PERCEPTION accurately predicts patient response and resistance to treatment using single-cell transcriptomics of their tumors

Abstract

Tailoring optimal treatment for individual cancer patients remains a significant challenge. Towards this, we developed PERCEPTION (PERsonalized Single-Cell Expression-Based Planning for Treatments In ONcology), a precision oncology computational pipeline. Our approach uses publicly available matched bulk and single-cell (sc) expression profiles from large-scale cell-line drug screens. These profiles help build treatment response models based on patients’ sc-tumor transcriptomics. PERCEPTION demonstrates success in predicting responses to targeted therapies in cultured and patient-tumor-derived primary cells, as well as in two clinical trials for multiple myeloma and breast cancer. It also captures the resistance development in lung cancer patients treated with tyrosine kinase inhibitors. PERCEPTION outperforms published state-of-the-art single-cell-based and bulk-based predictors in all clinical cohorts. PERCEPTION is accessible at https://github.com/ruppinlab/PERCEPTION. Our work, showcasing patient stratification using their tumor’s sc-expression profiles, will encourage the adoption of sc-omics profiling in clinical settings, enhancing precision oncology tools based on sc-omics.

Introduction

Precision oncology has made important strides in advancing cancer patient treatment in recent years, as described in several reviews^1–6. Much of the focus in the field has been on efforts to use FDA-approved sequencing assays to identify “actionable” mutations in cancer driver genes, to match patients to treatments¹. These efforts have been further boosted by the progress made in DNA-based liquid biopsies, which further can help guide and monitor treatment^7–9. However, a large fraction of cancer patients still do not benefit from such targeted therapies, and efforts are hence needed to find ways to analyze other molecular omics data types to benefit more patients. Addressing this challenge, recent studies have begun to explore the benefit of collecting and analyzing bulk tumor transcriptomics data to guide cancer patient treatment^10–17. Expression-based studies have demonstrated potential to complement DNA sequencing approaches in increasing the benefit of omics-guided precision treatments to patients.

One key limitation of current genomic and transcriptomic treatment approaches is that they are mostly based on bulk tumor data. Tumors are typically heterogeneous and composed of numerous clones, making treatments targeting multiple clones more likely to diminish the likelihood of resistance emerging due to clonal selection, and hence potentially enhancing the overall patient’s response¹⁸. Intra-tumor heterogeneity has been driving two major developments in recent years, the search for effective treatment combinations and the advent of single-cell profiling of the tumor and its microenvironment.

Large-scale combinatorial pharmacological screens have been performed in patient-derived primary cells, xenografts, and organoids and have already given rise to numerous combination treatment candidates^19–21. Concomitantly, the characterization of the tumor microenvironment via single-cell omics has already led to important insights regarding the complex network of tumor-microenvironment interactions involving both stromal and immune cell types¹⁸. It also offers a promising way to learn and predict drug response at a single-cell resolution. The latter, if successful, could guide the design of drug treatments that target multiple tumor clones disjointly^22–24 and help us understand the ensuing resistance to better overcome it. However, building such predictors of drug response at a single cell (sc) resolution is currently challenging due to the paucity of large-scale preclinical or clinical training datasets. Previous efforts, including a recent computational method termed Beyondcell that identifies tumor cell subpopulations with distinct drug responses from single-cell RNA-seq data for proposing cancer-specific treatments, have focused on preclinical models but lack validation in patients at the clinical level^24–28. Additional efforts to identify biomarkers of response and resistance at the patient level using single-cell expression are emerging for both targeted- and immuno- therapies, with remarkable results^29–31. However, to date, harnessing patients’ sc tumor transcriptomics for tailoring patients’ treatment in a direct, systematic manner has remained an important open challenge.

