Methods for high-throughput drug screens at clonal resolution

Abstract

The present invention provides methods for determining therapeutic response of cancer clones in a tumor cell sample. The invention further provides methods for determining the effect of a compound on one or more populations of cells using single-cell sequencing. The invention further provides methods of analyzing cell viability of a treated tumor cell sample over time for proactive hit selection and phenotypic profiling. The invention further provides software methods for training and using a statistical model or a machine learning algorithm for determining cell type or predicting cell type for a plurality of cells from a biological sample. The invention further provides software methods of analyzing cell culture image data for a drug screen. The invention further provides software methods analyzing cell culture data for selecting cell cultures for phenotypic profiling. The invention further provides software methods for selecting and pooling a subset of drug compounds from a library of drug compounds.

Description

METHODS FOR HIGH-THROUGHPUT DRUG SCREENS AT CLONAL RESOLUTIONFIELD OF THE INVENTION

[0001] The present invention provides methods for binary annotation of single-cells as sensitive or resistant to a compound prior to downstream single-cell phenotyping. The invention further provides methods for measuring cell viability over time in the presence of a compound to enable proactive selection of hits for downstream single-cell phenotyping. The invention further provides methods for measuring cell type-specific viability over time in the presence of a compound to enable proactive selection of disease-specific hits for downstream single-cell phenotyping. The invention further provides a software method for using live-cell imaging and machine learning for measuring cell type-specific viability over time with dramatically lower cost, time, and phototoxic effects. The invention further provides software methods for designing compressed drug screens, classifying cancer clones in heterogeneous tumor samples, and predicting drug response for each cancer clone. The invention further provides methods for determining treatment response in sub-populations of cancer cells. The present invention further provides a method for predicting treatment response in a subject with cancer.BACKGROUND OF THE INVENTION

[0002] Drug screens are commonly used in early drug discovery efforts to prioritize compounds for further investigation. Drug screens have started to be used for precision medicine applications to identify effective, often unexpected, drugs that work for a unique patient’s tissue sample. This is often called drug repurposing or functional precision medicine (FPM).

[0003] For most cancer patients and cancer drugs, there are no biomarker-guided strategies to select the best treatment regimens. Many patients undergo highly toxic chemotherapies with significant side effects, and only a subset of the population benefit. In fact, even for the most effective drug combinations that represent the standard-of-care, most patients are only responding to one of the drugs in the treatment cocktail, resulting in overtreatment, unnecessary toxicity and costs. See, Hwangbo et al. Additivity predicts the efficacy of most approved combination therapies for advanced cancer, Nat Cancer, 2023 Dec;4(12): 1693-1704, Epub 2023 Nov 16.

[0004] Some cancer patients have clinically actionable genomic alterations that can be treated with targeted therapies. However, the majority lack such markers and less than 10% of tested patients are matched to an effective targeted therapy. See, Letai, A. et al. Cancer Cell. 2021 Dec23, 40(1 ):26— 35. New strategies are needed to guide treatment selection for these patients. A further challenge in oncology is tumor heterogeneity. Even if a tumor has an actionable genetic alteration with an approved drug available, there are often clonal subpopulations within the tumor that have intrinsic or acquired resistance to said drug. It’s been demonstrated that the most resistant clone in a tumor is the primary determinant of clinical outcome to treatment. See, Sinha S. et al., PERCEPTION predicts patient response and resistance to treatment using single-cell transcriptomics of their tumors, Nat Cancer, 2024, Jun;5(6):938-952, Epub 2024 Apr 18.

[0005] Tumor heterogeneity exists not just between tumors (i.e. patients), but also within tumors. Intra-tumor heterogeneity within an individual patient limits the efficacy of a single cancer agent, including chemotherapies and targeted therapies. This has hampered the impact that molecular diagnostics and targeted therapies have had on patient outcomes. A large national trial revealed that the current “upper limit” for individuals with targetable genetic alterations is 38% but that many tumors with predicted genetic vulnerabilities failed to respond to the matched targeted therapy. The response rate in the NCI-MATCH trial in an intend on-to-treat analysis was below 5%. See, Letai, A. et al. Functional precision oncology: testing tumors with drugs to identify vulnerabilities and novel combinations, Cancer Cell, 2021 Dec 2340: 1 :26-35.

[0006] In contrast to molecular diagnostics, functional precision medicine is an alternative approach that uses ex vivo drug testing on cancer cells from a patient to identify effective therapies for their individual tumor. This approach has become standard for infectious disease, where antibiotics are tested on bacteria isolated from a patient and grown on agar plates. Despite the promise, functional precision medicine has yet to have the same impact for cancer. Many challenges make it difficult to predict the clinical response of a patient’s tumor through ex vivo drug testing. One important challenge is the tumor sample itself, which can have a mixture of cancer cells, healthy cells, and dead cells. This impacts the accuracy of measurements made on bulk cells, such as the CellTiter Gio assay. Single-cell technologies, such as flow cytometry or microscopy, enable more accurate measurements of drug effects specifically on cancer cells. It’s now possible to calculate the percent of cancer cell death relative to healthy cells, providing a measure of the therapeutic window of a tested drug.

[0007] Accurately predicting treatment response is a major goal in oncology research and cancer care. This requires rapid ex vivo drug screens on fresh tumor samples to match patients to the optimal drug or treatment combination based on their unique tumor biology. To overcome tumorheterogeneity, technologies with single-cell resolution must be applied, such as single-cell RNA sequencing. These technologies allow clonal subpopulations within a tumor sample to be identified. Measuring the differential effect of compounds on these clonal subpopulations enables personalized drug combinations to be designed that address the full complexity of a patient’s tumor.

[0008] Current methods leveraging single-cell technology known in the art, including the method described in US Patent Application No. 20100298255A1, US Patent Application No. 20210181183A1, and US Patent Application No. 20220011296A1, are able to measure the selective effect of a drug on individual cancer cells relative to healthy cells. However, they fail to address the fact that not all cancer cells are the same. Intra-tumor heterogeneity can negatively affect the predictive accuracy of these methods. If a drug is shown to kill 99% of cancer cells ex vivo, it may prove to be effective in the patient. Or, the remaining 1% of cancer cells may represent a unique population with a specific genetic alteration or phenotype that makes them resistant to the tested drug. Another drug may have appeared to be ineffective by only killing 1% of cancer cells, but it could have been highly effective at killing this unique population of cells. This information is critically important when predicting clinical outcomes and designing combination regimens, yet current methods in the art fail to provide it.

[0009] The MIX-seq method described in McFarland et al. and PCT Application No. PCT / US2020 / 048896 uses single-cell RNA sequencing technology to identify single nucleotide polymorphisms (SNP) and measure drug response based on transcription profiles. See, McFarland et al. Multiplexed single-cell transcriptional response profiling to define cancer vulnerabilities and therapeutic mechanism of action, Nat Commun 2020 Aug 27; 11(1):4296. This method has two major limitations. First, SNPs can only be used to differentiate between cancer cells from different donors, as they demonstrate by multiplexing cell lines from distinct donors. Cancer cells from the same donor (i.e. tumor) will share the same synonymous polymorphisms present in the cancer patient’s genome. This means MIX-seq is unable to identify clonal sub-populations of cancer cells from a primary tumor sample. Importantly, clonal subpopulations from a tumor sample will harbor non-synonymous mutations, including single nucleotide variants (SNV), copy number variations (CNV), insertions and deletions (indels), and more. These non-synonymous mutations have important effects on cancer cell biology, including phenotype, sensitivity and resistance to therapies. The second limitation of the MIX-seq methodis the use of transcription profiles to measure drug response. For certain drugs with mechanisms of action (MO A) that indirectly induce cell death, transcription profiles can be used successfully as a marker of cell viability. However, the authors demonstrate that for drugs that act directly on the apoptosis pathway, transcription profiles are unable to measure drug response. The BCL-2 inhibitors venetoclax and navitoclax represent an important class of FDA-approved drugs used in many cancer types, therefore making MIX-seq inappropriate for functional precision medicine.

[0010] A further limitation of MIX-seq and other single-cell phenotyping methods based on technologies such as scRNA-seq is that they are low throughput and very expensive due to the price of single-cell kits and DNA sequencing. For example, if a library of five compounds is screened across five concentrations each, the cost can rise to $10,000. This makes single-cell phenotyping technologies such as scRNA-seq prohibitively expensive for clinical applications.

[0011] An alternative approach to reduce costs is to use single-cell phenotyping technologies such as scRNA-seq on the baseline tumor sample only. These technologies generate rich phenotypic data at single-cell resolution that can be used to identify clonal subpopulations and predict drug response in a single patient. However, if a drug screen is not run on the patient’s unique tumor sample, the model is unlikely to accurately predict drug sensitivity and resistance for that patient’s unique set of clonal subpopulations.

[0012] There are now hundreds of cancer compounds approved by the Food and compound Administration (“FDA”), and the vast majority lack a predictive biomarker. Even for targeted therapies that are approved with a genetic biomarker, tumor heterogeneity is ignored, leading to high rates of treatment failure. There is a significant need for high-throughput drug screening methods with single-cell resolution that are cost-effective for clinical application. Such methods should effectively match a patient’s unique tumor to the optimal drug or combination, while generating high-content phenotypic data for training predictive models for future use.

[0013] Further, these methods should determine the effect of a compound on sub-populations of cancer cells for functional precision medicine.SUMMARY OF THE INVENTION

[0014] The present invention is directed to methods to run high-throughput phenotypic screens with single-cell resolution in a cost-effective manner. The present invention is further directed to methods for designing compressed drug screens, classifying cancer clones in heterogeneous tumor samples, and predicting drug response for each cancer clone.

[0015] In one aspect, the present invention is directed to methods of determining therapeutic response of cancer clones in a tumor cell sample, the method comprising the steps of: a. treating a subset of tumor cells from the tumor cell sample with a compound to create a treated subset and an untreated subset; b. labeling the treated and untreated subsets with an apoptosis reagent conjugated to a fluorophore or magnetic bead; c. using the apoptosis reagent to physically separate the tumor cell sample into a first fraction and a second fraction, wherein the first fraction are non-apoptotic cells and the second fraction are apoptotic cells; d. sequencing nucleic acids isolated from the first fraction to provide sequence reads from which cancer clones can be determined; e. sequencing nucleic acids isolated from the second fraction to provide sequence reads from which cancer clones can be determined; and f. determining cancer clones that are sensitive to the compound as cancer clones whose frequencies are lower in the first fraction relative to the frequencies of the same cancer clone in the untreated subset; and g. optionally, the step of determining cancer clones that are sensitive to the compound as cancer clones whose frequencies are higher in the second fraction relative to the frequencies of the same cancer clone in the untreated subset; and h. optionally, the step of determining cancer clones that are resistant to the compound as cancer clones whose frequencies are higher in the first fraction relative to the frequencies of the same cancer clone in the untreated subset; and i. optionally, labeling cells from the first fraction and cells from the second fraction with different oligonucleotide barcodes; and j . optionally, combining the cells from the first fraction and the cells from the second fraction into a single pool.

[0016] In another aspect, the present invention is directed to methods for determining the effect of a compound on a population of cells using single-cell sequencing, comprising: a. providing a sample containing a population of cells; b. incubating the population of cells in the presence of a test compound; c. staining the population of cells with one or more response marker reagents;d. physically separating the population of cells into a sensitive population and a resistant population based on the response marker reagents; and e. performing single-cell sequencing on both the sensitive population and resistant population of cells.

[0017] In another aspect, the present invention is directed to methods of analyzing cell viability of a treated tumor cell sample, the method comprising the steps of: a. measuring cell viability over a time series in a plurality of cell cultures comprising cells from the treated tumor cell sample, wherein the tumor sample is being treated with multiple compounds and wherein the times series comprises an initial timepoint (TO), one or more intervening timepoints (Th) and a final timepoint (Te); b. identifying one or more cell cultures from the treated tumor cell sample comprising an effective compound as a cell culture whose viability is below a viability threshold at a timepoint Th; c. transferring the one or more cell cultures treated with the effective compound at the intervening timepoint Th to a preservation buffer or a fixation buffer; d. repeating steps a-c until the final timepoint Te; and e. performing phenotypic analysis of the cells transferred at each timepoint Th, optionally, wherein the analysis occurs after the final timepoint Te, optionally, wherein the viability threshold is based on the ratio of malignant and normal cells in the tumor cell sample, optionally, wherein the viability threshold is based on a viability level measured in a third set of cell cultures containing non-malignant cells and optionally, wherein at least one of the one or more intervening timepoints Th is prior to secondary necrosis.

[0018] In another aspect, the present invention is directed to computer-implemented methods of training and using a statistical model or a machine learning algorithm for determining cell type and or predicting cell type for a plurality of cells from a biological sample, comprising: a. creating a first training set of transmitted light images of labeled cell types; b. receiving a second training set of matched transmitted light channel images and fluorescent channel images from a first set of cell cultures containing cells from a biological sample;c. training the statistical model or the machine learning algorithm for classifying cell types in the first set of cell cultures using cell type labels from the fluorescent image channel to label cell type in the transmitted light channel on the same matched images; and d. using the statistical model or the machine learning algorithm from step c to label cell types in transmitted light images from a second set of cell cultures containing cells from the same biological sample, and optionally, wherein the cells of the cell culture are alive during step b.

[0019] In a preferred embodiment of the present invention, the viability threshold is based on the ratio of malignant and normal cells in the tumor cell sample or a viability level measured in a third set of cell cultures containing non-malignant cells.

[0020] In another preferred embodiment of the present invention, at least one of the one or more intervening timepoints Th is prior to secondary necrosis.

[0021] In another aspect, the present invention is directed to a computer-implemented method of analyzing cell culture image data for a drug screen, the method comprising: a. accessing image data comprising a set of fluorescent images and matched non- fluorescent images from the same regions of interest in a plurality of cell cultures containing a cytotoxic compound, wherein the fluorescent image data comprises a first signal associated with a type of cell and a second signal associated with viability of a cell; b. using an image analysis algorithm to label each cell in each fluorescent image and matched non-fluorescent image as malignant or normal using the first signal; c. using an image analysis algorithm to label each cell in each fluorescent image and matched non-fluorescent image as viable or dying using the second signal; d. using the labeled cells from step c. in the matched non-fluorescent images to train one or more statistical models for a task of labeling each cell as normal or malignant and viable or dying in a non-fluorescent image; e. accessing image data comprising a second set of non-fluorescent images from a plurality of cell cultures treated with distinct drugs; f. using the statistical models to label each cell as normal or malignant and viable or dying in the second set of non-fluorescent images;g. quantifying the number of cells that are viable and malignant out of the total number of malignant cells in each image of the second set of non-fluorescent images; and h. calculating a drug response metric based on the quantification from step g. for each of the plurality of cell cultures.

[0022] In another aspect, the present invention is directed to computer-implemented methods for selecting a subset of drug compounds from a library of drug compounds comprising: a. receiving a request to evaluate a group of drug compounds; b. accessing a database comprising drug screen results for the group of drug compounds, wherein the drug screen results comprise test results from a plurality of samples treated with the group of drug compounds comprising cancer cell lines, tumor clones, and normal cell types and wherein the database comprises viability measurements, drug response labels, and RNA expression profiles for each pair of drug and sample; c. determining co-occurrence and mutual exclusivity of sensitivity labels of each drug compound of the group of drug compounds across each sample in the database; d. determining similarity in expression signatures of each drug compound of the group of drug compounds across each sample labeled as sensitive to each drug compound in the database; e. determining similarity of viability measurements over time of each drug compound of the group of drug compounds across each sample labeled as sensitive to each drug compound in the database; f. calculating an optimization score using the samples labeled as sensitive to each drug compound, wherein the optimization score minimizes overlap of samples labeled as sensitive in the database, similarity of expression signatures in samples labeled as sensitive for each drug compound, and similarity in viability measurements over time for samples labeled as sensitive for each drug compound; and g. assigning drug compounds into the subset of drug compounds, wherein the drug compounds in the subset of drug compounds have a similar effect on tumor cells and wherein assignment into the subset of drug compounds is based on the optimization score, wherein the subset of drug compounds is to be evaluated in parallel for an effect on tumor cells.

[0023] In a preferred embodiment of the present invention, the cells of the cell culture are alive during step b., immediately above.

[0024] In a further preferred embodiment of the present invention, each subset of drug compounds selected contains a unique set of drug compounds from every other subset of drug compounds selected and wherein each subset of drug compounds selected contains no more than one drug compound in common with any other subset of drug compounds selected.

[0025] In another aspect, the present invention is directed to computer-implemented methods for selecting a subset of drug compounds from a library of drug compounds, comprising: a. receiving a request to evaluate a group of drug compounds; b. accessing a database containing single-cell RNA expression data for a tumor sample; c. identifying one or more clonal populations in the tumor sample based on the single-cell RNA expression data; d. running one or more prediction models for the group of drug compounds on each clonal population, wherein the one or more models predict viability of the clonal population after treatment with each drug compound; and e. assigning one or more drug compounds from the group of drug compounds into the subset of drug compounds, wherein the drug compounds in the subset of drug compounds have an effect on distinct clonal population from the one or more clonal populations, wherein assignment into the subset of drug compounds is based on the predicted viability for each drug compound, wherein the subset of drug compounds is to be evaluated in parallel for an effect on tumor cells.

[0026] In a preferred embodiment, each subset of drug compounds selected contains a unique set of drug compounds from every other subset of drug compounds selected and wherein each subset of drug compounds selected contains no more than one drug compound in common with any other subset of drug compounds selected.BRIEF DESCRIPTION OF THE DRAWINGS

[0027] FIG. 1 illustrates a flow diagram for determining clonal drug response using a drug response annotation method.

