -80% confluence, cells were dissociated with trypsin and passaged.

[0140] Cell viability assay and drug IC50 calculus. The IC50 values of the drugs were determined by Alamar blue assay. Cells were seeded at 10,000 per well in 96-well plates. After 24 h, the cells were washed twice with phosphate buffered saline (PBS) and treated with drugs at different concentrations in cell culture media for 48 hrs. Each drug concentration has 6-8 replicates. After the drug treatments, cells were washed with PBS twice and the cell viability was determined by Alamar blue assay following the manufacturer’s protocol (Invitrogen) using plate reader. The IC50 values were determined by fitting the data using a dose response model with variable Hill slope built in Prism.

[0141] Drug combination and synergy score calculus. Cell viability assay was performed on cells treated by drug combinations to evaluate synergy scores. Everolimus-Lapatinib and Everolimus- Doxorubicin were chosen as model systems to study drug combinations. The concentration combinations of these two groups are shown in FIG. 9A - FIG. 9C and FIG. 10A - FIG. 10C. Each drug concentration combination has 5 replicates. To calculate synergy scores, an open-source package SynergyFinder was applied. The synergy scores were calculated based on the ZIP model available in the package.

[0142] Sample preparation for drug-treated cells with labeling. MDA-MB-231 cells were seeded on clean CaF2 substrates with 5*104cells per well in cell culture media (DMEM, 10% FBS, 1 % penicillin) overnight for control and other drug treatment conditions. Then the culture media was replaced by reamino acids DMEM with 50 M azido-palmitic acid, 50 M ds4 oleic acid, either single drug at its IC50 concentration or drug combinations at chosen concentrations. Drugs were prepared in 100% DMSO and diluted to 0.1% DMSO in labeling media. For the control group, only cell labeling media with 0.1% DMSO was added (without any drugs). Cells were treated for 48hrs. The culturing time was selected based on the trade-off between signal and experimental time. After that, cells were fixed by 4% PFA at room temperature for 15 min and washed three times with PBS buffer and five times with dd-H2O. The samples were then air-dried before FTIR imaging.

[0143] FTIR imaging. Agilent Cary 620 Imaging FTIR equipped with an Agilent 670-IR spectrometer and 128 x 128-pixels FPA mercury cadmium telluride (MCT) detector was used in the transmission mode. A background spectrum was collected on a clean CaF2 substrate using 128 scans at 8 enr1spectral resolution, suggesting that the IR absorbance was measured every 4 cm1. Cell spectra were recorded using 64-128 scans at 8 cm-1spectral resolution. A x25 IR objective (pixel size, 3.3 pm, 0.81 numerical aperture (NA)) was used for cell imaging.

[0144] Spontaneous Raman imaging. Spontaneous Raman imaging was performed using an upright confocal Raman microscope (Xplora, HORIBA Jobin Yvon). Cell samples were illuminated by 532 nm laser (80 mW on sample) through a 50x objective (air, NA 0.75, MPIan N, Olympus). Raman images were acquired using the point-by-point mapping mode with an acquisition time of 5s and 1 x accumulation for each point measurement. The step size was set as 7 pm. The grating was set as 1200 gr / mm. Both the slit size and the hole size was set as 100 pm. 500-1000 cells were imaged for each condition.

[0145] Raman image processing. Raw Raman spectra were first processed in the LabSpec 6 software (HORIBA). The “Despike” function was used to remove cosmic rays. The “Threshold” function was used to remove spectra with high background. The “Correction” function was used to subtract background from raw spectra by selecting a field of view without any cells. After that the spectral range from 2830 cm1to 3000 cnr1was integrated to generate a cell image, using custom-written MATLAB scripts. The cell image was used to segment single cells and generate a cell segmentation mask with the Otsu thresholding method using the Cell Profiler software. Based on the cell segmentation mask, single-cell spectra were generated by averaging all the spectra within each cell, using custom-written MATLAB scripts.