Aiming to address this challenge, here we present a precision oncology framework for PERsonalized Single-Cell Expression-based Planning for Treatments In ONcology (PERCEPTION). This approach builds upon the recent availability of large-scale pharmacological screens and sc-expression data in cancer cell lines to build machine learning-based predictors of drug response based on the gene expression of single cells. We first show that PERCEPTION can predict the response to single and combination treatments in three independent screens performed in cancer and patient-tumor-derived primary cells, based on their sc-expression profiles. Secondly, we show that PERCEPTION can stratify the responders vs. non-responders in two cohorts, multiple myeloma and breast cancer, with patients’ tumor sc-expression profiles and can capture the development of resistance using longitudinal tumor sc-expression profiles during treatment in a cohort of lung cancer patients. Notably, PERCEPTION markedly outperforms state-of-the-art single-cell-based and bulk-based predictors in all three sc clinical cohorts considered. Finally, we provide a guide for using PERCEPTION for a new clinical cohort with sc-expression to select patients for receiving treatment. In sum, we present a first-of-its-kind computational approach that showcases the exciting potential of sc-gene expression-based precision oncology.

Results

Overview of PERCEPTION

To predict patient response to therapy from the tumor’s sc-expression profile, we built a three-step machine learning pipeline called PERCEPTION (Figure 1A, a detailed description is provided in Methods). One of the key challenges in building a supervised machine learning model to predict clinical response using sc-expression is the lack of large-scale sc-expression data with clinical response labels. To overcome this, we utilized the concept of transfer learning, a machine learning technique where a model trained on one task (for which considerable data are available) is used as the starting point for a model on a second related task for which less training data are available). Transfer learning allows the second model to benefit from the knowledge learned by the first model.

Figure 1. Overview of the PERCEPTION framework and its performance during cross-validation.

(A) PERCEPTION builds drug-specific models in three steps: (i) Build bulk-expression response models based on drug response data measured in large-scale drug screens performed on cancer cell lines and their matched bulk expression. (ii) Building sc-expression models by tuning the bulk-expression models by determining the optimal number of genes used as predictive features that maximize its prediction performance based on sc-expression of cancer cell lines. (iii) In the third and final step, we predict the clinical response in patients following a three-step heuristic procedure: Given a patient scRNA-seq from the tumor, identify the major cancer cells clusters (called a transcriptional clone) and its mean expression. Use this mean expression as an input to the PERCEPTION model built at Step 2, yielding a predicted drug response for each transcriptional clone separately. The minimum response among all the clones is predicted to be the patient’s response. (B) The number of PERCEPTION predictive models of FDA-approved drugs (y-axis), when built from sc-expression (blue), bulk-expression (red), and pseudo-bulk, as a function of the Pearson correlation between predicted and observed response values (x-axis, the dashed horizontal line denotes the 0.3 threshold selected). (C) The distribution of predictive performance (x-axis) of the models. In the boxplots, the center line, box edges, and whiskers denote the median, interquartile range, and the rest of the distribution, respectively, as in standard box plots. Interestingly, the predictive performance is overall considerably higher for targeted therapies than for chemotherapies. A two-sided Wilcoxon rank sum test was performed to compare groups with N=44 drugs.

We built PERCEPTION response prediction models for each drug in three steps: Step 1: a bulk-expression model is trained to predict drug response in cell lines from the large-scale bulk-RNA-seq. In Step 2 (tuning), the bulk-expression models are tuned using the cell-lines’ sc-expression and drug response to build sc-expression models. In the final Step 3, we identify a heuristic strategy to predict clinical response by analyzing a clinical cohort with treatment response and sc-expression. For a given drug, we provide the input of its drug response and matched bulk-expression in cell lines for the first step, matched drug response and sc-expression in cell lines, and finally, sc-expression of the patient in the third step.

To gather cell lines data for building the predictors in the first two steps, we mined bulk expression³² and drug response profiles (PRISM) of 488 cancer cell lines (Table S1) from the DepMap database³³. The sc-expression profiles of these cell lines (N=205, Table S1) have been obtained from reference 34. Drug efficacy (also referred to as ‘viability’) is measured via the area under the curve (AUC) of the viability-dosage curve, where lower AUC values indicate increased sensitivity to treatment (Table S1).