[0028] FIG. 2 illustrates a user interface including information associated with the tumor cell populations and drug response based on the annotation method.

[0029] FIG. 3 shows a UMAP representation of cells from a heterogenous mixture of breast cancer cell lines treated with DMSO control.

[0030] FIG. 4 shows data comparing viability readouts from the annotation method (A) to viability reagents CellTiter-Glo (CTG) and Caspase-Gio (CG) for heterogeneous cell line mixtures screened with a library of drugs.

[0031] FIG. 5 shows data comparing viability readouts from the annotation method (A) to viability reagents CellTiter-Glo (CTG) and Caspase-Gio (CG) for deconvoluted breast cancer cell line mixtures screened with a library of drugs.

[0032] FIG. 6 shows data comparing viability readouts from the annotation method (A) to viability reagents CellTiter-Glo (CTG) and Caspase-Gio (CG) for deconvoluted leukemia cell line mixtures screened with a library of drugs.

[0033] FIG. 7 shows data comparing viability readouts from the annotation method with viability readouts from a prediction method using single-cell sequencing on a heterogeneous mixture of breast cancer cell lines screened with a library of drugs.

[0034] FIG. 8 illustrates a flow diagram for a proactive hit selection method.

[0035] FIG. 9 illustrates embodiments of a kinetic drug screen with proactive hit selection.

[0036] FIG. 10 shows data comparing real-time fluorescent viability reagents for proactive hit selection in cell cultures treated with staurosporine.

[0037] FIG. 11 shows data for 24 hour and 72 hour cell viability of a heterogeneous mixture of breast cancer cell lines treated with a drug library with diverse kinetic effects and mechanisms of action.

[0038] FIG. 12 shows data for cell viability over time for proactive hit selection for each of alpelisib (a) and everolimus (b) occurring at the earliest timepoint that breast cancer cell viability is observed to be below a predefined threshold relative to healthy control cell viability.

[0039] FIG. 13 shows data for single-cell sequencing cell recovery at the proactive hit timepoint (a) and the 72 hour endpoint (b) for the two earliest hits identified in the breast cancer cell line screen.

[0040] FIG. 14 shows data for single-cell sequencing clonal recovery at the proactive hit timepoint (a) and the 72 hour endpoint (b) for the two earliest hits identified in the breast cancer cell line screen.

[0041] FIG. 15 shows data comparing single-cell sequencing clonal drug response (annotation method) to a cell viability reagent at the proactive hit timepoint for each of alpelisib (a) and everolimus (c) and the 72 hour endpoint for each of alpelisib (a) and everolimus (d) for the two earliest hits identified in the breast cancer cell line screen.

[0042] FIG. 16 shows data for 24 hour and 72 hour cell viability of a heterogeneous mixture of leukemia cell lines treated with a drug library with diverse kinetic effects and mechanisms of action.

[0043] FIG. 17 shows data for cell viability over time for proactive hit selection for each of azacitidine (a) and venetoclax (b) occurring at the earliest timepoint that leukemia cell viability is observed to be below a predefined threshold relative to healthy control cell viability.

[0044] FIG. 18 shows data for single-cell sequencing cell recovery at the proactive hit timepoint and the 72 hour endpoint for each of azacitidine (a) and venetoclax (b) for the two earliest hits identified in the leukemia cell line screen.

[0045] FIG. 19 shows data for single-cell sequencing clonal recovery at the proactive hit timepoint and the 72 hour endpoint for each of azacitidine (a) and venetoclax (b) for the two earliest hits identified in the leukemia cell line screen.

[0046] FIG. 20 shows data comparing single-cell sequencing clonal drug response (annotation method) to a cell viability reagent at the proactive hit timepoint for each of azacitidine (a) and venetoclax (c) and the 72 hour endpoint for each of azacitidine (c) and venetoclax (d) for the two earliest hits identified in the leukemia cell line screen.

[0047] FIG. 21 shows a UMAP representation of cells from an AML tumor sample with normal cells (black), malignant clonal population 1 (dark gray), and malignant clonal population 2 (light gray) identified in the DMSO negative control wells.

[0048] FIG. 22 shows the normal and malignant cell population classification results for the single-cell sequencing data from the negative control DMSO wells for each tumor sample.

[0049] FIG. 23 shows data for cell viability over time for breast cancer tumor samples comparing tumor cells and healthy control cells treated with the same drug compounds and combinations.

[0050] FIG. 24 shows data for cell viability over time for proactive hit selection for each of alpelisib(a), alpelisib / fulvestrant mLiver002 (b) and alpelisib / fulvestrant mChest003 (c) occurring at the earliest timepoint that breast cancer tumor cell viability is observed to be below a predefined threshold relative to healthy control cell viability.

[0051] FIG. 25 shows data for single-cell sequencing clonal recovery at the proactive hit timepoint and the 72 hour endpoint for each of alpelisib(a), alpelisib / fulvestrant mLiver002 (b) and alpelisib / fulvestrant mChest003 (c) for the earliest hit identified for each breast cancer tumor sample screened.

[0052] FIG. 26 shows data comparing single-cell sequencing clonal drug response (annotation method) at the proactive hit timepoint to the 72 hour endpoint for each of alpelisib(a), alpelisib / fulvestrant mLiver002 (b) and alpelisib / fulvestrant mChest003 (c) for the earliest hit identified in each breast cancer tumor sample screened.

[0053] FIG. 27 shows data for cell viability over time for AML tumor samples comparing tumor cells and healthy control cells treated with the same drug compounds and combinations.

[0054] FIG. 28 shows data for cell viability over time for proactive hit selection for each of azacitidine (a), azacitidine / venetoclax (b) and idarubucin / cytaribine (c) occurring at the earliest timepoint that AML cell viability is observed to be below a predefined threshold relative to healthy control cell viability.

[0055] FIG. 29 shows data for single-cell sequencing clonal recovery at the proactive hit timepoint and the 72 hour endpoint azacitidine (a), azacitidine / venetoclax (b) and idarubucin / cytaribine (c) for the earliest hit identified for each AML sample screened.

[0056] FIG. 30 shows data comparing single-cell sequencing clonal drug response (annotation method) at the proactive hit timepoint to the 72 hour endpoint azacitidine (a), azacitidine / venetoclax (b) and idarubucin / cytaribine (c) for the earliest hit identified in each AML sample screened.

[0057] FIG. 31 illustrates a clinical report summarizing clonal drug response and treatment recommendation for a heterogeneous population of cancer cells.

[0058] FIG. 32 illustrates a flow diagram for training and using a model to measure cancer- selective drug response.

[0059] FIG. 33 illustrates embodiments of a high-throughput kinetic drug screen used to identify cancer-selective hits early and determine clonal drug response.

[0060] FIG. 34 illustrates a flow diagram for embodiments of a high-throughput kinetic drug screen used to identify cancer-selective hits early and determine clonal drug response.

[0061] FIG. 35 illustrates a workflow comparison between conventional drug screen analysis and embodiments of a high-throughput kinetic drug screen used to identify cancer-selective hits early and determine clonal drug response.

[0062] FIG. 36 illustrates a flow diagram for designing compressed drug pools for high- throughput drug screening on tumor cells.

[0063] FIG. 37 illustrates a flow diagram for using compressed drug pools in a high-throughput drug screen on tumor cells.

[0064] FIG. 38 illustrates a flow diagram for designing compressed drug pools using a prediction model.

[0065] Various embodiments are shown in the drawings for the purpose of illustration. However, those skilled in the art will recognize that alternative embodiments may be employed without departing from the principles of the present invention. Accordingly, while certain embodiments are shown in the drawings, the technologies described herein are amenable to various modifications.DETAILED DESCRIPTION OF THE INVENTION

[0066] In one aspect, the present invention provides methods for running a phenotypic drug screen with single-cell resolution. In a preferred embodiment of the present invention, the singlecell technology used for phenotyping is single-cell RNA sequencing (“scRNA-seq”). Derivative versions of the technology such as cellular indexing of transcriptomes and epitopes (CITE-seq) may also be used for purposes of multiplexing and / or measuring target surface proteins. This method enables each cell at the end of the drug screen to be annotated as sensitive or resistant using a binary classification step. This binary classification step physically separates cells into a pool of sensitive cells or a pool of resistant cells using one or more apoptotic markers. These pools are then kept separated as distinct samples for generating distinct cDNA libraries for sequencing or labeled with oligonucleotide barcodes and pooled back into a single sample for cDNA library preparation. This method overcomes current challenges in measuring drug response in single-cell sequencing data.

[0067] The primary method that has been used to measure cell viability in single-cell sequencing data are gene expression-based signatures associated with viability. In such methods, “viability-related” genes are identified by combining post-perturbation transcriptional datasets with cell viability data from the same set of cell lines and perturbations (i.e. compounds). For example, Szalai et al. identified a common “cell viability signature” (CVS) that enabled them to predict cell viability effectively across studies from different sources and types of perturbations. See, Szalai, B. et al. Signatures of cell death and proliferation in perturbation transcriptomics data- from confounding factor to effective prediction, Nucleic Acids Res 2019 Nov 4;47(19): 10010- 10026. The present invention was benchmarked against this method to demonstrate improved performance over expression-based viability signatures. See, Jones et al. further expanded on expression-based viability signatures by identifying a “global viability signature” (GVS) which revealed a robust association between drug sensitivity and expression for many genes, indicating that there is a consistent transcriptional signature related to viability effects. See, Jones, A. et al., Post-perturbational transcriptional signatures of cancer cell line vulnerabilities, preprint available March 5, 2020, at https: / / doi.org / 10.1101 / 2020.03.04.976217. Specifically, the global viability signature showed enrichment for genes involved in pathways relevant to cell viability, such as cell cycle regulation and apoptosis, suggesting that this approach is able to identify the biological pathways associated with fitness loss across a broad range of drug classes and different cell types. It was found that the global viability signature at 6- or 24-hours post-perturbation could be predictive of longer-term cell viability at 72 hours post-perturbation. See, US Patent Application Publication 2022 / 0340976A1. However, such expression-based methods can provide inaccurate and incomplete prediction of cell viability and drug response. In general, it was found that cancer cells with higher baseline proliferation will appear to be more sensitive across drugs screened, while cells with low baseline cell cycle signatures will appear to be more resistant. It was also found that the accuracy of a global viability signature was highly compound-dependent across cell lines, reinforcing that a single signature cannot capture the diversity of compound-specific responses. In addition, certain compounds caused distinct expression changes across cell lines. Furthermore, expression-based methods also provide incomplete prediction of cell viability for important cancer drug classes, such as Bcl-2 inhibitors, which cause no discernable change in gene expression post-perturbation. A primary advantage of the present invention is the unbiased labeling of cell viability in single-cell sequencing data, independent of compound or cell line.

[0068] Another method for measuring drug response in single-cell sequencing data is to use oligonucleotide-conjugated antibodies for apoptotic markers. These antibodies can bind toapoptotic markers (cytochrome C, Caspase 3 / 7, Annexin-V, etc.) and then the level of said apoptotic markers is measured based on the number of reads (i.e. “read count”) of the oligonucleotide barcode. However, these methods are prone to high background noise, because as cells die throughout the course of the drug screen, apoptotic markers are released extracellularly. These extracellular apoptotic markers are then bound by the oligonucleotide barcode and end up in proximity to viable cells during single-cell isolation and cell barcoding. The presence of these free-floating oligonucleotide barcodes in the same compartment (droplet, microwell, etc.) as a viable cell, results in said viable cell being labeled as sensitive (i.e. apoptotic). It is an advantage of the present invention that physical separation of sensitive (i.e. apoptotic) cells also captures free-floating extracellular apoptotic markers in the same fraction, leaving viable cells free of any background noise.

[0069] Accordingly, one aspect of the present invention provides methods for determining treatment response in a sub-population of cancer cells. The methods described herein can be used to measure the effect of compounds on distinct sub-populations of cancer cells from a subject's tumor for research, diagnostic, or therapeutic purposes.

[0070] In one embodiment of the present invention, cancer cells from a subject's tumor sample are split into two or more replicates and exposed to one or more compounds. At the end of the incubation, the cancer cells are analyzed to identify distinct sub-populations and measure the effect of the compounds on each cell. The results of the analysis can inform which subpopulations of cancer cells are sensitive or resistant to each compound tested.

[0071] Accordingly, another aspect of the present invention provides methods for running a high-throughput drug screen that measures cell viability over time to enable proactive selection of cells treated with an active drug hit for downstream phenotyping. Current drug screening methods have a defined endpoint for measuring drug response (‘endpoint assays’). In this invention, drug response is measured frequently throughout the course of the drug screen (‘kinetic assay’). Using a pre-defined, absolute, relative, or dynamic viability threshold, proactive hit selection can occur throughout the course of the drug screen. Once treatment wells pass this viability threshold, cells are collected and transferred to a preservation buffer designed to preserve the phenotypic state of the cells, such as the RNA or protein molecules present in the cell. Collecting cells from hit wells early in apoptosis, hours before secondary necrosis, preserves the biological profile and physical integrity of the cells for downstream phenotyping. Once cellsstart to undergo secondary necrosis, they lose cell membrane permeability, resulting in leakage of molecules. At this point, cells rapidly lose the fidelity of their original biological profile and will often not pass quality-control metrics used in single-cell sequencing computational pipelines. This method overcomes the challenge of different drugs having different mechanisms of action and pharmacokinetic effects on the cells. This method has the ability to cost-effectively screen a large library of drugs with diverse mechanisms of action and pharmacokinetic effects, and through proactive hit selection at early apoptotic timepoints, rich and high-quality phenotypic data can be generated with single-cell resolution for the top drugs having an effect on the tumor cells. Using cheaper and higher throughput technology for the drug screen prior to more expensive and lower throughput phenotyping (e.g. single-cell sequencing) allows large drug library to be screened cost-effectively, and only cells treated with drugs of interest (i.e. ‘hits’) are then collected for downstream phenotyping.

[0072] A primary advantage of this proactive hit selection method is the preservation of the phenotypic biology of the drug treated cancer cells early in the apoptotic process, hours before secondary necrosis starts. This timepoint occurs before cell membrane integrity is lost and phenotypic molecules (i.e. RNA or protein) start to permeate out of the cell and / or degrade. There are several additional advantages of this invention, including cost-savings, time-savings, and reduced phototoxic damage. Using a plate reader to measure apoptosis enables quick and frequent measurements of viability across one or more microplates in a large drug screen. Only measuring apoptosis in treatment wells after a certain timepoint also reduces the phototoxic damage from the fluorescent apoptosis reagent in the treatment wells. Since the apoptotic reagent will only be labeling apoptotic cells, phototoxic damage is further reduced for the viable cells. Measuring apoptosis first in the cytotoxic control wells also reduces the time it takes the plate reader to scan the plate. The full plate only needs to be scanned at timepoints after the cytotoxic control wells pass a certain fluorescence (i.e. viability) threshold. In certain embodiments of the present invention, healthy control cells are screened in parallel to identify cancer-selective drug hits. In other embodiments of the present invention, an imaging system is used in combination with a plate reader to measure cancer- selective drug effects within each treatment well.

[0073] In another aspect, the present invention provides methods for using live-cell imaging and in silico labeling for measuring cell type-specific viability over time with dramatically lower cost, time, and phototoxic effects. Live-cell imaging using fluorescent antibodies for cell typemarkers enables cancer-selective viability to be directly measured for each cell (i.e. object) in an image, leading to more accurate drug screen results. However, fluorescent reagents for live-cell imaging are expensive and cause phototoxic damage. To overcome this challenge, fluorescent antibodies for cancer cell type markers can be used with a first fluorescent channel, and / or fluorescent antibodies for normal cell types can be used in a second fluorescent channel. These fluorescent antibodies are only used in cytotoxic control wells containing heterogeneous cells from the tumor sample treated with a known cytotoxic compound. Matched fluorescent and transmitted light images are captured for these control wells, and statistical methods are used to learn distinct visual features of cancer versus normal cells in the transmitted light images based on the matched fluorescent images where each cell (i.e. object) is labeled with cell type markers. The learned visual features are then used across the treatment wells containing tumor cells to identify cancer versus normal cells in transmitted light images. This removes the effect of phototoxic damage in the treatment wells, enabling frequent viability measurements throughout the course of the drug screen. It also reduces the time required to capture viability measurements, since only transmitted light images need to be captured to identify cell type. Most importantly, it reduces the cost of the fluorescent antibody reagents by up to 99% by limiting their use to the cytotoxic control wells. Cell viability or apoptosis can be measured in the control and treatment wells through various methods, including cell death morphology (e.g. nuclear disintegration, cell membrane permeability, blebbing, etc.) or fluorescent apoptosis reagents (e.g. Cytochrome C, Caspase 3 / 7, Annexin V, etc ).

[0074] Fluorescent antibodies are used in the art to label cancer, normal, or both cell types from a tumor sample. In some embodiments of the present invention, a cocktail of fluorescent antibodies with a first fluorophore can be used to label multiple normal cells commonly found in tumors. These normal cells include, but are not limited to, immune cells (CD45+), fibroblast cells (FAP+), and endothelial cells (CD31+). A cocktail of fluorescent antibodies with a second fluorophore can be used to label cancer cells. For example, in solid tumors, cancer cells of epithelial origin typically express EpCam and / or one more cytokeratins (e.g. pan-cytokeratin, cytokeratin 8, 18, 19, etc.). Mesenchymal markers (e.g. vimentin, FN1, CDH2, and SERPINE1 / PAI1, etc.) can also be included to identify cancer cells undergoing epithelial-to- mesenchymal transition (EMT). In hematological cancers, cancer cells can be identified using immunophenotyping markers from the pathology report, which identify one or more blast-specific marker(s). These are only a few examples of cancer cell markers. Any cancer cell marker known in the art may be used in methods of the present invention.