[0146] Calculation of the protein-to-lipid ratio, lipid unsaturation ratio, protein synthesis rate, saturated lipid synthesis rate, and unsaturated lipid synthesis rate. The protein-to-lipid ratio is defined as the intensity at 2930 cm1divided by the intensity at 2850 cm1. The lipid unsaturation ratio is defined as the intensity at 3050 cm1divided by the intensity at 2850 cm1. The protein synthesis rate is defined as the peak area of 1600-1630 cm1divided by the sum of the peak area of 1600-1630 cm1and the peak area of 1634-1700 cnr1. The saturated lipid synthesis rate is defined as the peak area of from 2080-2130 cm’1divided by the peak area of from 2825-2875 cnr1. The unsaturated lipid synthesis rate is defined as the peak area of from 2130-2230 cm’1divided by the peak area of from 2825-2875 cnr1.

[0147] Normalization of vibrational spectra

[0148] Chemical imaging spectra typically have two major sources of technical noise that are removed with different normalization strategies. The first is a scattering effect that introduces a drift in the spectra and is removed with a baseline correction method. The second is introduced by the variable depth of the measured tissue and can be thought of as similar to the library size in transcriptomic settings. These are typically removed by computing cell-specific normalization constant as the average over the whole spectra, or specific peaks. Many choices exist and are often left to the user depending on the experimental and analytical settings.

[0149] Five normalization methods were explored:

[0150] Amide I normalizes the area of the Amide I peak (1600-1800 cm1) to 1

[0151] Amide II normalizes the area of the Amide II peak (1470-1570 cm1) to 1

[0152] CH normalizes the area of the (a)symmetric CH stretching (2815-3015 cm’1) to 1

[0153] Signal normalizes the area of the whole spectra to 1

[0154] Min-Max linearly scales and shifts the whole spectra to lie in [0, 1]

[0155] Unless otherwise stated, IR datasets were analyzed with the signal normalization and Spontaneous Raman datasets were analyzed with the CH normalization.

[0156] Recovering individual cell perturbation effects with optimal transport

[0157] While VIB-OT can profile cells in a non-destructive manner, it is not possible to measure the same cell in multiple treated states. As a result, instead of access to a set of paired cellular measurements in the treated and untreated states, complex heterogeneous perturbation responses were reasoned with only observations of treated and untreated populations. Here, computational approaches were relied on to recover individual cellular responses, which, like all computational approaches, required making some assumptions on the nature of these perturbations. The main assumption made on perturbation effects was that they do not fully alter cellular states, but instead affect only some targeted sub-programs of a cell and leave the rest of the profile largely unaltered. Individual cellular responses were recovered with optimal transport (OT), a mathematical framework that aligns probability distributions. With this approach, observations of cells were viewed from a profiled population as samples from an underlying probability distribution. Given a control and perturbed population, the perturbation effect was modeled as one that transforms the distribution over control states into the distribution over treated states. Many such functions will fulfil this transformation, for instance a trivial random assignment, but optimal transport identifies the transformation that alters cell states as efficiently as possible, with respect to some cost function. This principle of least effort is well-aligned with our assumption that perturbations only alter some targeted sub-programs of a cell as it essentially means that the coupling can be induced by utilizing the unaltered cellular features.

[0158] Let |i and be the probability distributions underlying the control and treated cellular states and T: |i => v be a function that maps cells from the control state to treated in the treated state. The optimal transport problem can be formulated as:

[0159] MinT:T#n=v f ||x - T(x)| |2d|i(x) where the pushforward constraint, T#|i=v, ensures that T transforms |i into v.

[0160] The optimal transport problem was solved via an efficient solution to the Kantorovich formulation, as implemented in the PythonOT package. This solution induces soft assignments between control and treated cells, when a discrete assignment is necessary, for example when labels are propagated, a hard assignment is induced by a weighted sample over all assigned states.

[0161] Other Embodiments

[0162] While the invention has been described in connection with specific embodiments thereof, it will be understood that it is capable of further modifications and this application is intended to cover any variations, uses, or adaptations of the invention following, in general, the principles of the invention and including such departures from the invention that come within known or customary practice within the art to which the invention pertains and may be applied to the essential features hereinbefore set forth, and follows in the scope of the claims.

[0163] Other embodiments are within the claims.

Claims

CLAIMS1 . A method of analyzing a biological sample comprising:(a) obtaining one or more infrared images of the biological sample using infrared microscopy and / or one or more Raman images of the biological sample using Raman microscopy; and(b) applying, by a computer, a machine learning model to analyze the one or more infrared images and / or the one or more Raman images.