For a given drug, PERCEPTION uses the above data to build a drug-specific response predictor in cell lines via the following two steps: Step 1: Building Bulk-expression models: We first build a linear model with elastic net regularization of drug response using the bulk expression and drug response data available for 318 PRISM cancer cell lines from 21 cancer types (Extended Figure 1A). Step 2: Building SC-Expression Models: The goal of this step is to build sc-expression-based prediction models of drug response. To this end, we determine the number of genes used as predictive features (hyperparameter tuning) that maximize the ability to predict the response from sc-expression data, analyzing the 169 cancer cell lines where both scRNA-seq profiles and drug response data are available (Extended Figure 1B). To evaluate the performance of a sc model in a cell line, PERCEPTION predicts the response to a given drug for each of its individual cells, and the mean response over all those individual cells is taken as the predicted sc-based response of that cell line to that specific drug. The output of this machine learning pipeline is hence a drug-specific sc response model and a quantification of its predictive accuracy from sc-expression in cell lines. We evaluate this model’s performance in an unseen test subset of the cell lines, employing a standard leave-one-out (one cell line) cross-validation procedure (Methods). As described in Methods, the models for some drugs will be deemed sufficiently predictive and the models for other drugs will not. Only drugs with predictive models from Step 2 are considered in Step 3.

In the third and final Step 3, we predict the clinical response in patients, the main goal of our study. This is done using the following heuristic procedure: (a) We first identify the major cancer cells clusters in the patient’s tumor using the sc-expression (transcriptional clones, a cluster of single cells whose transcription profile looks similar). (b) We then compute the mean expression of each transcriptional clone and use this as an input to the predictive drug-specific models yielded from Step 2 to predict drug response for each transcriptional clone (if a combination of drugs is used in the treatment, we take the maximum predicted killing among those drugs as the predicted killing effect on that clone, following the independent drug action (IDA) principle³⁵. (c) Finally, the overall patient response is predicted as the minimum response among all clones, taking the stance that the clone predicted to be most resistant will likely determine the overall clinical response. As we describe later in Results, this prediction strategy was determined in a multiple myeloma patient cohort by studying five different potential strategies and was then fixed and applied as is to two other patient cohorts. For any new drug in a new cancer type cohort, the response model (Steps 1 and 2) should also be built using all the cell lines available in the screen (pan-cancer model), as we found that this pan-cancer model construction performs better during cross-validation than building cancer type-specific models that use cell lines belonging only to the patient’s cancer type (Extended Figure 1C).

We test and demonstrate PERCEPTION’s performance in predicting the response to monotherapy and combination treatments in screens performed in cancer and patient-tumor-derived primary cells. Then, focusing on patient data, the main goal of this investigation, we study its ability to predict treatment response in two clinical cohorts and to predict the emergence of resistance in a third clinical cohort. We additionally compare PERCEPTION’s prediction performance with published state-of-the-art single-cell-based and bulk-based methods. Finally, we provide a guide for using PERCEPTION to predict response in new data sets.

Cross-validation and independent performance in cell lines

We applied PERCEPTION to build response models for 133 U.S FDA-approved oncology drugs tested in the PRISM drug screen (Table S2, Extended Figure 1D) and computed their performance to predict response in a leave-one-out cross-validation and 10-fold-cross validation. Prediction performances for each of these drugs are provided in Figure 1B. We deemed models to be sufficiently predictive if the Pearson correlation between their predicted (mean sc-response per cell line) vs. the observed viability on the test data was greater than 0.3. This threshold was chosen as it corresponds to the mean cross-screen replicate correlation observed among three major pharmacological screens and confirmed by us as well as previously reported (average cross-platform correlation across GDSC³⁶, CTD³⁷ and PRISM³⁸ ~ 0.30). We were able to build predictive models for 33% of the drugs tested (44 out of 133 drugs, Table S2, Figure 1B). The mean performance of PERCEPTION leave-one-out cross-validation and 10-fold-cross validation are 0.39 and 0.36 for 44 drugs with predictive models (Figure 1B). Studying this subset where we are able to build predictive models, we found that the drugs in this subset are more likely to be targeted therapy (mean Pearson Rho = 0.43 vs. 0.35 for chemo), have a higher variance in response profile (Wilcoxon P = 5E-7) and bimodality index in their response profile during training (reflecting the presence of both sensitive and resistant cell lines).