[0075] A distinct fluorescence channel can then be used to measure bulk apoptosis (i.e. fluorescence) in each well using a fluorescent plate reader and a fluorescent apoptotic marker reagent (e.g. Cytochrome C, activated Caspase 3 / 7, Annexin- V, etc.). The level of apoptosis measured in the cytotoxic control wells provides an indicator to start measuring apoptosis in the treatment wells. Prior to this timepoint, there is no need to measure apoptosis in the treatment wells, increasing the throughput of the workstation or lab given a limited set of instruments. The level of apoptosis measured in the cytotoxic control wells also provides a relative fluorescence value correlated to all cells dying (i.e. normal and cancer cells). For a drug to have a cancer- selective effect, the level of apoptosis should be at or below the cancer fraction threshold measured in the cytotoxic control wells. For example, if the cancer fraction was calculated to be 50% of cells in the cytotoxic control wells, then a cancer- selective drug should result in the fluorescence intensity of the apoptotic reagent in the treatment well to be 50% or less relative to the fluorescence intensity of the apoptotic reagent in the cytotoxic control wells. In a preferred embodiment of the present invention, the measurement instrument (e.g. fluorescent plate reader) is part of an automated workstation that is automatically programmed at the start of the drug screen to transfer liquids from wells containing a hit compound for downstream phenotyping. The hit transfer step would occur proactively when an apoptosis fluorescent intensity threshold is crossed, and a cancer-selective hit is confirmed using imaging and in silico labeling.

[0076] A primary advantage of the methods of the present invention over plate reader fluorescence is a more accurate measurement of cancer-selective drug effect in each treatment well. Counting cell types in each well is much more accurate than relative fluorescence ratios across cytotoxic control and treatment wells or measuring cell viability in healthy control wells. Another advantage of this method is the accurate measurement of cancer fraction when cancer fraction differs between wells, such as drug screens using tumor slices or tumor fragments (i.e. tumoroids). A further advantage of the methods of the present invention is the more accurate measurement of cancer-selective drug response when measured within each treatment well by using the normal cells (e.g. fibroblasts, immune cells, endothelial cells) and cancer cells together in the same well from the tumor sample. Imaging each well, however, consumes more time and compute resources than fluorescence measurement in plate reader mode. In some embodimentsof the present invention, fluorescent apoptosis reagents (Caspase 3 / 7, Annexin V, etc.) are used in each well and plate reader measurements are used to nominate cancer- selective hits, in accordance with the plate reader method described above. When a well is nominated as a hit in plate reader mode, images for the nominated hit wells can then be captured to confirm a cancerspecific drug response by directly measuring cell type. Only confirmed hits (i.e. drugs selectively killing cancer cells in the treatment wells) would then undergo proactive hit selection, wherein the cells from the confirmed hit wells are transferred to a preservation buffer, as described above. Another advantage of combining this in silico labeling method with the plate reader method described above is the ability for the plate reader to measure the apoptotic fluorescence signal from adherent and floating solid tumor cells that can become detached while undergoing apoptosis in response to the drug treatment.

[0077] In another aspect, the present invention provides methods for designing compressed drug libraries when limited tumor material is available. The method comprises accessing a database of drug screen results on cancer cell lines, cancer tumor sample clones, and normal cell types. The database contains both viability measurements and RNA expression profiles for each drugsample (cell line, tumor clone, normal cell type) pair. The method uses set overlap minimization for multiple sets of features. A first feature set is cross-sensitivity, which quantifies the number of cell types (cell lines, cancer tumor clones, healthy cell types) that are labeled as sensitive for a pair of drugs. The method minimizes the overlap (co-occurrence) of cell types that are sensitive to drugs in the same pool. A second feature set is pharmacokinetic effect, with two or more effect labels (fast, slow, etc.). The method minimizes the overlap of drugs with the same pharmacokinetic effect label in the same pool. A third feature set is expression signature, which identifies the top differentially expressed genes in cell types that are labeled as sensitive to a drug. The method minimizes the overlap of differentially expressed genes for drugs in the same pool. In some embodiments of the present invention, a fourth feature set is used to minimize toxic drug combinations. This feature set is based on normal (i.e. non-malignant) cell types that are labeled as sensitive for a pair of drugs (when screened in combination), but not for the individual drugs. While past compression methods are designed using theoretically relevant factors (chemical structure, mechanism of action, drug target, etc.) that could be broadly applicable to diverse compound libraries, the method described herein has a significant advantage by leveraging past drug screen results to optimize drug pooling based on concretedifferences in drug effects. See, US Patent Application No. 2022 / 04340A1 . A secondary advantage of this method is that it improves with each new drug screen result added to the database, creating a flywheel effect to increase the level of compression (i.e. drug pool size) continuously over time.

[0078] In another aspect the present invention provides methods for designing compressed drug libraries using a prediction model. The method comprises accessing a database of single-cell RNA sequencing data for a tumor sample, and using said data to identify clonal populations and predict drug response for each clone-drug pair. The method minimizes the probability that two drugs screened together in a pool will have an effect on the same clonal population. This is achieved by first identifying each clonal population in the tumor sample using the single-cell RNA sequencing data, as described in more detail in the embodiments below. Once clonal populations have been identified, the RNA expression data for the cells from said clonal population are input into a prediction model, which uses the RNA expression profile or signature to predict cell viability after treatment with a drug compound. Such prediction models can be trained on previous drug screen results, including paired RNA expression and cell viability data from treated cancer cells. The viability data is measured at the end of the drug screen, while the RNA expression can be measured before (i.e. baseline) or after (i.e. post-perturbation) the drug screen. The model learns features in the RNA expression data predictive of the cell viability post-treatment and uses those features to predict cell viability for the same drug compound on an unseen clonal population from a tumor sample.

[0079] A primary advantage of using predictive models to design compressed drug pools is to improve the probability of finding effective drugs for each distinct clonal population in a tumor sample without increasing the number of treatment conditions required. Pools with hit drugs effective for the major clone are also more likely to have drugs effective for the minor clones from the tumor sample. Through deconvolution, the effect of each drug compound on each clonal population can be confirmed, and personalized drug combinations can be designed. A secondary advantage of this method is each new drug screen performed on a tumor sample can be used to improve the prediction models over time, which results in better compressed drug pool design and personalized drug combinations.

[0080] In some embodiments of the present invention, the step of classifying cancer cells and cancer clones is based on gene expression phenotype, inference of genetic characteristics fromtranscriptional data, or both. For example, gene expression phenotype can be used to generate unbiased clusters of phenotypically similar cells using well-known computational pipelines such as the R package Seurat. Cell clusters can be visualized using dimensionality reduction techniques such as t-Distributed Stochastic Neighbor Embedding (t-SNE) or Uniform Manifold Approximation and Projection (UMAP). Cell clusters can then be labeled manually or automatically using known cell type markers from databases such as the CellMarker2.0 database. Computational tools such as CellTypist or scType can be used for automated annotation. Normal human cells can be accurately identified based on these cell type markers, typically resulting in high confidence scores indicating a high level of confidence in the cell type annotation. Clusters with low scores can be labeled as “unknown” or “other” cell types. Within these clusters, malignant cells can be identified using additional bespoke cell markers based on the specific tumor type. For example, acute myeloid leukemia (AML) blasts are immunophenotyped as part of the pathology process, and blasts often express a combination of established marker genes, including CD33, CD34, CD38, PR0M1, ENG, CD99 and KIT33. Leukemic stem cells (LSCs) can be further differentiated using distinct marker combinations, such as CD34+CD38-. For solid tumors, such as breast cancer, malignant epithelial cells often express EpCAM and / or cytokeratins (KRT8, KRT18, and KRT19). Malignant epithelial cells undergoing epithelial-to- mesenchymal transition (EMT) can express mesenchymal cell markers, including Vimentin, FN1, CDH2, and SERPINE1 / PAI1. Using normal human cell markers in combination with tumor-specific markers can be used to classify each cell from a tumor sample as normal or malignant. In addition to cell type markers, genetic inference methods and reference-based methods can be used to classify normal and malignant cell types. Inferable genetic characteristics include chromosome number (i.e. aneuploidy), copy number variation (CNV) across genomic loci, or genetic sequence alterations (i.e. single nucleotide variants, insertions, deletions, etc.). Tools exist to infer genetic alterations using single-cell RNA sequencing data, such as inferCNV, CopyKAT, and SCEVAN. Reference-based methods can also be used to differentiate between normal and malignant cells in a sample. In such reference-based methods, a classification model is trained on a reference dataset of cells labeled as normal or malignant. Reference-based method tools exist for this purpose, such as scATOMIC.

[0081] Within the malignant cell clusters, distinct cancer clones can be further differentiated based on gene expression phenotype, inference of genetic characteristics from transcriptionaldata, or both. For example, in an AML sample, one clone may have a typical blast phenotype described above, while another clone may have a distinct immunophenotype resembling a leukemic stem cell (LSC). In solid tumors, the major clone might display a typical epithelial phenotype (e.g. EpCAM and / or pan-cytokeratin), while a minor clone might be undergoing EMT and display mesenchymal markers. In addition to known cell type markers and expression profiling, inference of genetic alterations provides an orthogonal and powerful means to identify distinct cancer clones. Malignant cell populations undergo a branched evolutionary process over time in response to environmental and therapeutic pressures. In this process, a single cancer clone can evolve into two genetically distinct populations through the acquisition, loss, and / or modification of genetic material. In some embodiments of the present invention, cancer clones are identified by applying genetic inference tools such as InferCNV, CopyKAT, and SCEVAN to single-cell RNA expression data. In some embodiments,, an ensemble approach (e g. majority voting) or consensus approach (e g. agreement between orthogonal methods) comprising cell type markers, gene expression profiles, and multiple inference tools can be used to classify malignant cell clusters and label one or more clusters as distinct clones for the purpose of drug screening and response prediction.

[0082] In order to simulate a heterogeneous tumor sample while preserving a source of ground truth for verification of the results produced by the present invention, a mixture of cell lines was co-cultured and screened. The breast cancer cell lines MDA-MB-231 (70% of co-cultured cells), MCF-7 (20% of co-cultured cells), and MCF-7taml (10% of co-cultured cells) were used to simulate a solid tumor sample, and the B-cell leukemia cell lines SUP-B15 (70% of co-cultured cells), REH (20% of co-cultured cells), and NALM-6 (10% of co-cultured cells) were used to simulate a hematologic tumor sample. First, to verify accurate identification of distinct cancer clones (i.e. cell lines) at the end of each drug screen, the proportion of each clone in the negative control wells containing 0.1% dimethyl sulfoxide (DMSO) was compared to the known proportion of each cell line added to the co-culture at the start of the drug screen. The ratio of cell lines input into the co-culture simulates the ratio of distinct clones identified in a baseline tumor sample. As described above, consensus and ensemble approaches were used to generate cell clusters and label cell clusters to identify distinct cancer clones (i.e. cell lines). Gene expression profiles were used to generate cell clusters, and a combination of cell type markers (CellTypist), genetic inference tools (SCEVAN), and classifier models (scATOMIC) were usedto label each cluster to identify malignant clusters and distinct cancer clones. A consensus approach was used to label each cluster. When consensus did not occur, majority voting was used to label each cluster based on the orthogonal methods described above. To further demonstrate the performance of the present invention, drug screens were performed on primary tumor samples. For solid tumors, breast cancer tumor samples with known receptor status and genetic alterations of therapeutic relevance were screened to compare the clonal drug response results to the genetic background and receptor status of the tumor samples. For hematologic tumors, acute myeloid leukemia (AML) samples were screened to compare the malignant cell cluster annotation to the blast fraction in the pathology report, the malignant cell cluster phenotype to the blast markers in the pathology report, and the clonal drug response to the treatment history.

[0083] In another embodiment the present invention provides approaches to phenotypically screen large libraries of compounds in a resource-efficient manner. Current compound screens use multiple concentrations of each compound to measure a dose-response curve over a fixed period of time. The effect of each compound is measured at a fixed point in time, the end of the compound screen. The differential effect of each dose for a given compound measured at the end of the screen results in a dose-response curve that can be used to calculate a compound response score. One common example of compound response scores is half maximal inhibitory concentration (“IC50”) which measures the potency of a compound in inhibiting cell growth. Another common example is area under the curve (“AUG”) which measures the effect of a compound across all concentrations tested. Additional curve-fitting equations can be used to take into account additional features of the dose-response curve, including the slope and upper and lower limits. While these methods are effective at accurately measuring a compound response score that can be compared across compounds, they require many distinct concentrations (usually 5 to 10) to construct the dose-response curve. Each concentration requires distinct cells for the compound to be tested on, and each concentration is usually tested in technical replicates. If 1,000 cells are tested per replicate, and a 5 -point dose curve is used for screening, then one compound requires 10,000 cells. This becomes a major constraint when running compound screens on precious tissue samples with limited cellular material.

[0084] An underappreciated dimension of a compound screen is time. As described above, most compound screens measure the effect of each compound at a fixed period of time, the end of thecompound screen. While exposure time is always considered and closely measured in patients for pharmacokinetic studies, it is ignored and normalized away in compound screens. Current compound screens select a standard exposure time for all compounds with the intent that the full effect of the compound will occur by the end of the screen when the effect is measured.

[0085] In the present invention, we describe methods to measure the effect of each compound in a screen over time and make multiple selection steps for further analysis. These methods provide several advantages over current compound screens. First, they allow compounds that are highly effective to be identified earlier in the screen, which enables earlier phenotypic analysis of the cells exposed to those compounds, preserving important features of biology before the cells are completely killed. Second, they allow dose-time-response curves to be constructed, which generates more information about the effect of a compound on the cells with fewer concentrations needed. Third, the dose-time-response curves can be used to identify differential compound response scores that would not be detectable using dose-response curves at the end of a screen. These differential compound response scores can improve the comparison of compounds with similar kinetics within a screen, as well as the comparison of a specific compound between screens (i.e. different tissue samples being screened).

[0086] In a preferred embodiment of the present invention, a first instrument is used to collect measurements over a series of programmed timepoints for each well in a microplate.

[0087] The software program receives the measurement data from the first instrument and calculates a response output for each treatment well using the measurement from one or more control wells at each timepoint. For example, a negative control containing no compounds or an inert compound is used to measure background signal and quantify the effect size for each treatment well. In another example, a positive control containing a toxic or highly potent compound is also used to normalize the effect size for each treatment well relative to a known effect. Once measurement data from multiple timepoints is received, the software program also calculates the slope of the response output over multiple timepoints for each treatment well. The value of the response output and the slope are used by the software program to execute one or more proactive selection operations.

[0088] In a preferred embodiment of the present invention, the first instrument is a microplate reader. In such an embodiment, the microplate reader can use a variety of measurement methods based on luminescence or fluorescence. For example, the level of metabolism of all the cells inthe well can be measured using a luminescent or fluorescent reagent. In such an example, the dose-time-response curve would have a negative slope relative to control wells. In another example, the proportion of apoptotic or dead cells can be measured using a fluorescent reagent that binds to apoptotic markers or a fluorescent dye that is cell impermeable and binds to DNA, respectively. In such an example, the dose-time response curve would have a positive slope relative to control wells.

[0089] In another embodiment of the present invention, the first instrument is a microplate reader with imaging capabilities. In such an embodiment, additional measurement methods are available. For example, whole-well imaging can be used to measure confluence, which is related to cell growth and viability. In another example, single-cell imaging can be used to count each cell in the well and measure its morphology. In some embodiments of the present invention, a standalone microscope is used for whole-well imaging or single-cell imaging.

[0090] In a preferred embodiment of the present invention, live-cell imaging is used to identify cell type and calculate a response output for each well. The calculation is based on the percent of viable malignant cells out of the total malignant cell count in each image of a well, or image of an area in the well, or sample from the well. For example, multiple images of each well can be taken, and the malignant cells are identified, counted, and added together for the well. In another example, one area within each well is imaged and malignant cells are identified and counted. In all examples, visual features are used to differentiate between malignant and normal cell types and viable and non-viable cells. For example, nuclear and cell morphology, size, and intensity can be used to differentiate between normal and malignant cell types. Nuclear morphology and cell membrane integrity can also be used to identify viable and non-viable cells. In some embodiments of the present invention, brightfield imaging is used to identify cell type and viability based on the features described above. In other embodiments of the present invention, fluorescent dyes are used to enhance the identification of cell type and viability. For example, a cell permeable fluorescent dye like calcein can be used to identify viable cells and a low-toxicity nuclear dye can be used to enhance the visual features of the nucleus for identifying malignant and normal cell types. In another example, a fluorescent marker like Annexin-V can be used to stain the membranes of non-viable cells.

[0091] In another embodiment of the present invention, a software program calculates the slope of the response output over multiple timepoints for each treatment well once measurement datafrom multiple timepoints is received. The value of the response output and the slope are used by the software program to execute one or more selection operations. For example, a response output threshold of 80% (relative to the positive control well) can be used to select treatment wells during the compound screen. In such an example, the selected treatment wells are further analyzed to identify the phenotype of the minority of cells that remain viable. An advantage of this invention is that it enables hits to be analyzed earlier in the compound screen before all the cells are dead. For compound screens using cells from primary tissue samples, phenotypic drift is also reduced the earlier the cells are analyzed. In another example, the slope of the response output over multiple timepoints is used to select treatment wells. This can occur during the screen, when the slope flattens out and approaches an asymptote. At this point, minimal further effect is expected. In order to reduce experimental time and phenotypic drift, the selected treatment wells are further analyzed to identify the phenotype of the cells that remain viable. In a further example, the slope of the response output over multiple timepoints is used to select treatment wells near or at the scheduled end of the compound screen. This can be especially useful for slow-acting compounds. In order to predict the potency of slow-acting compounds, the steepness of the curve at the late timepoints can serve as a surrogate measure without having to continue to run the compound screen. An advantage of the present invention is the ability to run short (e.g. 24 or 48 hour) compound screens across compounds with diverse mechanisms and kinetic effects, while preserving baseline biology for further phenotypic analysis.