2. A method of identifying a biological state in a biological sample comprising:(a) obtaining one or more infrared images of the biological sample using infrared microscopy and / or one or more Raman images of the biological sample using Raman microscopy; and(b) applying, by a computer, a machine learning model to analyze the one or more infrared images and / or the one or more Raman images, wherein an output of the analysis indicates the biological state.

3. A method of determining the efficacy of an agent in effecting a change in a biological sample comprising:(a) obtaining one or more infrared images and / or one or more Raman images of the biological sample after administration of the agent, wherein the one or more infrared images are obtained using infrared microscopy and / or the one or more Raman images are obtained using Raman microscopy; and(b) applying, by a computer, a machine learning model to analyze the one or more infrared images and / or the one or more Raman images, wherein an output of the analysis indicates that the agent is effective in effecting a change in the biological sample.

4. The method of any one of claims 1 -3, wherein prior to obtaining the one infrared images and / or the one or more Raman images, one or more infrared probes and / or one or more Raman probes are added to the biological sample.

5. The method of claim 4, wherein the one or more infrared probes and / or the one or more Raman probes comprise a probe for detecting unsaturated fatty acid uptake.

6. The method of claim 5, wherein the probe for detecting unsaturated fatty acid uptake comprises ds4 oleic acid.

7. The method of claim 5, wherein the one or more infrared probes and / or the one or more Raman probes comprise a probe for detecting saturated fatty acid uptake.

8. The method of claim 7, wherein the probe for detecting saturated fatty acid uptake comprises azido palmitic acid.

9. The method of claim 4, wherein the one or more infrared probes and / or the one or more Raman probes comprise a probe for detecting protein synthesis.

10. The method of claim 9, wherein the probe for detecting protein synthesis comprises13C labeled amino acids.11 . The method of any one of claims 1 -10, wherein each of the one or more infrared images comprises a collection of pixels, wherein each pixel comprises an infrared spectrum, and / or wherein each of the one or more Raman images comprises a collection of pixels, wherein each pixel comprises a Raman spectrum.

12. The method of claim 11 , wherein the method further comprises determining a chemical composition of the biological sample at each pixel.

13. The method of claim 1 -12, wherein the infrared microscopy produces a spectrum from 400 cm'1to 4000 cm'1and / or the Raman microscopy produces a spectrum from 500 cm'1to 3500 cm1.

14. The method of claim 13, wherein the spectrum is from 900 cm1to 1800 cm1.

15. The method of any one of claims 1 -14, wherein the one or more infrared images and / or the one or more Raman images comprises one or more segmentation images at a spectrum from2830 cm1to 3000 cm1.

16. The method of claim 15, wherein the computer applies a cell segmentation method to the one or more segmentation images to generate an output of a segmented cell.

17. The method of claim 16, wherein each of the one or more infrared images comprises a collection of pixels, wherein each pixel comprises an infrared spectrum, and / or wherein each of the one or more Raman images comprises a collection of pixels, wherein each pixel comprises a Raman spectrum, and wherein:(a) the infrared spectra of a subset of the collection of pixels of the one or more infrared images comprising the segmented cell are averaged to generate an average infrared spectrum of the segmented cell; and / or(b) the Raman spectra of a subset of the collection of pixels of the one or more Raman images comprising the segmented cells are averaged to generate an average Raman spectrum of the segmented cell.

18. The method of claim 16 or 17, wherein the cell segmentation method is an Otsu thresholding method.

19. The method of claim 17 or 18, wherein:(a) the average infrared spectrum of the segmented cell is normalized by scaling the area of the average infrared spectrum to 1 ; and / or(b) the average Raman spectrum of the segmented cell is normalized by scaling the area of a subspectrum of from 2815 cm-1to 3015 cm-1to 1 .

20. The method of any one of claims 1 -19, wherein the machine learning model was developed using a combination of one or more reference infrared images and / or one or more reference Raman images.21 . The method of any one of claims 1 -20, wherein the machine learning model is an optimal transport model.

22. The method of claim 20 or 21 , wherein the one or more reference infrared images and / or the one or more reference Raman images were obtained using infrared microscopy and / or Raman microscopy, respectively, from a reference sample comprising reference cells having reference cell states.

23. The method of claim 22, wherein the optimal transport model is applied to the one or more infrared images and / or the one or more Raman images by minimizing the cost function for transforming the infrared spectra and / or the Raman spectra of the biological sample to an infrared spectra and / or a Raman spectra of the reference cells.