Studying the predictive accuracy of these 44 predictive models in a cross-validation manner for different kinds of transcriptomics inputs, including sc, bulk, and pseudo-bulk-expression (generated by summing up the gene-mapped reads across single cells, Methods), we reassuringly find that the predictive performance of PERCEPTION for sc-expression as inputs on these cell lines is comparable to that performance obtained using bulk-expression or pseudo-bulk as inputs (Figure 1C). Importantly, we note that a model built on only scRNA-seq without any pretraining on bulk-RNA-seq has markedly lower prediction accuracy (Pearson Rho = 0.22 vs. 0.39 for the 44 predictive drugs in the left-out test cell-lines) highlighting the importance of pre-training on bulk. We visualized PERCEPTION’s predicted killing levels at single-cell resolution for eight FDA-approved drugs with the high-confidence mechanism of action and the activity of the pathway they are targeting and provided in Extended Figure 2.

We next asked what are the identities of the genes that these 44 models are using to predict drug response. An average of 76 genes are used as features in the above models after regularization, where the key pathways enriched include apical junction pathway, including genes like ABCB1, encoding multi-drug resistance 1 (MDR1) a transporter implicated in resistance to many drugs, cell-cycle-related targets and more (Extended Figure 1E).

We next evaluated PERCEPTION’s performance on three independent large-scale cell-line screens, two cultured (Nair et al.³⁹ and GDSC) and one patient-derived (PDC), to stratify the resistant vs sensitive cell lines (Top vs Bottom 33% by viability, respectively). We built PERCEPTION models for each drug across the three screens individually. We note that we were unable to build predictive models for any drugs in the PDC screens using PRISM data and thus used GDSC data ~800 cell lines. Detailed methods on how PERCEPTION models were built and used are provided in Methods 3.1–3.3. PERCEPTION was able to stratify the resistant vs. sensitive with an average AUC of 0.81 (AUC of 0.87 for cultured cell lines, Extended Figures 3A–G, and 0.75 for PDCs, Extended Figure 3H–K). A detailed performance evaluation including drug-level performance measures is provided in Extended Figures 4–5. Predicted and observed viability are also strongly correlated in all three datasets (Pearson Rho=0.36 for Nair, 0.28 for GDSC and 0.64 for PDCs, Extended Figures 5–6). We note that a control PERCEPTION model that is not tuned on sc-expression yielded a modestly inferior performance in this test (Average AUC=0.71, for cultures cell lines =0.81, for PDCs=0.62).

Predicting treatment response in a multiple myeloma trial

After showing that PERCEPTION’s cell-line-based model can predict response of monotherapy and combination in cultured and patient-derived cell lines, we next turn to learn how can we utilize the cell-lines based models to predict patient response using their pre-treatment SC transcriptomics from their tumors. To this end, we mined the largest such dataset published to date, including data for 41 multiple myeloma patients. The patients were treated with a DARA–KRD combination of four drugs - daratumumab (monoclonal antibody targeting CD38), carfilzomib (proteasome inhibitor), lenalidomide (immunomodulator), and dexamethasone (anti-inflammatory corticosteroid)²⁹. The sc-expression and clonal (transcriptional cluster) composition and treatment response labels, as determined in the original study²⁹, were available for 28 tumor samples of these patients (Figure 2A). Patient response was measured via tumor size estimates in radiological images.

Figure 2: PERCEPTION predictions of DACA-KRD combination therapy in multiple myeloma patients.