[0092] In another embodiment of the present invention, each selection operation involves two treatment wells containing distinct doses of the same compound. The high dose and low dose provide two distinct dose-time-response curves to rank hits (i.e. compare potency of different compounds). The high dose and low dose also provide distinct opportunities to identify sensitive and resistant populations during downstream phenotypic analysis. For example, the high dose provides more information regarding resistant cells that remained viable up until the selection operation. The low dose provides more information regarding highly-sensitive cells that lost viability early, prior to the selection operation. In some embodiments of the present invention, the software program sends instructions to a user or automated instrument based on the response output or slope of the high dose. In other embodiments of the present invention, the selection operation is based on the low dose. In further embodiments of the present invention, the selection operation is based on both doses. In all embodiments of the present invention, the logic behindthe selection operation is configurable by a user of the software program and can be configured at the compound and dose level.

[0093] In other embodiments of the present invention, the selection operation involves liquid handling. In such an embodiment, the software program instructs a user or automated liquid handling instrument to collect the volume from the selected wells. The cells from those wells can then be preserved using a preservation buffer, or fixed using a fixation buffer, prior to further analysis. In another embodiment of the present invention, the selection operation involves injection of a liquid into the selected wells. In such an embodiment, the software program instructs a user or instrument to inject a volume of liquid into the selected wells. For example, the liquid could be a preservation or fixation buffer that slows or stops the effect of the compound and preserves or fixes the cells. In a preferred embodiment of the present invention, the selection operation involves the automated injection step described above, and it occurs throughout the compound screen as the software program identifies treatment wells ready for selection. In such an embodiment, the compound screen is conducted fully autonomously by the software program, and at the end of the compound screen, volume from the selected wells is then collected for further phenotypic analysis. In such an embodiment, the first instrument automatically performs the injection steps when it receives instructions from the software program. A user can then receive a notification at the end of the compound screen once all selection operations have been completed. In some embodiments of the present invention, one or more steps involved in the phenotypic analysis are also automated using further equipment and instrumentation.

[0094] In another preferred embodiment, the present invention provides methods for running a drug screen on a plurality of cells in a microwell plate and measuring a drug response for each drug using microscope images. These methods provide a user the ability to run high-content phenotypic screens with large drug libraries in a resource-efficient manner.Methods of Measuring Clonal Therapeutic Response

[0095] Shown in FIG. l is a method of the present invention for measuring the therapeutic response of clonal cancer cell populations from a tumor sample using single-cell RNA sequencing. The method comprises treating the cells from a tumor sample with one or more drug compounds (100). In some embodiments of the present invention, the tumor cells are screenedwith a library of cancer drugs. In such embodiments, at the end of the drug screen the cells in each well are stained with a multiplexing reagent comprising an oligonucleotide barcode used to label the treatment condition for each cell in the drug screen (110). In some embodiments of the present invention, the multiplexing reagent is an oligonucleotide-conjugated antibody that binds to surface targets present on all human cells. In other embodiments of the present invention, the oligonucleotide barcode can be conjugated to another binding moiety such as a lipid molecule. In the embodiments described above, each treatment condition is stained with a unique oligonucleotide barcode sequence that allows the treatment condition (i.e. drug compound, dose, negative control, etc.) to be identified for each sequenced cell. Once cells have been stained with the multiplexing reagent, they can be pooled into one sample for drug response annotation (120).

[0096] As shown in FIG. 1, drug response annotation is achieved by staining treated tumor cells with an apoptosis reagent conjugated to a magnetic bead or fluorophore (130). For example, Annexin-V conjugated to magnetic microbeads will bind to phosphatidylserine present on the cell membrane of apoptotic cells and cellular debris. After staining with the apoptotic reagent, the cells are physically separated into a first fraction of non-apoptotic cells and a second fraction of apoptotic cells based on binding of the apoptosis reagent (140). In some embodiments of the present invention, physical separation is performed using a magnetic column. In such embodiments, the first fraction of non-apoptotic cells passes through the magnetic column into a first container while the second fraction of apoptotic cells is retained in the column based on binding of the apoptosis reagent. The second fraction can then be released from the column into a separate container by removing the magnet from the column. In other embodiments of the present invention, the physical separation is performed using fluorescence activated cell sorting by a flow cytometry instrument. Cells are sorted into a first container for the first fraction or a second container for the second fraction based on the absence or presence of the fluorescent apoptosis reagent on each cell, respectively.

[0097] The apoptotic and non-apoptotic fractions may be kept as separate samples for single-cell sequencing (150). This minimizes background noise (e g. cellular debris) for the non-apoptotic fraction, but also increases reagent cost (e.g. single-cell kits). Each fraction may also be stained with additional oligonucleotide-conjugated multiplexing reagents not used in the first multiplex staining described above. In such embodiments, the two fractions can be combined into one sample for single-cell sequencing In either case, proactive hit selection helps reduce the level ofnecrosis, RNA release, and background noise, which improves phenotypic clone identification and enables embodiments where the two fractions are combined. Proactive hit selection methods are described in more detailed embodiments below.

[0098] As shown in FIG. 1, the single-cell RNA sequencing data is used to phenotypically cluster cells. First, cells that don’t pass one or more quality control criteria are filtered out (160), as described above. Then, clusters are analyzed to determine normal cell type or malignant clone (170), as described in more detail above. In some embodiments, the next step is to use the multiplex barcode described above to identify the treatment condition for each cell (171). In other embodiments, this is step is performed before clustering, and cell populations clusters are formed for each unique treatment condition. The next step is to overlay the apoptosis annotation for each cell across the cell population clusters (172). Finally, each clone can be determined to be “sensitive” or “resistant” to a drug based on the number of apoptotic and non-apoptotic cells within a clonal population relative to the negative control (DMSO) and / or baseline tumor sample (180). A “sensitive” clone is defined as a clone present at a lower frequency in the non-apoptotic fraction relative to the negative control or baseline tumor sample. Conversely, a “resistant” clone is defined as a clone present at a higher frequency in the non-apoptotic fraction relative to the negative control or baseline tumor sample.

[0099] Comparing the frequency of each clone in the apoptotic and non-apoptotic fraction to the negative control or baseline tumor sample provides a more robust measure of clonal drug response compared to using the ratio of apoptotic and non-apoptotic cells in the treatment wells alone. This is because cell death and cell loss inevitably occur throughout the experimental process, and certain cell types and clones may be affected to different degrees. The baseline tumor sample can be considered “ground truth” for clonal frequency since minimal handling steps are performed prior to single-cell sequencing. In some embodiments of the present invention, a portion of the cell suspension after tumor dissociation can be stored in a preservation buffer during the drug screen, and then sequenced together with the treated tumor cells at the end of the drug screen. In other embodiments of the present invention, single-nucleus RNA sequencing (snRNA-seq) can be performed on the intact tumor tissue to minimize cell death and cell loss during tumor dissociation.

[0100] The negative control can be used to measure cell death and cell loss after all experimental steps have been completed. In general, optimized culture conditions and experimental techniquesshould result in most cells present in the non-apoptotic fraction in the negative control at a similar frequency to the baseline tumor sample. However, certain cell types can be differentially affected by specific experimental steps. For example, adherent cell types (e.g. normal fibroblasts, malignant epithelial cells) from a solid tumor may be preferentially lost during culture, washing, filtering, and / or physical separation steps. If this occurs, the frequency will be lower in the negative control relative to the baseline tumor sample. If differential cell death of a specific clone or cell type occurs due to ex vivo culture conditions, it will also be evident in the negative control at the end of the drug screen. For example, a quiescent clone from a hypoxic niche of a solid tumor may undergo stress-induced cell death in response to reactive oxygen species (ROS) caused by the high oxygen and nutrient levels in culture.

[0101] Within the DMSO negative control wells, the frequency of each cell type and clone in the apoptotic fraction is an indicator of cell death during the experiment, while the frequency in the non-apoptotic fraction indicates the level of cell loss during the experiment (relative to the baseline tumor sample). The frequency of each clone in the apoptotic and non-apoptotic fraction in a treatment well may be compared to the apoptotic and non-apoptotic fractions of the negative control. In such embodiments, a “resistant” clone is present at a similar frequency in both the apoptotic and non-apoptotic fractions as the negative control. Conversely, a “sensitive” clone is enriched in the apoptotic fraction of the treatment wells relative to the apoptotic fraction of the negative control, and / or depleted in the non-apoptotic fraction of the treatment well relative to the non-apoptotic fraction of the negative control. When the negative control is mostly non- apoptotic and highly similar to the baseline tumor sample, the treatment well can be compared to the baseline tumor sample, the non-apoptotic negative control fraction, or both. In the present invention, if a cancer clone is much more sensitive to a drug than the other clones from the tumor sample and underwent secondary necrosis at a faster rate, then it will be present at a low frequency in the treatment well and almost exclusively in the apoptotic fraction. Thus, the present invention provides a robust measure of clonal drug response based on the relative frequency of each clone in the apoptotic and non-apoptotic fractions in both the treatment wells and the negative control or baseline tumor sample. The present method also provides a highly accurate measure of the cancer-selective drug response relative to the normal cell types present in the tumor sample, as described in more detail below.

[0102] As shown in FIG. 2, the present invention results in clusters of distinct cell populations based on their RNA expression profile, as described above. The present invention also results in each cell being annotated “apoptotic” or “non-apoptotic” through drug response annotation method, which is a physical separation step after drug treatment and prior to single-cell sequencing. This annotation step provides a definitive label for each sequenced cell, which can then be used to measure clonal drug response. As shown in FIG. 2, malignant (211 and 221) and normal (212 and 222) cell populations can be visualized on one axis (i.e. the Y axis), and apoptotic and non-apoptotic cells can be visualized on the other axis (i.e. the X axis). To measure cancer-selective effect of a drug (210), the relative frequency of malignant cells (211) in the apoptotic fraction (top left quadrant) and non-apoptotic fraction (top right quadrant) can be compared to the frequency of normal cells (212) in the apoptotic fraction (bottom left quadrant) and non-apoptotic fraction (bottom right quadrant). A highly-selective cancer drug will have more malignant cells in the apoptotic fraction (top left quadrant) and more normal cells in the non-apoptotic fraction (bottom right quadrant). These ratios can be further compared to the baseline tumor sample or the DMSO negative control (220). The relative frequency of malignant cells (221) in the apoptotic fraction (top left quadrant) and non-apoptotic fraction (top right quadrant) can be compared to the frequency of normal cells (222) in the apoptotic fraction (bottom left quadrant) and non-apoptotic fraction (bottom right quadrant). These ratios for the DMSO control provide important information on basal cell death and differential cell loss. Cell population clusters in the DMSO control (220) can also be used as an unperturbed reference for identifying or confirming clonal populations in the treatment wells (210).

[0103] As shown in FIG. 2, cells can be further clustered by cell type (normal cells) or clonal population (malignant cells). These distinct populations are visualized as clusters with unique colors. Within the normal cell populations (212 and 222), common cell types found in tumors can be seen, such as fibroblasts, immune cells and endothelial cells. Within the malignant cell populations (211 and 221), three distinct clonal populations can be seen in this example.

[0104] As shown in FIG. 2, clonal population 1 appears to be sensitive to Drug A due its lower frequency in the non-apoptotic fraction compared to the negative control (“DMSO”), while clonal populations 2 and 3 appear to be resistant to Drug A. It is advantageous to sequence and analyze the apoptotic fraction to confirm the existence of a sensitive clone (i.e. the top left quadrant shown in FIG. 2) compared to assuming the existence of a sensitive clone based on itsabsence or lower frequency in the non-apoptotic fraction relative to the negative control or baseline tumor sample. For example, a clone may be present at a lower frequency in a certain treatment well based on factors unrelated to the compound in the treatment well, such as differential or random cell death or cell loss.Methods of Proactive Hit Selection

[0105] To measure clonal drug response as described above, clonal populations must be identifiable using single-cell sequencing after the drug screen. Single-cell sequencing requires intact cells to capture the expression profile based on the RNA molecules present inside the cell membrane and nucleus. A challenge with using single-cell sequencing to measure cell viability at the end of a drug screen is the fact that dying cells eventually lose cell membrane integrity and release their molecules extracellularly, changing their expression profile and losing their phenotype. To address this reality, the expression-based viability signatures described above are typically measured early in the drug screen, such as 6 or 24 hours after start of treatment. See, US Patent Application No. 2022 / 0340976A1. For most cancer drugs, sensitive cancer cells will start to undergo the early apoptosis process within that time frame. At this timepoint, the cell membrane remains intact while the cell starts to undergo programmed cell death as defined by apoptotic markers such as Cytochrome C release, Caspase 3 / 7 activation, and Annexin-V binding to phosphatidylserine present on the cell membrane. This point early in the cell death process provides an opportunity to both annotate apoptotic cells and identify their clonal identity using their mostly intact expression profile. 6-12 hours after the start of early apoptosis, cells begin to undergo secondary necrosis, resulting in cell membrane permeability and release of RNA molecules extracellularly. After this point in time, it becomes difficult to identify the clonal identity of a cell, and in many cases, these cells will not pass quality control thresholds necessary for single-cell sequencing analysis. Thus, a short (i.e. 24 hour) drug screen is preferred for measuring viability using single-cell sequencing. However, not every cancer drug has the same pharmacokinetic effect on cell viability. Certain cancer drugs will impart their full effect on cells within 24 hours, while others can take up to 72 hours to cause apoptosis and cell death. Thus, no single endpoint will be optimal across a library of cancer drugs with diverse mechanisms of action.

[0106] As shown in FIG. 8, the present invention provides a method for proactively selecting hits throughout the course of a drug screen to preserve the fidelity of the phenotype of the treated cells for clonal identification and measurement of clonal drug response. The method comprises treating tumor cells with a library of drug compounds (100) and measuring cell viability over a series of timepoints (700). When an effective drug hit is identified based on a cell viability threshold (710) a liquid transfer step (720) is immediately performed to move the cells from the treatment wells to a preservation buffer to quench the drug effect and preserve the transcriptomic phenotype of the cells for clonal analysis. This enables multiple proactive hit selection operations (710, 720) to occur throughout the course of the drug screen for drug compounds that have different pharmacokinetic effects on cancer cells. In some embodiments, the drug screen ends (730) after a certain number of hits have been selected and transferred to the preservation buffer. In other embodiments, the drug screen continues until a predetermined endpoint (730). After the cells have been transferred from all the hit wells to the preservation buffer, they can be processed together for single-cell sequencing (740). This reduces the cost, complexity, and batch effects associated with performing single-cell sequencing at separate timepoints for each hit. The present invention also avoids the need for cryopreserving cells at each timepoint, which negatively affects cell viability and phenotype after the freeze / thaw cycle.

[0107] As shown in FIG. 9, proactive hit selection occurs by measuring cell viability over time (700). When an effective drug hit is identified based on a cell viability threshold (710), a liquid transfer step (720) is immediately performed to move the cells from the treatment wells to a preservation buffer. Multiple proactive hit selection operations (710, 720) can occur throughout the course of the drug screen. In some embodiments, replicate treatment wells for identified hits can continue to be screened (i.e. not transferred to preservation buffer) to measure the full effect of the drug compounds until the end of the drug screen (730).Methods of Measuring Cancer-Selective Therapeutic Response

[0108] Plate reader assays utilizing luminescent or fluorescent cell viability reagents, like the ones described above, measure cell viability at the well level (i.e. bulk measurement). For tumor samples processed into a cell suspension, each well will contain a heterogeneous mix of malignant and normal cells that contribute to the overall cell viability measurement. It’s impossible to measure cancer- selective cell viability using plate readers, unless cancer cellmarkers are used to enrich malignant cells from the tumor sample prior to plating. This requires a full separate drug screen to be performed on normal cells from a healthy control sample (e.g. peripheral blood sample), potentially doubling the cost and time required to perform the drug screen. This also reduces the physiologic relevance of the drug screen, since normal cells from the tumor microenvironment (e.g. fibroblasts, macrophages, etc.) are known to have an effect on cancer cell response to drugs. The most physiologically relevant culture systems are tumor fragments or spheroids, which maintain the three-dimensional architecture of the tumor microenvironment during the drug screen. Enrichment is not possible with these culture systems, making bulk cell viability measurements unreliable.

[0109] Current microscopy -based drug screens use fluorescent reagents to determine cell type and measure cell viability. For example, a cancer cell can be identified using markers from a hematologic pathology report, or solid tumor markers such as EpCAM and pan-cytokeratin. Cell viability can be measured using a variety of different fluorescent reagents, such as nuclear dyes or calcein AM. Fluorescent reagents specific for apoptotic markers such as Caspase 3 / 7 or Annexin-V can also be used.