24. The method of claim 22 and 23, wherein the reference cells comprising cells of the biological sample prior to addition of an agent for effecting a change.

25. The method of any one of claims 3-24, wherein the agent is a small molecule, a nucleic acid, a protein, a gene editing agent, or a cell.

26. The method of claim 25, wherein the nucleic acid is an RNAi, an shRNA, an mRNA, or an antisense oligonucleotide.

27. The method of claim 25, wherein the protein is a transcription factor, an antibody, a peptide, a cytokine, a hormone, an enzyme, an antibody drug conjugate, or a fusion protein.

28. The method of claim 25, wherein the gene editing agent is clustered regularly interspaced short palindromic repeats (CRISPR), a zinc finger nuclease, a base editor, a prime editor, or transcription activator-like effectors (TALEN).29, The method of claim 25, wherein the cell is a chimeric antigen receptor T (CAR-T) cell.

30. The method of any one of claims 1 -29, wherein the biological sample comprises a cell, a tissue, and / or a cellular component.31 . The method of any one of claims 1 -30, wherein the biological sample was obtained from a human.

32. The method of any one of claims 1 -31 , wherein the biological sample is obtained from a subject with a disorder.

33. The method of claim 30, wherein the biological sample comprises human cancer cells.

34. The method of claim 33, wherein the biological sample comprises human breast cancer carcinoma cells.

35. The method of any one of claims 3-34, wherein the effected change comprises treating a disorder, altering a gene or protein expression profile, inducing cell differentiation, inducing a cell state change, inducing cell death, or a combination thereof.

36. A computer whose input data is images obtained using infrared microscopy and / or images obtained using Raman microscopy, wherein the computer is programmed with a machine learning model developed using a combination of images obtained using infrared microscopy and / or Raman microscopy.

37. The computer of claim 36, wherein the machine learning model is an optimal transport model.

38. A system comprising:(a) an infrared imaging component configured for infrared microscopy and / or a Raman imaging component configured for Raman microscopy; and(b) a computer configured to receive images from the infrared imaging component and / or images from the Raman imaging component, wherein the infrared imaging component and / or the Raman imaging component are each operatively coupled to the computer, wherein the computer is programed with a machine learning model whose input data is the images from the infrared imaging component and / or the Raman imaging component.

39. A method of predicting the response of a subject to a drug comprising:(a) obtaining one or more infrared images and / or one or more Raman images of a biological sample after administration of the drug, wherein the biological sample was obtained from the subject, and wherein the one or more infrared images are obtained using infrared microscopy and / or the one or more Raman images are obtained using Raman microscopy; and(b) applying, by a computer, a machine learning model to analyze the one or more infrared images and / or the one or more Raman images, wherein an output of the analysis is predictive of the response of the biological sample to the drug.

40. A method of predicting the response of a subject to a drug combination comprising:(a) obtaining one or more infrared images and / or one or more Raman images of a biological sample after administration of the drug combination, wherein the biological sample was obtained from the subject, and wherein the one or more infrared images are obtained using infrared microscopy and / or the one or more Raman images are obtained using Raman microscopy; and(b) applying, by a computer, a machine learning model to analyze the one or more infrared images and / or the one or more Raman images, wherein an output of the analysis is predictive of the response of the biological sample to the drug combination.41 . A method of determining the mechanism of action, dosage, or target of a drug comprising:(a) obtaining one or more infrared images and / or one or more Raman images of a biological sample after administration of the drug, and wherein the one or more infrared images are obtained using infrared microscopy and / or the one or more Raman images are obtained using Raman microscopy; and(b) applying, by a computer, a machine learning model to analyze the one or more infrared images and / or the one or more Raman images, wherein an output of the analysis indicates the mechanism of action, dosage, or target of the drug.

42. A method of determining the mechanism of action, dosage, or target of a drug combination comprising:(a) obtaining one or more infrared images and / or one or more Raman images of a biological sample after administration of the drug combination, and wherein the one or more infrared images are obtained using infrared microscopy and / or the one or more Raman images are obtained using Raman microscopy; and(b) applying, by a computer, a machine learning model to analyze the one or more infrared images and / or the one or more Raman images, wherein an output of the analysis indicates the mechanism of action, dosage, or target of the drug combination.

Images