(A)Distribution of abundance of the transcriptional clones (y-axis) in each multiple myeloma patient (x-axis), where the color code for the clones is provided at the top. (B) Predicted viability of the combination at a clonal level for each patient, where the response status is provided at the bottom strip of each facet. The left-to-right order of patients is the same as in panel A. (C) The stratification performance in distinguishing responders vs. non-responders from the clone-level predicted response information (y-axis) of five different strategies (x-axis). (D) The predicted combination response in 28 multiple myeloma patients stratified by responder (N=21) vs. non-responder (N=7) status. A two-sided Wilcoxon rank sum test was performed to compare groups. Box plot shows median (center), 25th and 75th percentiles (bounds of box), and minima and maxima (whiskers). (E) Receiver Operating Characteristic curve displaying the predicted combination response. The area under this curve, provided at the bottom right corner, denotes the overall stratification power in distinguishing responders vs. non-responders.

As explained above in the PERCEPTION overview, to predict the clinical response from tumor’s sc-expression, PERCEPTION first finds the major transcriptional clones (provided by original publication²⁹) and predicts the treatment response for each clone separately (Methods, response is defined as the predicted reduction in viability after treatment). Figure 2B shows the predicted viability of the combination at a clonal level for each patient. We designed and tested five different strategies to predict the clinical response from clone-level killing to find the most optimal strategy. We tested their performance for stratifying responders (N=7) vs non-responders (N=21) (Figure 2C, Methods). Briefly, the clinical response of a patient is determined by computing one of the following strategies: 1. Weighted average response: an average of response across all the clones weighted by their abundance in the tumor; 2. Unweighted average response: an average of response across all the clones; 3. Most-sensitive clone response: the response of the most-sensitive clone, that is the clone with the highest predicted response; 4. Unweighted Most-resistant clone response: the response of the most-resistant clone, that is, the clone with the least response. 5. Most-resistant clone response: the response of the most-resistant, weighted by its abundance proportion. This resulting AUCs for these strategies were 0.59, 0.55, 0.64, 0.75 and 0.83, respectively (Figure 2C). This analysis revealed that the fifth strategy best predicts the clinical response. In cell lines, this strategy also stratified resistant vs. sensitive, however, with lower performance (AUC=0.79, Extended Figures 6I) than the mean-response strategy (AUC=0.89).

As an illustrative example using most-resistant clone strategy (Figure 2B), in patient Kydar19, there are three clones: c1, c2 and c3. Here, c2 and c3, two low abundance clones, are predicted to be relatively responsive to the treatment, whereas c1, the most abundant clone, is predicted to be resistant. In this case, c1 will likely drive the patient response and thus the patient will be predicted to be resistant or with a low response to the treatment. The resulting predicted response scores from this strategy are significantly higher in responders vs. non-responders (Figure 2D), successfully predicting the treatment response (ROC-AUC of 0.83, Figure 2E). This may be the case as the most resistant clone is most likely to be selected upon treatment and end up dominating the tumor, thus best reflecting clinical response. From here onwards, we fixed this most-resistant clone response strategy for predicting clinical response and tested it in two additional cohorts. The top pathways enriched among the gene features used by the PERCEPTION model are surfactant metabolism and O-linked glycosylation of mucins.

Predicting CDK inhibition response in a breast cancer trial

Using the most-resistant clone response prediction approach described in the previous subsection, we next tested PERCEPTION’s ability to predict patient response in the FELINE breast cancer clinical trial⁴⁰. This clinical trial includes three treatment arms: endocrine therapy with letrozole (Arm A), an intermittent high-dose combination of letrozole and CDK inhibitor ribociclib (Arm B), and a continuous lower dose combination of the latter (Arm C). Sc-expression and treatment response labels were available for 33 patients (Arms A, B, C having 11 samples each; Table S7). Patient response was determined via tumor growth measurements from mammogram, MRI, and ultrasound of the breast.