[0110] In the present invention, live-cell imaging is used to measure cancer- selective drug response in each treatment well containing heterogeneous tumor cells. Fluorescent antibodies for cancer markers and / or normal cell types are used to label malignant and normal cells, respectively. These live-cell fluorescent antibodies are added to each well at the start of the drug screen and don’t require any washing steps. Only cells expressing the target surface antigen bind and internalize the antibody. Before internalization, the antibodies don’t produce any fluorescent signal due to a pH-sensitive fluorophore. After cell binding and internalization, the lower pH inside the cell activates the fluorophore, labeling the cell as malignant or normal depending on the target the antibody binds to. A cocktail of antibodies for multiple cancer markers can be used with a first pH-sensitive fluorophore, and a cocktail of antibodies for different normal cell type markers can be used with a second pH-sensitive fluorophore. This results in each cell being fluorescently labeled as “malignant” or “normal” in the well.[0U1] However, there are significant limitations to live-cell imaging assays. These pH-sensitive fluorescent antibodies for live-cell imaging are expensive, costing up to $5 per well. Acquiring multiple images from distinct fluorescent channels for each well is also time-consuming, and each fluorescent image causes phototoxic damage.

[0112] In the present invention, a drug screen is performed using the fluorescent reagents described above only in a subset of wells. For example, a screen of 100 cancer drugs is performed on tumor cells from surgical or biopsy sample. As part of the screen, a positive control is included in the form of a highly cytotoxic compound, such as staurosporine. Transmitted light images, such as brightfield or phase contrast images, are captured for each well at baseline and at one or more timepoints. Fluorescent images are also captured for the same regions of interest at the same timepoints for the positive control wells containing the cytotoxic compound. Image analysis algorithms are applied to the matched fluorescent and transmitted light images to label each cell as normal or malignant, and apoptotic or non-apoptotic based on the fluorescent signals. The labeled cells in the transmitted light images are then used to train one or more statistical models (e.g. machine learning or deep learning models) to predict cell type label and apoptosis label for unseen cells in other transmitted light images.

[0113] The type of cell and response output for each cell is measured in one or more microscopy-based measuring operations to differentiate between malignant and normal cell types and apoptotic and non-apoptotic cells in each well. Each measurement results in a drug response calculation for each drug in the screen. The calculation is based on the percent of viable malignant cells out of the total malignant cell count in each image of a well, or image of an area in the well, or sample from the well. For example, multiple images of each well can be taken, and the malignant cells are identified, counted, and added together for the well. In another example, one area within each well is imaged and malignant cells are identified and counted. In all examples, visual features are learned by a statistical model and used to differentiate between malignant and normal cell types and viable and non-viable cells. For example, nuclear and cell morphology, size, and intensity can be used to differentiate between normal and malignant cell types. Nuclear morphology and cell membrane integrity can also be used to identify viable and non-viable cells. In some embodiments, brightfield imaging or phase contract imaging are used to identify cell type and viability based on the features described above.

[0114] As shown in FIG. 32, the present invention provides a method for training and using an in-silico labeling model to measure cancer-selective cell viability. In this method, the fluorescent antibodies are only added to the cytotoxic control wells (3020), and in silico labeling is used to identify malignant and normal cells in transmitted light images for the treatment wells. Matched fluorescent and transmitted light images captured from the cytotoxic control wells (3030) areused to train a computer vision model (3040). In some embodiments of the present invention, the computer vision model is pre-trained on a large dataset of transmitted light images of normal and malignant cell types with known labels (3010). The model learns general feature extraction of cell objects and cell types. The model can also learn normal cell type features that are common across samples and specific to certain cell types commonly found in tumors (e.g. fibroblasts, macrophages, endothelial cells, etc.). The model is then further trained using the matched fluorescent and transmitted light images captured from the positive control wells (3040). In some embodiments of the present invention, this tumor sample-specific training step utilizes transfer learning or fine-tuning techniques. The model learns visual features unique to the tumor’s malignant cells labeled with the first fluorophore, and the visual features specific to the normal cell types labeled with the second fluorophore. In some embodiments of the present invention, a second fluorophore is not needed to identify the normal cell types. In such embodiments, the normal cell types are identified as non-fluore scent, or the pre-trained model is sufficiently capable of labeling normal cell types in transmitted light images. These learned features specific to the malignant and normal cell types allow the model to then label each cell in the transmitted light images across all the treatment wells (3050). This reduces the cost and time required to run the drug screen by over 90% and prevents any phototoxic damage in the treatment wells. This method has the additional advantage of measuring cancer-selective drug response in more complex culture systems such as tumor fragments, where enrichment is not possible and each well may have a distinct mixture of normal and malignant cells. In some embodiments of the present invention, a fluorescent apoptosis reagent is added to all the wells to measure cell viability (3062), as described above. In such embodiments, in silico labeling is used to classify malignant and normal cells in each transmitted light image, and the apoptosis reagent is used to label each cell as apoptotic or non-apoptotic in the matched fluorescent image. These two steps result in a measurement of cancer-selective drug response (3060). In other embodiments of the present invention, the fluorescent apoptosis reagent is only added to the positive control wells containing the cytotoxic compound and fluorescent antibody cocktails (3062). In such an embodiment, the in-silico labeling model can be trained on matched fluorescent and transmitted light images from the positive control wells to learn visual features of apoptosis, and the model is used to label cells as apoptotic in the treatment wells based on the learned visual features of apoptosis (3060). In further embodiments of the present invention, the model is pre-trained on adatabase of transmitted light images of labeled malignant and normal cell types undergoing apoptosis (3061), and the model uses these learned features to label cells as apoptotic in the transmitted light images of the treatment wells. In such an embodiment, cancer-selective drug response is measured just using the transmitted light images (3060)

[0115] Methods for High-Throughput Cancer-Selective Clonal Drug Response

[0116] FIG. 33 illustrates embodiments of a high-throughput kinetic drug screen used to identify cancer-selective hits early and determine clonal drug response. A tumor specimen is isolated from a patient in order to provide cancer cells for drug screening (100). The tumor specimen can be isolated from a primary or metastatic site through conventional means, including blood draw, fine needle aspiration, tissue biopsy, or surgery. The cancer cells are incubated in a tissue culture system (100). The tumor specimen may have also contained non-cancerous cells, such as fibroblasts, epithelial or immune cells. Alternatively, healthy cells from a distinct sample can be incubated with said compounds, either together with the cancer cell or in parallel in a separate tissue culture system.

[0117] As shown in FIG. 33, a plate reader (3110) is utilized to take quick and frequent bulk measurements of cell viability in each well over time (700). These bulk measurements are used to identify potential drug hits, which are then nominated for further image analysis to measure cancer-selective drug response (3060) as described above. For drug hits confirmed to be cancer- selective by imaging (3060), clonal drug response is then measured as described above. Briefly, cells from each confirmed drug hit well are stained with an apoptotic marker binding reagent conjugated to a magnetic bead or fluorophore. In a preferred embodiment, treated cells are stained with Annexin-V magnetic microbeads and passed through a magnetic column (140). Viable (non-apoptotic) cells pass through the column into a first collection container. After removing the magnet, Apoptotic cells are then passed through into a second container. Singlecell sequencing is then performed on the cells (150). Cells from the first fraction (i.e. container) and second fraction (i.e. container) can be kept separate during this process, or labeled with a oligo-conjugated multiplexing reagent and pooled. For single-cell sequencing, the RNA molecules in each cell are labeled with a unique oligonucleotide identifier barcode that allows the sequencing reads to be attributed to a single cell. For example, the cells can be partitionedphysically using a microfluidic system such as the lOx Chromium system (150). Once partitioned, the RNA molecules are labeled with the unique oligonucleotide barcode. In another example, a combinatorial barcoding system can be used to instead of a microfluidic system. In such a system, the cells are separated into multiple wells and the RNA molecules are labeled with a first oligonucleotide barcode. The cells from each well are then pooled and re-separated into distinct wells and labeled with a second barcode. This process is performed multiple times to increase the probability that each cell is labeled with a unique combination of barcodes. Once the individual cells are labeled, all the cells are further processed through a series of standard library preparation steps for next-generation sequencing. The cells are then sequenced using, for example, an Illumina sequencing instrument (1 0).

[0118] A next-generation sequencing bioinformatics pipeline is then used to analyze all the reads, align them across the human genome, and assign them to an individual cell. Individual cells can then be analyzed to classify malignant and normal cells. Then, within the malignant cell population, each cell is assigned to a certain clonal population based on genomic mutational profile, transcriptomic expression profile, phenotypic markers, or a combination of the above. Once cell classification and clonal identification are complete, each cell is labeled as apoptotic or non-apoptotic. Cells that passed through the magnetic column into the first container will be labeled as apoptotic, while cells that pass through the column into the second container are labeled as non-apoptotic. The end result of the analysis is a population of normal cells and malignant clonal populations, with each cell labeled as apoptotic and non-apoptotic. The relative ratio of apoptotic versus non-apoptotic cells in each normal and malignant cell population provides a clonal drug response measurement (180).

[0119] As shown in FIG. 33, the present invention utilizes plate reader mode (3110) to proactively identify and nominate hits by quickly performing a fluorescent plate scan (apoptosis reagent) to measure bulk cell viability across the entire plate in ~6 minutes (~1 second per well). As described above, plate readers and imaging systems have distinct advantages and disadvantages for high-throughput drug screens. When imaging and plate reading modes are combined for proactive hit selection, the throughput of the drug screen can be dramatically increased. The time it takes to image a 384-well plate with three fluorescent channels (cancer markers, normal markers, and apoptosis reagent) can take 5-10 seconds per well and 30-60 minutes per plate. With the in-silico labeling method described above, only transmitted lightimages (+ / - one fluorescent channel for apoptosis) are required, which reduces the imaging time by up to 75%.

[0120] For solid tumor samples, the time savings relative to imaging is even more dramatic since plate reader mode quickly captures bulk fluorescence throughout the well, including solid tumor cells in three-dimensional format (spheroids or tumor fragments) or solid tumor cells that have lost adherence to the bottom of the well due to apoptosis. Treatment wells above a certain fluorescence threshold represent potential hits, which can then be imaged to measure cancer- selective apoptosis as described above. Cancer-selective hits are then proactively selected by transferring to the preservation buffer, while broadly toxic drugs are maintained in the drug screen. This combines all the advantages of the methods described above into one high- throughput and affordable drug screen that identifies cancer-selective drugs with single-cell clonal resolution.

[0121] FIG. 34 illustrates a flow diagram for embodiments of a high-throughput kinetic drug screen used to identify cancer-selective hits early and determine clonal drug response. Tumor cells are treated with a library of drug compounds (100), and bulk cell viability is measured frequently over time (700). Bulk cell viability is used to nominate potential drug hits based on a viability threshold (710), which can be a pre-defined, absolute, relative, or dynamic viability threshold. In some embodiments, the viability threshold can take into account a reference toxicity level for each drug compound, or compare viability levels in wells containing tumor cell cells with wells containing healthy control cells in the same screen. Nominated drug hits are then imaged to measure cancer-selective drug response (3060), as described above. Confirmed cancer-selective drug hits are then transferred to a preservation or fixation buffer (720) until the end of the drug screen. The hit identification (710), confirmation (3060), and transfer (720) steps are repeated until a certain number of hits are selected or until the end of the drug screen (730). Once the drug screen is complete, single cell sequencing is then performed on the cells from the selected hit wells along with the baseline tumor sample or DMSO negative control (740).

[0122] FIG. 35 illustrates a workflow comparison between conventional drug screen analysis and embodiments of a high-throughput kinetic drug screen used to identify cancer-selective hits early and determine clonal drug response. In most conventional drug screens, tumor cells are treated with a library of drug compounds (100) until a predefined assay endpoint (730). At the end of the assay, bulk cell viability (3210) or cancer-selective drug response (3211) is measured.For example, reagents like CellTiter-Glo and Caspase-Gio can be used to measure bulk cell viability (3210). Flow cytometry and microscopy using fluorescent reagents specific for cancer cell markers and apoptosis and / or necrosis markers can be used to measure cancer- selective drug response. Annexin-V and cell impermeable nuclear dyes such as 7- Aminoactinomycin D (7- AAD) are often used in flow cytometry to identify dying cells in early and late apoptosis. Cell impermeable nuclear dyes or cell permeable cytoplasmic dyes such as calcein AM are often used to identify apoptotic and viable cells in microscopy, respectively. Cell and nuclear morphology can also be used to measure cell viability in images without fluorescent reagents. In the present invention, tumor cells are treated with a library of drug compounds (100) and bulk cell viability is measured frequently over time (700) during the course of the assay. Bulk cell viability is used to nominate potential drug hits based on a viability threshold (710), which can be a pre-defined, absolute, relative, or dynamic viability threshold. In some embodiments, the viability threshold can take into account a reference toxicity level for each drug compound, or compare viability levels in wells containing tumor cell cells with wells containing healthy control cells in the same screen. Nominated drug hits are then imaged to measure cancer-selective drug response (3060), as described above. Confirmed cancer-selective drug hits are then transferred to a preservation or fixation buffer (720) until the end of the drug screen. The hit identification (710), confirmation (3060), and transfer (720) steps are repeated until a certain number of hits are selected or until the end of the drug screen (730). Compared to conventional drug screens, a primary advantage of the present invention is the early identification of cancer-selective drug hits and preservation of cell biology for downstream phenotypic analysis. Once the drug screen is complete, single cell sequencing is then performed on the cells from the selected hit wells along with the baseline tumor sample or DMSO negative control (740). This enables clonal drug response to measured across a library of drug compounds with diverse pharmacokinetic profiles and mechanisms of action.Methods of Pooling Drugs for a Drug Screen

[0123] Despite the extensive advantages of the methods described above, a major limitation to functional precision medicine remains the limited tumor tissue available from advanced cancer patients. While the clonal drug response enabled by the methods described above can be highly valuable for advanced cancer patients with aggressive and heterogeneous tumors, the number ofdrugs that can be screened limits the number of effective drugs and combinations that can be identified. In order to expand the number of drugs that can be screened on a finite number of tumor cells, pools of cancer drugs can be screened together and deconvoluted to identify the most active drug in a pool. This is achieved by combinatorial pool design, in which each pool contains a unique set of drugs, and no two drugs are present in the same sets of pools. At the end of the drug screen, the pattern of pools containing an active drug can be deconvolved based on which drug most closely matches the pattern. The compression and deconvolution are limited by the expected hit rate. For approved cancer drugs, the hit rate can be above 10% for any given tumor sample, limiting the compression level (i.e. number of drugs in each pool that can be deconvolved).

[0124] As shown in FIG. 36, a method for designing compressed drug libraries optimized for cancer drug screens on tumor cells is based on a reference database of past cancer drug screens (3320). The reference database (3320) contains drug-sample pairs comprising cancer drug compounds previously screened on cancer cell lines and tumor samples. Each drug-sample pair has cell viability measurements , and measurements below a certain threshold result in a “sensitive” label for the drug-sample pair (3331). This label can then be used to determine the cross-sensitivity level between two cancer drugs, which represents the level of co-occurrence versus mutual exclusivity in the sensitivity label across samples in the database. Each drugsample pair also has post-perturbation expression signatures associated with effective drug response (3332). In some embodiments, these expression signatures are identified using singlecell RNA sequencing. In other embodiments, the expression signatures are identified using bulk RNA sequencing or other “omic” phenotyping technologies. Finally, the reference database (3320) contains kinetic profiles of effective drug responses for each drug-sample pair (3333). Drugs with distinct pharmacokinetic profiles can be pooled together, and the kinetic drug response of the pool of drugs can be informative of which drug compound in the pool is having the largest effect on the tumor cells.

[0125] As shown in FIG. 36, sensitivity labels (3331), expression signatures (3332), and kinetic profiles (3333) are used together to calculate an optimization score (3330) for the purpose of designing compressed drug pools for drug screening. In some embodiments, cross-sensitivity is the first feature used to design compressed drug pools. In some embodiments, gene expression signature is the second feature used to design compressed pools of cancer drugs. Within the setof sensitive samples for a given drug compound, a gene expression signature is determined for the drug based on the differentially expressed genes in the sensitive samples. Differentially expressed genes could be determined by comparing the post-treatment expression profde of a sensitive sample to its pre-treatment expression profde, or by comparing its expression profde to the expression profde of non-sensitive samples (pre- or post-treatment). In some embodiments, pharmacokinetic profde of each drug compound is the third feature used to design compressed pools of cancer drugs . Within the set of sensitive samples for a given drug, a pharmacokinetic label is determined for the drug based on the measured viability over time. For example, a drug that has a TIC50 value (time to achieve 50% inhibition at the IC50 concentration) of 72 hours would be labeled as slow, 48 hours would be labeled as moderate, and 24 hours would be labeled as fast.

[0126] In the present invention, a set of drug compounds (3310) are received for the purpose of designing a compressed drug pool for screening drugs on tumor cells. Subsets of the drug compounds (3310) are allocated into pools (3340), wherein each pool has a unique subset of drugs, and no two drugs are in the exact same set of pools. In some embodiments, drug pools are designed by first minimizing the cross-sensitivity between drugs potentially allocated into the same pool(s). This set of drugs is further divided into multiple subsets by minimizing the overlap of gene expression signatures for the samples labeled as sensitive. Finally, each subset of drugs is further divided into unique pools based on the pharmacokinetic label of each drug. For example, a subset of 15 drugs is divided into unique pools of five drugs each, with each pool containing one or two drugs with the same pharmacokinetic label. The 15 drugs could result in five total pools, where each pool comprises a unique combination of drugs with distinct pharmacokinetic labels, and no two drugs are present in the same three (out of five) pools. A primary advantage of this method is the use of past drug screen results to identify cancer drugs that often co-occur as hits and distribute them into distinct groupings. A further advantage of this method is to increase the compression level and improve the deconvolution of hits by pooling drugs from a group based on distinct gene expression signatures and pharmacokinetic profiles when samples are sensitive to said drugs.

[0127] As shown in FIG. 37, optimized drug pools can be designed, compressed, and deconvoluted in a drug screen on a tumor sample. Compressed drug pools (3340) are designed such that each pool has a unique subset of drugs, and no two drugs are in the exact same set ofpools. These compressed drug pools comprise two or more drug compounds which are used to treat the same tumor cells in the same subset of wells in a compressed drug screen (3410). Hits are identified based on bulk cell viability or cancer-selective drug response (3420), as described above. Once all the hits have been identified (3420), hit compounds are deconvoluted based on their presence in drug pools identified as a hit, and absence in drug pools not identified as hits (3430). This pattern matching increases the probability that a drug exclusively present in hit pools is having the largest effect in said pools, since each pool comprises a unique subset of drugs, and no two drugs are in the same set pools. After deconvolution, the effect of each hit pool is assigned to the drug compound most likely to be having the largest effect in said pool (3440). After a hit compound has been identified for each hit pool, hit compounds can be ranked according to the magnitude of effect (i.e. cell viability level) for the cell cultures treated with a pool containing said hit compound (3450). A primary advantage of the present invention is the generation of a ranked list of hit compounds, equivalent to screening drugs individually, but with far higher throughput given a limited amount of tumor cells available for screening.Methods of Pooling Drugs for a Drug Screen Using a Prediction Model

[0128] In the compression method described above, pools with an effective drug will only be identified when the majority of cells in a tumor sample are sensitive to the most active drug in the pool. Since pools of drugs are screened, certain pools may also contain a drug that is effective for a minor clone in the tumor sample. However, this would have to occur by chance, and for pools with an active drug only effective for a minor clone, the pool would not appear to be a hit during the drug screen. In order to improve the probability of identifying hits to major and minor clones in a tumor sample, predictive models can be used to design compressed drug pools.

[0129] As shown in FIG. 38, such prediction models (3530) can be trained on previous drug screen results in the database described above (3320), including paired RNA expression and cell viability data from treated cancer cells. The viability data is measured at the end of the drug screen, while the RNA expression can be measured before treatment (i.e. baseline), after treatment (i.e. end of the drug screen), or both. The model (3530) learns features in the RNA expression data predictive of the cell viability post-treatment and uses those features to predict cell viability for the same drug compound on an unseen clonal population from a tumor sample.Single-cell RNA sequencing is performed on the tumor sample prior to the drug screen (3510) to identify clonal populations present in the tumor (3520). Once clonal populations have been identified, the RNA expression data for the cells from said clonal population are input into a prediction model (3530), which uses the RNA expression profile or signature to predict cell viability after treatment with a set of drug compounds (3310). The models can be used to predict cell viability for each clone-drug pair across all drugs with sufficient training data. The method then minimizes the probability that two drugs screened together in a pool will have an effect on the same clonal population, based on the results of the prediction model. This results in drug pools (3340) designed to include a first drug compound predicted to be effective against the major clone present in the tumor sample, which would lead to the drug pool being identified as a hit based on bulk or cancer-selective drug response, as described above, Within these hit pools, the prediction model (3530) increases the probability that the other drug compounds present in the pool are also effective, but specifically for the one or more minor clones identified in the tumor sample. Key advantages of the present invention not only include increased drug screen throughput given a limited amount of tumor cells, as described above, but also the identification of hit drug compounds for minor clones present in the tumor sample. A further advantage of the present invention is the identification of novel drug combinations for a given tumor sample, with associated toxicity data based on the measured cell viability of normal cells treated with the hit drug pool containing the effective novel drug combination.Definitions

[0130] The practice of the present invention may employ, unless otherwise indicated, conventional techniques of molecular biology, cell biology, synthetic biology, biochemistry, organic chemistry, medicinal chemistry, high-throughput screening, virtual screening, bioinformatics, machine learning, and generative neural networks, which are within the skill of the art. Such conventional techniques include, but are not limited to, automated liquid handling, robotic plate handling, emulsion droplet generation, cell transfection and transduction, fermentation and incubation, cell sorting and analysis, nucleic acid sequencing, biophysical measurement, and the like. Specific examples of appropriate techniques are provided for reference below. However, other equivalent conventional techniques can also be used. Such conventional techniques can be found in standard laboratory manuals.

[0131] Phenotypic analysis suitable for use in the present invention includes, but is not limited to, determining identity of cell surface markers, cell morphology, cell size, and cell expression profile at the RNA or protein level.

[0132] As used herein the term “therapeutic response” refers to any change in phenotype or lack thereof of any cell after contact with an exogenous compound.

[0133] As used herein the term “exogenous compound” refers to any compound that is introduced to the cell regardless of whether that compound was already intrinsic to the cell.

[0134] As used herein the term “labeling” refers to modifying a cell in such a way as the cell may be differentiated from cells that have not been labeled or that have been labeled in a distinct manner.

[0135] As used herein the term “apoptotic reagent” refers to any compound capable of causing apoptosis of a cell.

[0136] As used herein the term “apoptotic cell” refers to any cell that has underwent or began to undergo the process of apoptosis.

[0137] As used herein the term “apoptosis” refers to a regulated process by which a cell becomes irreversibly nonviable.

[0138] As used herein the term “necrotic” refers to an unregulated process by which a cell becomes irreversibly nonviable.

[0139] As used herein the term “sensitive” refers to a cell that upon contact with a compound becomes apoptotic and or necrotic. For example, a cell that is sensitive to compound A becomes apoptotic and or necrotic upon contact with compound A.

[0140] As used herein the term “resistant” refers to a cell that upon contact with a compound remains viable. For example, a cell that is sensitive to compound A remains viable upon contact with compound A.

[0141] As used herein the term “oligonucleotide barcodes” refers to specific DNA sequences capable of identifying a particular cell type.

[0142] As used herein the term “pool” refers to a group of cells that is comprises cells from two distinct groups.

[0143] As used herein the term “initial timepoinf ’ refers to the instant of time in which one or more compounds first make contact with a cell or a medium in which a cell resides.

[0144] As used herein the term “final timepoint” refers to the endpoint of the assay or drug screen wherein no further measurements are collected.

[0145] As used herein the term “transmitted light image” refers to any image created by passing light through an object.

[0146] As used herein the term “computer-implemented” refers to a method by which a computer, a computer network or a similar programmable apparatus is required to complete one or more steps via a computer program.

[0147] Unless otherwise expressly stated, it is in no way intended that any method set forth herein be construed as requiring that its steps be performed in a specific order. Accordingly, where a method claim does not actually recite an order to be followed by its steps or it is not otherwise specifically stated in the claims or descriptions that the steps are to be limited to a specific order, it is no way intended that an order be inferred, in any respect. This holds for any possible non-express basis for interpretation, including: matters of logic with respect to arrangement of steps or operational flow; plain meaning derived from grammatical organization or punctuation; and the number or type of embodiments described in the specification.

[0148] The disclosed embodiments are simply exemplary embodiments of the inventive concepts disclosed herein and should not be considered as limiting, unless the claims expressly state otherwise.

[0149] The following examples are intended to illustrate the present invention and to teach one of ordinary skill in the art how to use the formulations of the invention. They are not intended to be limiting in any way.EXAMPLESExample 1. Experimental Systems

[0150] The breast cancer cell lines MDA-MB-231, MCF-7, and MCF-7taml cells were cultured in DMEM supplemented with 10% fetal bovine serum (FBS), 1% amphotericin B, and 1% penicillin-streptomycin. Incubation was performed at 37°C and 5% CO2. The cells were grown in a T25 culture flask and evaluated with an inverted microscope for viability and contamination. When the cells filled at least 80% of the flask, they were removed with TrypLE Express and passaged (harvested, washed in HBSS, and cultured as described above).

[0151] The acute leukemia cell lines SUP-B15, REH, and NALM-6 cells were cultured in RPMI 1640 containing 2 mM stable glutamine (1-Ala-l-Gln dipeptide) supplemented with 10 or 20%fetal bovine serum (FBS), 20 mM HEPES, 1 mM sodium pyruvate) lx MEM non-essential amino acids, and 1% penicillin-streptomycin. Incubation was performed at 37°C and 5% CO2. The cells were grown in a T25 culture flask and evaluated with an inverted microscope for viability and contamination. When the cells reached a density of 2 million cells / mL, they were passaged (harvested, washed in HBSS, and cultured as described above).

[0152] Tumor samples from three metastatic breast cancer patients were acquired from a biobank. Each sample comprised at least 10 tissue fragments, each at least one cubic millimeter in size, viably frozen in cryovials. Sample mBreastOOl was a sample resected from the breast of a donor with metastatic apocrine carcinoma. Sample mBreastOOl was reported to be hormone receptor negative (HR-negative) and HER2 negative (HER2-negative). Sample mBreastOOl was reported to have the following genetic alterations: ABL1 p.Lys266Arg, KIT p.Met541Leu, and PIK3CA p.Glu545Lys. Sample mLiver002 was a sample resected from the liver of a donor with metastatic adenocarcinoma. Sample mLiver002 was reported to be estrogen receptor positive (ER-positive 67%), progesterone receptor positive (PR-positive 19%), and HER2 negative (HER2 -negative). Sample mLiver002 was reported to have the following genetic alterations: KDR p.Gln472His and PIK3CA p.Glu545Lys. Sample mChest003 was a sample resected from the chest wall of a donor with metastatic ductal adenocarcinoma. Sample mChest003 was reported to be hormone receptor positive (ER-positive and PR-positive) and HER2 amplified (HER2-positive). There was no genetic alteration information available for sample mChest003.

[0153] Tumor samples from three acute myeloid leukemia patients were acquired from a biobank. Each sample comprised at least 10 million cells viably frozen in cryovials. Sample AMLBM001 was a bone marrow sample with a reported blast fraction of 80% from a donor reported to be resistant to induction therapy (idarubicin + cytarabine). Sample AMLBM002 was a bone marrow sample with a reported blast fraction of 89% from a donor reported to be resistant to induction therapy (idarubicin + cytarabine). Sample AMLPB003 was a peripheral blood sample with a reported blast fraction of 58% from a donor reported to be resistant to venetoclax and azacitidine (two courses venetoclax; one course azacytidine; two courses azacytidine plus venetoclax).

[0154] For drug screens, the following table of drug compounds and concentrations were used in one or more of the experimental systems described above:

[0155]

[0156] For drug screens, cell lines were cultured and diluted to plating density in either supplemented DMEM (breast cancer cell lines and samples) or RPMI 1640 (leukemia cell lines and samples) media prepared as described above. Primary tumor samples were thawed and processed into single-cell suspensions using the protocol recommended by the biobank vendors. All cells were diluted to a plating density of 100k cells / mL. Cells were treated with soluble compounds at the stated concentrations in a final volume of 100 uL per well. Each drug compound was screened at two concentrations to confirm a dose-dependent treatment effect. Cells were screened in standard tissue culture incubation conditions (37 °C, 5% CO2) in a 96- well sterile opaque tissue culture plate. For 24-hour luminescence viability assays, white opaque plates were used with the CellTiter-Glo® and Caspase-Gio® 3 / 7 luminescent reagents (Promega). The luminescence reagents were added at the end of the 24-hour drug screen. For 72- hour fluorescence viability assays, black opaque plates were used with the CellEvent™ Caspase-3 / 7 Red fluorescence reagent (Thermo Fisher) and CellTox™ Green fluorescent reagent (Promega). For proactive hit selection during the 72 hour drug screen, Caspase Red and CellTox Green reagents were added to each well at the start of the drug screen and relative fluorescence units (RFU) were measured by plate reader for each well every 24 hours. For luminescence and fluorescence assays, raw data were reduced by subtracting mean background signal of blank wells not containing the luminescent or fluorescent reagent, respectively. For luminescent and fluorescent reagents, viability was calculated for each treatment condition based on the percent of maximal assay response using 100 uM staurosporine positive control wells. 100% of the maximal assay response resulted in a viability level of zero. For single-cell sequencing data, viability was calculated as the percent of total cells (passing quality control criteria postsequencing) annotated as non-apoptotic for each treatment condition. For all viability measurements, including luminescent reagents, fluorescent reagents, and the drug response annotation method, viability was normalized relative to the negative control wells (DMSO). This normalization step quantified and corrected for basal cell death in culture during the drug screen. Hits were identified based on the calculated viability level for the cancer wells (containing cell lines or tumor samples) compared to the healthy control wells for the same treatment condition. For the experiments described below, a hit was defined as a cancer well with 20% lower viability compared to healthy control peripheral blood mononuclear cells (PBMC) treated with the same compound.

[0157] At the earliest timepoint that a hit was identified, the full volume of the treatment wells containing said drug hit were transferred to an Eppendorf tube and spun down at 400 x g for 5 minutes. Supernatant was discarded and cells were resuspended in the OMICS-Guard preservation buffer (BD) and stored at -4 °C until the end of the drug screen. For adherent cells, TrypLE™ Express Enzyme (Thermo Fisher) was used prior to liquid transfer of the hit wells. At least two biological replicates for each treatment condition (compound-concentration) were used in each screen. If replicates resulted in distinct results, they were not used for further experimental steps (i.e. single-cell sequencing). During a first pooling process, replicate wells for each treatment condition were collapsed into one pooled sample. This involved the transfer of cells from replicate wells into one tube containing preservation buffer, as described above. A second pooling process was then performed to combine cells across treatment conditions into one sample for multiplexing. Cells from up to 12 treatment conditions were pooled and thentagged with distinct cell multiplexing oligonucleotides (BioLegend Total Seq) to increase throughput and decrease batch effects. After the second pooling step, cells from across the treatment conditions were stained with magnetic microbeads (Miltenyi Dead Cell Removal Kit) to annotate each sequenced cell as “apoptotic” or “non-apoptotic”. Non-apoptotic cells are present in the fraction of cells that flow through the magnetic column, while apoptotic cells and cellular debris (e.g. blebbing cells, necrotic cells, membrane fragments, etc.) are retained in the magnetic column. After magnetic separation, cells were counted (including trypan blue non- viable cells) and resuspended in IX PBS w / 0.04% BSA at a concentration of 1,000 cells per microliter. Cells were captured on a lOx Chromium controller using Single Cell 3' reagent chemistries (lOx Genomics). Samples were sequenced on a NovaSeq 6000 (Illumina) and sequencing data were processed using Cell Ranger v5.0.1 (lOx Genomics). The cellranger mkfastq pipeline was used to generate all demultiplexed FASTQ files from the raw sequencing data. Read counts were aligned to human GRCh38 genome reference. Cells were removed from the analysis if <200 or >8,000 distinct genes were detected, <1,000 counts were detected, or >15% of reads mapped to mitochondrial genes were detected. Multiplexing required a demultiplexing step during computational processing. Cells that could not be confidently assigned to a specific treatment condition were also removed from further analysis.Example 2. Measuring Clonal Drug Response

[0158] As shown in FIGS. 3-7, the present invention is effective at measuring clonal drug response in a heterogenous population of cancer cells. Cells treated with the highest dose of each drug for 24 hours were annotated and sequenced to measure clonal drug response across a library of cancer drugs with diverse mechanisms of action. The negative control DMSO wells were also sequenced. No cell line monocultures were sequenced. Drug response in each cell line monoculture was used as ground truth to confirm clonal identity and drug screen results in the heterogeneous mixture of cell lines. After sequencing is complete, the first step for measuring clonal drug response is to first identify cancer clones as described above. As shown in FIG. 3, cells from a heterogeneous mixture of breast cancer cells treated with DMSO for 24 hours show clustering by clonal identity (i.e. cell line). A UMAP was generated to visualize cell clustering (300). Color indicates identified clonal population, with clone 1 labeled as black (310), clone 2 labeled in dark gray (320), and clone 3 labeled in light gray (330). Cells were clustered based on gene expression profile and clonal populations were identified using the method describedabove. Clone 1 (310) was identified to be MDA-MB-231 cells based on negative receptor status, and clone 2 and clone 3 were differentiated hormone receptor pathway activity. Clone 2 (320) was identified to be MCF-7 cells based on high expression of genes involved in the hormone receptor pathway, while clone 3 (330) was identified to be MCF-7taml cells based on lower expression of genes involved in the hormone receptor pathway. These clonal identities were confirmed using drug response as ground truth for each cell line in monoculture. For the heterogeneous mixture of leukemia cell lines, no cell type markers were used to label each clonal population. Drug response was used as ground truth to confirm clonal identity by comparing drug screen results in monoculture to drug screen results in the mixture of leukemia cell lines.

[0159]

[0160] Once clonal populations are identified, the next step is to overlay apoptosis annotation on each clonal population (i.e. labeled cluster). Finally, each clone can be determined to be “sensitive” or “resistant” to a drug based on the number of apoptotic and non-apoptotic cells within a clonal population relative to the negative control and / or baseline tumor sample. Based on the low apoptotic fraction in the negative control wells (<5%), the treatment wells were compared only to the non-apoptotic fraction from the DMSO wells. This is equivalent to normalizing the viability reagents to the signal in the DMSO wells. As shown in FIG. 3, the cancer clones (i.e. cell lines) were identified in the negative control co-cultures at frequencies similar to the baseline input cell line ratios, demonstrating there was no differential cell loss or cell death observed between the cancer clones (i.e. cell lines).

[0161] As shown in FIG. 4, the present invention is accurate at measuring cell viability in singlecell sequencing data. Cells treated with the highest dose of each drug for 24 hours were annotated and sequenced to measure clonal drug response across a library of cancer drugs with diverse mechanisms of action. The negative control wells were also sequenced. After sequencing is complete, the first step for measuring clonal drug response is to first identify cancer clones as described above. The accuracy of the present invention was demonstrated using gold standard cell viability reagents in cell line mixtures (co-cultures) and cell line monocultures as ground truth (400). CellTiter-Glo (CTG) and Caspase-Gio (CG) viability reagents were used to provide two orthogonal measures of cell viability. In general, the drug response annotation method (A) resulted in a slightly higher viability level compared to CTG and CG reagents. CTG measures the quantity of adenosine triphosphate (ATP), a metabolite directly correlated to the number andmetabolic activity of viable cells. As cells start to undergo apoptosis, they stop generating ATP, resulting in reduced CTG signal. However, ATP loss can be delayed compared to the fraction of cells initiating apoptosis, resulting in higher viability measured by CTG compared to an apoptotic marker like CG. Moreover, ATP levels can be affected by growth inhibition without cell death, resulting in lower viability measured by CTG compared to apoptotic markers like CG. More specifically, the Caspase-Gio reagent (CG) and the drug response annotation method (A) represent direct comparison of apoptosis using orthogonal apoptotic markers (Caspase 3 / 7 and Annexin- V, respectively). The overestimation bias of the annotation method is likely due to the preferential loss of dying cells over viable cells throughout the experimental process. For example, apoptotic and necrotic cells can be more adherent and form clumps of cells, resulting in cell loss during liquid handling or retainment in the magnetic column even after removal of the magnet. Cells that stick together will also result in doublets that are filtered out during quality control processing of the single-cell sequencing data. Necrosis also leads to cells being filtered out based on one or more quality control criteria post-sequencing. For these reasons, the preferential loss of apoptotic cells leads to a consistent overestimation of cell viability by the drug response annotation method (A) relative to the viability reagents (CTG and CG) used in this comparison study. However, as shown in FIG. 4, this bias is not significant and can be accounted for due to consistency in the directionality of the bias. Furthermore, consistent overestimation of viability is compound-agnostic and provides a conservative estimate of drug effect. This results in fewer false positives (i.e. hits) than false negatives, which is preferrable for precision medicine applications. Assuming sufficient hits are identified in a drug screen of a tumor sample, it is more important to have confidence that each hit is truly having an effect on cell viability.

[0162] As shown in FIGS. 5 and 6, the present invention is accurate at measuring clone-level viability in single-cell sequencing data. The accuracy of the present invention was demonstrated using gold standard cell viability reagents in cell line monocultures as ground truth. However, viability reagents are not capable of measuring clone-specific drug response in a heterogenous sample of cells (i.e. tumor sample or mixture of cell lines). Therefore, single-cell sequencing was used to measure clone-specific response in the cell line mixture.

[0163] As shown in FIG. 5, certain drugs had differential effects on the distinct breast cancer clones (i.e. cell lines) in the heterogeneous sample (500). For example, clone 1 (MDA-MB-231) and clone 2 (MCF-7) were much more sensitive to azacitidine compared to clone 3 (MCF-7taml). Clone 1 was much more sensitive to paclitaxel compared to clone 2 and clone 3. Clone 2 was much more sensitive to alpelisib compared to clone 1 and clone 3. Clone 2 was also much more sensitive to fulvestrant compared to clone 1 and clone 3. This reflects the negative hormone receptor status of MDA-MB-231 (clone 1) and the known resistance of MCF-7taml (clone 3), since MCF-7taml is a cell line generated through long-term culture of MCF-7 cells in the presence of tamoxifen (a hormone therapy). In general, MCF-7taml displayed higher resistance across the drug library, which is a common observation in the clinic after a patient has been treated with multiple lines of therapy. Interestingly, high levels of Nectin-4 expression were observed for MCF-7taml and the parental cell line MCF-7 in the single-cell RNA sequencing data. Given these results, a combination of enfortumab (an antibody-drug conjugate targeting Nectin-4) and paclitaxel would be expected to have a strong therapeutic effect on this simulated heterogeneous tumor sample. In addition to measuring clonal drug response, the present invention also provides the advantage of measuring the expression and inferred activity of therapeutically-relevant targets and pathways.

[0164] As shown in FIG. 6, the present invention is accurate at measuring cell viability in singlecell sequencing data across solid tumor and hematologic cell types. Certain drugs had differential effects on the different leukemia clones (i.e. cell lines) in the heterogeneous sample (510). For example, clone 3 (NALM-6) was much more sensitive to alpelisib and resistant to azacitidine compared to clone 1 (SUP-B15) and clone 2 (REH). Clone 1 (SUP-B15) was much more sensitive to idarubicin, paclitaxel, and venetoclax compared to clone 2 (REH) and clone 3 (NALM-6). Given these results, a personalized combination of alpelisib, azacitidine, and venetoclax would be expected to have a strong therapeutic effect on this simulated heterogeneous tumor sample.

[0165] As shown in FIG. 7, the drug response annotation method was much more accurate at measuring clone-specific viability compared to expression-based viability signatures using the same single-cell sequencing data. It is evident that expression-based viability signatures are less accurate across compounds and cell lines (600). Expression-based viability values were generated using the CEVICHe web portal, which is an R Shiny app developed by Szalai et al. and available at https: / / saezlab.shinyapps.io / ceviche / . See, See, Szalai, B. et al. Signatures of cell death and proliferation in perturbation transcriptomics data-from confounding factor to effective prediction, Nucleic Acids Res 2019 Nov 4;47(19): 10010-10026. More specifically, the CTRP-L1000-24h linear model in the portal was used for MCF-7 cell line for select drugs with available viability predictions. Limited predictions were available for MDA-MB-231 cells, and no predictions were available for the MCF-7taml cell line or the leukemia cell lines. Expressionbased viability signatures are limited based on the training datasets used, thus, expression-based viability was only available for a subset of compounds and cell lines. As discussed in the original manuscript, expression-based viability signatures cannot be confidently predicted for compounds or cell lines not represented in the training datasets. Thus, the expression-based viability predictions used in this comparison represent the most accurate and confident predictions made by the model.

[0166] Subsequent to identifying clones (i.e. cell lines) in the single-cell sequencing data, clonelevel viability was measured using the present invention as described above and clone-level viability was predicted using the model available through the web portal described above. Clonelevel viability prediction values were accessed for matching doses at the 24-hour timepoint. For alpelisib, the viability prediction for the same dose (2 uM) was not available, so the average of the 1.11 uM and 3.33 uM dose was used as the viability prediction value. When multiple viability predictions were available for the same compound and dose at 24 hours, the average of the predictions was used. As shown in FIG. 7, expression-based cell viability prediction is much less accurate than the present invention. The general direction of prediction error is to overestimate viability. This is especially true for compounds that don’t act directly on the cell cycle (alpelisib, azacitidine, bortezomib), since the viability-related genes used in the prediction are predominantly cell cycle genes. For some compounds that work directly on cell cycle targets, such as cytarabine, the viability prediction was more accurate. For other compounds working directly on cell cycle targets, such as paclitaxel, the model underestimated viability. This is possibly due to a strong cell cycle suppression signature, but a lower level of apoptosis and cell death at the 24-hour timepoint. For drugs that act on epigenetic regulator targets, such as panobinostat, the model varies dramatically. The viability prediction for 1 uM panobinostat at 24 hours was much lower than the other methods, moreover, the prediction for 10 um panobinostat at 24 hours resulted in a range from 0.205 - 0.952 (data not show). This range is far above the viability range predicted for 1 uM panobinostat at 24 hours (0.065 - 271). Epigenetic regulators like panobinostat can cause broad changes in expression, likely driving the variability and inaccuracy in expression-based viability signatures. Finally, drugs that work directly onapoptotic targets, such as venetoclax, caused no change in viability (1 .0) according to the prediction model. This is in line with additional literature on expression-based viability signatures. See, McFarland et al. Nat Commun 2020 Aug 27; 11(1):4296.

[0167] A benefit of using gene expression signatures to predict cell viability is that viability can be measured directly in the single-cell sequencing data, with no need for additional reagents. Expression changes also occur at timepoints far earlier than cell death (e.g. 6 hours). However, the comparison above suggests that gene expression-based viability signatures are less accurate than the present invention, based on gold standard cell viability reagents used as ground truth. Although there is consistency in viability-related transcriptional signatures among certain subsets of compounds, these signatures vary substantially across the full panel of perturbations analyzed. This wide variety of perturbation-specific signatures highlight the heterogeneity that exists in expression programs induced by selective treatments, and the need for analyzing the relationship between expression and viability on a perturbation-specific level. Thus, expression signatures are a useful signal for understanding mechanism of action for a specific perturbation, but an unreliable and often inaccurate means for inferring cell viability in gene expression data. For many compounds the expression signature was cell-type specific, and for some drug-cell line combinations a significant transcription effect was correlated with no measurable impact on cell viability. Furthermore, these expression-based models perform even worse for new drugs with limited or no training data for discovering expression-based signatures. Finally, certain drug classes have specific mechanisms of action that do not result in transcriptional signatures at all, such as Bcl-2 inhibitors (e.g. venetoclax). This makes expression-based viability readouts inaccurate and often useless for measuring clonal drug response in drug screens.Example 3. Proactive HU Selection

[0168] As shown in FIG. 10, the real-time cell viability reagents Caspase Red and CellTox Green were added to each well at the start of the drug screen to measure apoptosis over time. Fluorescent reagents like the Caspase Red and CellTox Green reagent described above are fluorescently quenched until they bind to their nucleic acid substrate inside of cells. For CellTox Green, the impermeable reagent simply needs to pass through the disrupted membrane of a necrotic cell to emit a fluorescent signal. Loss of cell membrane integrity is an event that occurs late in the apoptosis process, when secondary necrosis occurs. For Caspase Red, the cell permeable reagent can pass through intact cell membranes and bind to activated Caspase 3 / 7 toemit a fluorescent signal. Caspase 3 / 7 activation is an event that occurs early in the apoptotic process. Caspase-based fluorescence reagents consist of a fluorogenic DNA dye that is covalently linked to a peptide that contains the caspase-3 / 7 DEVD recognition sequence. The inert, non-fluore scent reagent crosses the cell membrane where it is cleaved by activated Caspase-3 / 7 resulting in the release of a DNA dye and fluorescent staining of the nuclear DNA. The fluorescent signal is the result of an irreversible cleavage event of the fluorogenic substrate by active caspase-3 / 7, which DNA staining causes the fluorescent signal to persist even after caspase 3 / 7 is inactivated or degraded. The fluorescent signal increases with each cell undergoing apoptosis in the well, providing an accurate cell viability readout of relative fluorescence units (RFU) that can be measured frequently across treatment wells. However, excitation and emission of fluorescence reagents in each well of a high-throughput drug screen can be time-consuming. Frequent fluorescent measurements also cause phototoxic damage, confounding the effect of the drug treatment. Therefore, in some embodiments, only the positive control wells containing a high dose of a cytotoxic compound are measured at frequent timepoints early in the drug screen. Once a threshold of fluorescence (i.e. apoptosis) is reached in the cytotoxic control wells, fluorescence can be frequently measured across the treatment wells. This ensures early apoptotic timepoints are detected for rapid acting drugs in the library without unnecessary phototoxic damage.

[0169] As shown in FIG. 10, viability levels (800) were measured using Caspase Red and CellTox Green at 24 hours and 72 hours for the breast cancer cell line mixture, leukemia cell line mixture, and healthy control PBMCs as described above. Each sample was treated with 10 uM of staurosporine. Viability levels measured using Caspase Red were much lower at the 24 hour timepoint compared to CellTox Green. This is because Caspase Red fluorescence is activated early in the apoptosis process, while CellTox Green fluorescence is activated during the secondary necrosis process, as described above. By the 72 hour timepoint, viability levels measured using Caspase Red and CellTox green were similar for all samples, indicating that apoptosis was complete and cells were now necrotic. For this reason, Caspase Red (and other fluorescent reagents for early apoptotic markers) is superior for the proactive selection of hit compounds for downstream phenotyping.

[0170] As shown in FIGS. 11-20, the 24-hour drug screen described above was extended to 72- hours to allow the drug compounds in the library to demonstrate their full and distinctpharmacokinetic effects on the breast cancer and leukemia cell line mixtures. The Caspase Red reagent was added to each well at the start of the drug screen. Fluorescent measurements were taken every 12 hours in the positive control wells containing staurosporine. At 24 hours, the fluorescent intensity in the positive control wells indicated substantial apoptosis was occurring. Fluorescence measurements were then taken across all the treatment wells every 24 hours until the end of the 72-hour drug screen. When an effective drug hit was identified based on a cell viability differential of 20% compared to healthy PBMCs, a liquid transfer step was immediately performed to move the cells from the treatment wells to a preservation buffer that quenches the drug effect and preserves the transcriptomic phenotype of the cells for clonal analysis. If the drug hit was identified prior to 72 hours, replicate treatment wells continued to be screened to compare the results to the proactive hit selection wells. The earliest two hits in each drug screen were sequenced to demonstrate the performance of proactive hit selection relative to the 72 hour endpoint.

[0171] As shown in FIG. 11, the pharmacokinetic effects of the drug library were compared at 24 and 72 hours in the heterogeneous mixture of breast cancer cell lines (900). Certain drugs, including azacitidine, bortezomib, fulvestrant, and venetoclax, had an immediate effect. These drugs had similar viability levels measured at 24 and 72 hours. Other drugs had a much slower effect, including cytarabine, everolimus, idarubucin, paclitaxel, and panobinostat. These drugs resulted in much lower viability levels at 72 hours compared to 24 hours. As shown in FIG. 16, the pharmacokinetic profde of the drug library was consistent across breast cancer and leukemia cell lines (1400). These results demonstrate the potential for proactive hit selection of more rapid acting drugs to preserve the phenotype of sensitive cells undergoing cell death earlier in the course of the drug screen across solid tumor and hematologic cancers.

[0172] As shown in FIG. 12, the two earliest hits identified in the 72 hour breast cancer cell line screen were alpelisib (1010) and everolimus (1020). These two drug compounds achieved 20% lower viability at 48 hours compared to healthy control PBMCs. At 48 hours, cells from these hit wells were transferred into preservation buffer until the end of the 72 hour drug screen.

[0173] As shown in FIG. 13, proactive hit selection at the 48 hour timepoint resulted in superior cell recovery after single-cell sequencing for both alpelisib (1110) and everolimus (1120). At 48 hours, more total cells were recovered and more cells passed the quality control criteria for single-cell analysis. This results in superior measurement of clonal drug response by providingmore cells that can be included in the drug response analysis, especially for sensitive cancer cells responding to the drug compound.

[0174] As shown in FIG. 14, clone 2 (MCF-7) cells were more sensitive to both alpelisib (1210) and everolimus (1220), A preferential loss of clone 2 was observed after 48 hours likely due to secondary necrosis, highlighting the advantage of proactive hit selection for recovering sensitive cancer clones earlier in the drug screen. As shown in FIG. 15, this preferential loss of the sensitive clone 2 (MCF-7) cells results in discordant viability measurements using the drug response annotation method. As more cells from a specific clonal population are lost during the drug screen, the viability level for the sensitive clonal population is skewed higher due to the bias for non-apoptotic cells to survive until the end of the drug screen. Proactive hit selection at the 48 hour timepoint not only improved cell recovery, but also accuracy of the drug response annotation method when compared to viability measured using the Caspase Red reagent. This was demonstrated for both alpelisib, where drug response annotation accuracy was far superior at the 48 hour timepoint (1310) compared to the 72 hour endpoint (1311), and for everolimus, with superior accuracy observed at the 48 hour timepoint (1320) compared to the 72 hour endpoint (1321).

[0175] As shown in FIGS. 17-20, proactive hit selection improves clonal drug response measurement across solid tumor and hematologic cancer cells. As shown in FIG. 17, the two earliest hits identified in the 72 hour leukemia cell line screen were azacitidine (1510) and venetoclax (1520). These two drug compounds achieved 20% lower viability at 24 hours compared to healthy control PBMCs. At 24 hours, cells from these hit wells were transferred into preservation buffer until the end of the 72 hour drug screen.

[0176] As shown in FIG. 18, proactive hit selection at the 24 hour timepoint resulted in superior cell recovery after single-cell sequencing for both azacitidine (1610) and venetoclax (1620). At 24 hours, more total cells were recovered and more cells passed the quality control criteria for single-cell analysis. This results in superior measurement of clonal drug response by providing more cells that can be included in the drug response analysis, especially for sensitive cancer cells responding to the drug compound.

[0177] As shown in FIG. 19, clone 1 (SUP-B15) and clone 2 (REH) cell populations were more sensitive to azacitidine (1710), A preferential loss of clones 1 and 2 were observed after 24 hours likely due to secondary necrosis, highlighting the advantage of proactive hit selection forrecovering sensitive cancer clones earlier in the drug screen. As shown in FIG. 20, this preferential loss of the sensitive clones 1 and 2 results in discordant viability measurements using the drug response annotation method. As more cells from these clonal population are lost during the drug screen, the viability level for the sensitive clonal populations are skewed higher due to the bias for non-apoptotic cells to survive until the end of the drug screen. Proactive hit selection not only improved cell recovery, but also accuracy of the drug response annotation method for azacitidine when comparing the 24 hour timepoint (1810) to the 72 hour endpoint (1811).

[0178] As shown in FIG. 19, clone 1 (SUP-B15) and clone 3 (NALM-6) cell populations were more sensitive to venetoclax (1720), A preferential loss of clones 1 and 3 were observed after 24 hours likely due to secondary necrosis, highlighting the advantage of proactive hit selection for recovering sensitive cancer clones earlier in the drug screen. As shown in FIG. 20, this preferential loss of the sensitive clones 1 and 3 results in discordant viability measurements using the drug response annotation method. As more cells from these clonal population are lost during the drug screen, the viability level for the sensitive clonal populations are skewed higher due to the bias for non-apoptotic cells to survive until the end of the drug screen. Proactive hit selection not only improved cell recovery, but also accuracy of the drug response annotation method for venetoclax when comparing the 24 hour timepoint (1820) to the 72 hour endpoint (1821).

[0179] As shown in FIGS. 21-30, 72-hour drug screens were performed on breast tumor and acute myeloid leukemia (AML) samples. Since “ground truth” for primary tumor samples is more difficult to assess, drugs were selected for screening based on the known receptor status and therapeutically relevant genetic alterations for the breast tumor samples. For the AML samples, drugs were selected for screening based on the known treatment histories and associated treatment responses. The Caspase Red reagent was added to each well at the start of the drug screen. Fluorescent measurements were taken every 12 hours in the positive control wells containing staurosporine. At 24 hours, the fluorescent intensity in the positive control wells indicated substantial apoptosis was occurring. Fluorescence measurements were then taken across all the treatment wells every 24 hours until the end of the 72-hour drug screen. When an effective drug hit was identified based on a cell viability differential of 20% compared to healthy PBMCs, a liquid transfer step was immediately performed to move the cells from the treatment wells to a preservation buffer that quenches the drug effect and preserves the transcriptomic phenotype of the cells for clonal analysis. If the drug hit was identified prior to 72 hours,replicate treatment wells continued to be screened to compare the results to the proactive hit selection wells. The earliest hit in each drug screen was sequenced to demonstrate the performance of proactive hit selection relative to the 72 hour endpoint. DMSO wells from the proactive hit timepoint and 72 hour endpoint were also sequenced for each tumor sample.

[0180] As shown in FIG. 21 cells from an AML bone marrow sample (AMLBM001) treated with DMSO for 72 hours show clustering by cell type (normal or malignant) and clonal identity (major clone 1 and minor clone 2). A UMAP was generated to visualize cell clustering (1900). Color indicates identified cell type, with normal cells labeled as black (1910), major clone 1 labeled in dark gray (1920), and minor clone 2 labeled in light gray (1930). Cells were clustered based on gene expression profile and cell types were identified using the method described above. In brief, reference-based classification methods and blast markers from the pathology report (2000) expressed at the RNA level were used to label cell clusters as normal or malignant. This was necessary because unlike solid tumors, AML has lower levels of copy number alterations that can be inferred for malignant cell classification. Certain normal hematologic cell types can also have non-diploid copy numbers, such as megakaryocytes, making CNV challenging for classifying AML blast cells. After malignant cell classification, inferred CNV was able to identify distinct AML clones, as shown in FIGS. 21 and 22.

[0181] As shown in FIG. 22 three breast cancer and three AML tumor samples were screened and sequenced (2000). Clinical information was available for each tumor sample, including information from the pathology report related to receptor status, genetic alterations, blast markers, and blast fraction. Treatment history was also available for the AML samples. This clinical information represented ground truth for interpreting and validating the drug screen results. Normal cells and the top two clones were identified and quantified in each sequenced tumor sample (2010). Additional clonal population clusters could be further segmented, but the drug screen results did not appear to change with further segmentation (data not shown). No time-dependent transcriptional response was observed to affect the clustering and cell type identification in the DMSO treated negative control wells.

[0182] As shown in FIGS. 23-26, tumor samples from metastatic breast cancer patients were screened with alpelisib (0.2 uM, 2 uM), alpelisib / fulvestrant combination (0.2 uM / 1 uM, 2 uM / 10 uM), and paclitaxel (0.1 uM, 1 uM). As shown in FIG. 23, the drug screen results (2100) showed mBreastOOl to be sensitive to alpelisib and paclitaxel, but not fulvestrant (no significantdifference between alpelisib and alpelisib / fulvestrant). The observed alpelisib sensitivity and fulvestrant resistance is concordant with the reported genetic alteration PIK3CA p.Glu545Lys and negative hormone receptor status, respectively (2000). mLiver002 was observed to be sensitive to alpelisib and even more sensitive to the alpelisib / fulvestrant combination. This is in concordance with the reported genetic alteration PIK3CA p.Glu545Lys and positive hormone receptor status (2000). mChest003 was observed to be resistant to alpelisib but sensitive to fulvestrant, which is in concordance with the reported positive hormone receptor status (2000). mChest003 was also reported to be HER2-amplified, which may explain the lower sensitivity to fulvestrant and higher resistance to paclitaxel (2000).

[0183] As shown in FIG. 24, proactive hit selection occurred at 24 hours for all three metastatic breast cancer tumor samples. For mBreastOOl, alpelisib was identified as a hit at 24 hours (2210) and cells were transferred to preservation buffer until the 72 hour endpoint. For mLiver002, alpelisib / fulvestrant combination was identified as a hit at 24 hours (2220). For mChest003, alpelisib / fulvestrant combination was also identified as a hit at 24 hours (2230).

[0184] As shown in FIG. 25, cancer cells were classified and clustered into two distinct clonal populations for measuring clonal drug response. Proactive hit selection at 24 hours resulted in improved cancer cell recovery for all three metastatic breast cancer tumor samples. For mBreastOOl, both identified cancer clones appeared to be sensitive to alpelisib and demonstrated improved cell recovery at 24 hours compared to the 72 hour endpoint (2310). For mLiver002, one cancer clone appeared sensitive to the alpelisib / fulvestrant combination, while the other clone appeared resistant. Proactive hit selection resulted in improved cell recovery for the sensitive clone at 24 hours compared to the 72 hour endpoint (2320). For mChest003, one of the cancer clones appeared to be much more sensitive to the alpelisib / fulvestrant combination, resulting in improved cell recovery at 24 hours compared to the 72 hour endpoint (2330).

[0185] As shown in FIG. 26, proactive hit selection at 24 hours not only resulted in improved cell recovery of the sensitive cancer clones, but also more accurate viability measurements using the drug response annotation method. For mBreastOOl, both sensitive clones demonstrated much lower viability at 24 hours compared to 72 hours (2410), indicating cell loss after 24 hours skewed viability measurements higher. For mLiver002 (2420) and mChest003 (2430), the sensitive clone demonstrated much lower viability at 24 hours compared to 72 hours. Theseresults demonstrate that proactive hit selection results in superior cell recovery and clonal drug response measurements in cell lines and tumor samples.

[0186] As shown in FIG. 27-30, blood or bone marrow samples from acute myeloid leukemia (AML) patients were screened with azacitidine (1 uM, 10 uM), azacitidine / venetoclax combination (1 uM / 1 uM, 10 uM / 10 uM), and idarubicin / cytarabine combination (0.1 uM / 1 uM, 1 uM / 10 uM). As shown in FIG. 27, the drug screen results (2500) showed AMLBM001 to be sensitive to azacitidine, but not venetoclax, idarubucin, or cytarabine. The observed idarubicin / cytarabine resistance is concordant with the reported treatment history of relapse after induction therapy with idarubucin / cytarabine (2000). AMLBM002 was observed to be sensitive to venetoclax, but resistant to azacitidine, idarubucine, and cytarabine. The observed idarubicin / cytarabine resistance is concordant with the reported treatment history of relapse after induction therapy with idarubucin / cytarabine (2000). AMLPB003 was observed to be resistant to azacitidine and venetoclax, but sensitive to idarubucin / cytarabine. The observed azacitidine / venetoclax resistance is concordant with the reported treatment history of relapse after multiple courses of azacitidine / venetoclax therapy. The observed sensitivity to idarubucin / cytarabine may be explained by a lack of induction therapy in this donor, perhaps due to a clinical decision based on the frailty of the patient. Newly diagnosed AML patients that are too frail to receive induction therapy often receive azacitidine / venetoclax treatment instead.

[0187] As shown in FIG. 28, proactive hit selection occurred at 24 hours for azacitidine and azacitidine / venetoclax for AMLBM001 (2610) and AMLBM002 (2620), respectively. Proactive hit selection occurred for idarubucin / cytarbaine at 48 hours for AMLPB003 (2630). The low dose of idarabucin / cytarabine was selected for AMLPB003 due to the high toxicity of this combination.

[0188] As shown in FIG. 29, cancer cells were classified and clustered into two distinct clonal populations for measuring clonal drug response. The fraction of cancer cells identified in each AML sample (2010) was in agreement with the reported blast fraction from the pathology report (2000). Cancer fraction measured in the single-cell sequencing data was slightly higher likely due to non-malignant cell loss during freeze / thaw cycles for these cryopreserved samples. Proactive hit selection at 24 or 48 hours resulted in improved cancer cell recovery for all three AML tumor samples. For AMLBM001, both identified cancer clones appeared to be sensitive to azacitidine and demonstrated improved cell recovery at 24 hours compared to the 72 hourendpoint (2710). For AMLBM002, one cancer clone appeared sensitive to the azacitidine / venetoclax combination, while the other clone appeared resistant. Proactive hit selection resulted in improved cell recovery for the sensitive clone at 24 hours compared to the 72 hour endpoint (2720). For AMLPB003, one of the cancer clones appeared to be slightly more sensitive to the idarubucin / cytarabine combination, but both clones demonstrated improved cell recovery at 48 hours compared to the 72 hour endpoint (2730).

[0189] As shown in FIG. 30, proactive hit selection not only resulted in improved cell recovery of the sensitive cancer clones, but also more accurate viability measurements using the drug response annotation method. For AMLBM001, both sensitive clones demonstrated much lower viability at 24 hours compared to 72 hours (2810), indicating cell loss after 24 hours skewed viability measurements higher. For AMLBM002 (2820), the sensitive clone demonstrated much lower viability at 24 hours compared to 72 hours. For AMLPB003, both clones demonstrated lower viability at 48 hours compared to 72 hours, with the discordance in viability over time increasing for the more sensitive clone (2830). These results demonstrate that proactive hit selection results in superior cell recovery and clonal drug response measurements in across solid tumor and hematologic cancer samples.

[0190] FIG. 31 illustrates a clinical report (2900) summarizing clonal drug response and treatment recommendation for a heterogeneous population of cancer cells. In the report, the number of cancer clones and proportion of each clone can be listed (2910) to indicate the level of heterogeneity of the tumor sample. This helps inform the number of agents that may be needed to effectively treat this patient’s cancer. The report also ranks the top drug hits for each clone identified in the tumor sample (2920). The clinician can use these rankings to design a personalized treatment combination unique to the patient’s tumor heterogeneity (2930). These personalized treatment combinations can also be designed based on evidence of drug interactions or toxicity when certain agents are used in combination. This information can help the clinician decide between two potentially effective combination options, or even choose to sequence the top ranked drugs in order to minimize the risk of toxicity.

Claims

WHAT IS CLAIMED IS:

1. A method of determining therapeutic response of cancer clones in a tumor cell sample, the method comprising the steps of a. treating a subset of tumor cells from the tumor cell sample with a compound to create a treated subset and an untreated subset; b. labeling the treated and untreated subsets with an apoptosis reagent conjugated to a fluorophore or magnetic bead; c. using the apoptosis reagent to physically separate the tumor cell sample into a first fraction and a second fraction, wherein the first fraction are non-apoptotic cells and the second fraction are apoptotic cells; d. sequencing nucleic acids isolated from the first fraction to provide sequence reads from which cancer clones can be determined; e. sequencing nucleic acids isolated from the second fraction to provide sequence reads from which cancer clones can be determined; and f. determining cancer clones that are sensitive to the compound as cancer clones whose frequencies are lower in the first fraction relative to the frequencies of the same cancer clone in the untreated subset.

2. The method of claim 1, further comprising the step of determining cancer clones that are sensitive to the compound as cancer clones whose frequencies are higher in the second fraction relative to the frequencies of the same cancer clone in the untreated subset.

3. The method of claim 1, further comprising the step of determining cancer clones that are resistant to the compound as cancer clones whose frequencies are higher in the first fraction relative to the frequencies of the same cancer clone in the untreated subset.

4. The method of claim 1, further comprising the steps of: g. labeling cells from the first fraction and cells from the second fraction with different oligonucleotide barcodes; and h. combining the cells from the first fraction and the cells from the second fraction into a single pool.

5. A method for determining the effect of a compound on a population of cells using singlecell sequencing, comprising: f. providing a sample containing a population of cells; g. incubating the population of cells in the presence of a test compound;h. staining the population of cells with one or more response marker reagents; i. physically separating the population of cells into a sensitive population and a resistant population based on the response marker reagents; and j. performing single-cell sequencing on both the sensitive population and resistant population of cells.

6. A method of analyzing cell viability of a treated tumor cell sample, the method comprising the steps of: a. measuring cell viability over a time series in a plurality of cell cultures comprising cells from the treated tumor cell sample, wherein the tumor sample is being treated with multiple compounds and wherein the times series comprises an initial timepoint (TO), one or more intervening timepoints (Th) and a final timepoint (Te); b. identifying one or more cell cultures from the treated tumor cell sample comprising an effective compound as a cell culture whose viability is below a viability threshold at a timepoint Th; c. transferring the one or more cell cultures treated with the effective compound at the intervening timepoint Th to a preservation or a fixation buffer; d. repeating steps a-c until the final timepoint Te; and e. performing phenotypic analysis of the cells transferred at each timepoint Th,7. The method of claim 6, wherein the analysis occurs after the final timepoint Te.

8. The method of claim 6, wherein the viability threshold is based on the ratio of malignant and normal cells in the tumor cell sample.

9. The method of claim 6, wherein the viability threshold is based on a viability level measured in a third set of cell cultures containing non-malignant cells.

10. The method of claim 6, wherein at least one of the one or more intervening timepoints Th is prior to secondary necrosis.

11. A computer-implemented method of training and using a statistical model or a machine learning algorithm for determining cell type and or predicting cell type for a plurality of cells from a biological sample, comprising: a. creating a first training set of transmitted light images of labeled cell types;b. receiving a second training set of matched transmitted light channel images and fluorescent channel images from a first set of cell cultures containing cells from a biological sample; c. training the statistical model or the machine learning algorithm for classifying cell types in the first set of cell cultures using cell type labels from the fluorescent image channel to label cell type in the transmitted light channel on the same matched images; and d. using the statistical model or the machine learning algorithm from step c to label cell types in transmitted light images from a second set of cell cultures containing cells from the same biological sample.

12. The method of claim 11, wherein the cells of the cell culture are alive during step b.

13. A computer-implemented method of analyzing cell culture image data for a drug screen, the method comprising: a. accessing image data comprising a set of fluorescent images and matched non- fluorescent images from the same regions of interest in a plurality of cell cultures containing a cytotoxic compound, wherein the fluorescent image data comprises a first signal associated with a type of cell and a second signal associated with viability of a cell; b. using an image analysis algorithm to label each cell in each fluorescent image and matched non-fluorescent image as malignant or normal using the first signal; c. using an image analysis algorithm to label each cell in each fluorescent image and matched non-fluorescent image as viable or dying using the second signal; d. using the labeled cells from step c. in the matched non-fluorescent images to train one or more statistical models for a task of labeling each cell as normal or malignant and viable or dying in a non-fluorescent image; e. accessing image data comprising a second set of non-fluorescent images from a plurality of cell cultures treated with distinct drugs; f. using the statistical models to label each cell as normal or malignant and viable or dying in the second set of non-fluorescent images; g. quantifying the number of cells that are viable and malignant out of the total number of malignant cells in each image of the second set of non-fluorescent images; andh. calculating a drug response metric based on the quantification from step g. for each of the plurality of cell cultures.

14. A computer-implemented method for selecting a subset of drug compounds from a library of drug compounds comprising: a. receiving a request to evaluate a group of drug compounds; b. accessing a database comprising drug screen results for the group of drug compounds, wherein the drug screen results comprise test results from a plurality of samples treated with the group of drug compounds comprising cancer cell lines, tumor clones, and normal cell types and wherein the database comprises viability measurements, drug response labels, and RNA expression profdes for each pair of drug and sample; c. determining co-occurrence and mutual exclusivity of sensitivity labels of each drug compound of the group of drug compounds across each sample in the database; d. determining similarity in expression signatures of each drug compound of the group of drug compounds across each sample labeled as sensitive to each drug compound in the database; e. determining similarity of viability measurements over time of each drug compound of the group of drug compounds across each sample labeled as sensitive to each drug compound in the database; f. calculating an optimization score using the samples labeled as sensitive to each drug compound, wherein the optimization score minimizes overlap of samples labeled as sensitive in the database, similarity of expression signatures in samples labeled as sensitive for each drug compound, and similarity in viability measurements over time for samples labeled as sensitive for each drug compound; and g. assigning drug compounds into the subset of drug compounds, wherein assignment into the subset of drug compounds is based on the optimization score, and wherein the subset of drug compounds is to be evaluated in parallel for an effect on tumor cells.

15. The method of claim 14, wherein each subset of drug compounds selected contains a unique set of drug compounds from every other subset of drug compounds selected and whereineach subset of drug compounds selected contains no more than one drug compound in common with any other subset of drug compounds selected.

16. A computer-implemented method for selecting a subset of drug compounds from a library of drug compounds, comprising: a. receiving a request to evaluate a group of drug compounds; b. accessing a database containing single-cell RNA expression data for a tumor sample; c. identifying one or more clonal populations in the tumor sample based on the single-cell RNA expression data; d. running one or more prediction models for the group of drug compounds on each clonal population, wherein the one or more models predict viability of the clonal population after treatment with each drug compound; and e. assigning one or more drug compounds from the group of drug compounds into the subset of drug compounds, wherein the drug compounds in the subset of drug compounds have an effect on distinct clonal populations from the one or more clonal populations, wherein assignment into the subset of drug compounds is based on the predicted viability for each drug compound, wherein the subset of drug compounds is to be evaluated in parallel for an effect on tumor cells.

17. The method of claim 16, wherein each subset of drug compounds selected contains a unique set of drug compounds from every other subset of drug compounds selected and wherein each subset of drug compounds selected contains no more than one drug compound in common with any other subset of drug compounds selected.

Images