We could build a (borderline) predictive PERCEPTION response model for only the CDK4/6 inhibitor ribociclib (with a Pearson R=0.26, P=1.5E-03), and thus we focused our analysis on the combination arms B and C that include it (Figure 3A). We processed the sc-expression profiles of the tumor cells as previously described⁴⁰ and identified 38 transcriptional clusters/clones that are shared across the patients (Extended Figure 7A–C, Methods). Patient response was predicted based on the pretreatment samples, following the exact same strategy employed in the multiple myeloma case. As the number of patients in each arm (B and C) is quite small we predicted the response of the patient pre-treatment samples in aggregate. The resulting predicted viability of the non-responders is higher than that of the responders (Wilcoxon rank-sum test, one-sided P=0.05, Figure 3B), as expected. PERCEPTION successfully stratified the responders vs. non-responders with a ROC-AUC of 0.776 (Figure 3C). Aligning to our known mechanism of action of ribociclib - inhibition of CDK4/6 activity, leading to cell cycle arrest, PERCEPTION’s signature comprising 72 genes is enriched in pathways involved in cell cycle, specifically, TNF receptor family (P=0.004) and regulation of p53 (P=0.004).

Figure 3: PERCEPTION prediction of the combination therapy in the FELINE clinical trial.

(A) Transcriptional clone composition (y-axis) in each breast cancer patient tumor studied in the combination arms B and C (x-axis), where the color code for the clones is provided at the top. In the x-axis, the labels are a combination of the patient id and the time point at which the sample was collected (“_S” - day 0 and “_E” - day 180). (B) The predicted combination response in 14 breast cancer patients (samples collected at day 0), stratified by their responder (N=7) vs. non-responder (N=7) status. Two-sided Wilcoxon rank sum tests were performed to compare groups. Box plot shows median (center), 25th and 75th percentiles (bounds of box), and minima and maxima (whiskers). (C) Receiver Operator Characteristic curve displaying the predicted combination response. The area under this curve, provided at the right bottom corner, denotes the overall stratification power in distinguishing responders vs. non-responders.

Capturing emergence of resistance in lung cancer patients

We next tested if PERCEPTION can capture the development of clinical resistance during targeted therapy treatment in patients. To this end, we analyzed a published cohort with scRNA-seq profiles of 24 non-small cell lung cancer (NSCLC) patients with 14 pre- and 25 post-treated biopsies (Extended Figure 8A–F, Table S8)⁴¹. In total, patients in this cohort were treated with four different tyrosine kinase inhibitors, including erlotinib (a first-generation EGFR inhibitor), dabrafenib (a serine/threonine kinase inhibitor), osimertinib (a third generation EGFR inhibitor), and trametinib (a MEK inhibitor). Based on the notion that the resistance to these targeted therapies frequently increases as the treatment time grows, we hypothesized that the predicted response for a given post-treatment biopsy would decrease (reflecting an increase in resistance to that treatment) as time elapses from the treatment start.

To study this hypothesis, for each post-treatment biopsy, we defined its estimated “Extent of Resistance” to a given treatment as the difference between its PERCEPTION predicted response vs. the baseline predicted response. The latter was computed as the mean predicted viability across all pre-treatment biopsies (as the majority of the samples were not matched, precluding an overall pairwise matched comparison). We found that the extent of resistance to treatment increases with the elapsed time since the start of treatment, but only in the patients reported to acquire resistance (progressive disease, Spearman Rho=0.634, P=0.026, Figure 4A, N=17). We also found that this positive correlation between the elapsed treatment time and the estimated extent of resistance holds true when patients receiving different drugs are analyzed separately (Extended Figure 9A), when controlling for prior treatments (Extended Figure 9B), when individual patients are analyzed separately (Extended Figure 9C) and when controlling for tumor stage (Extended Figure 9D). The extent of predicted resistance is significantly higher in post-treatment biopsies collected from the patients with progressive disease vs. residual disease (Wilcoxon rank sum P<0.002, stratification ROC-AUC=0.88, Figure 4B). Notably, we do not observe this strong positive correlation but rather a negative trend in patients that responded well to the treatment (residual disease, N=7, Spearman Rho= −0.67, P=0.11, Figure 4A). The observed increase in the predicted extent of resistance to treatment with elapsed treatment time occurred specifically in patients that acquire resistance.

Figure 4: Predicting the development of resistance to tyrosine kinase inhibitors (TKIs) in lung cancer patients.

(A) The extent of predicted reduced killing (as a corollary of resistance) to a treatment from the baseline (X-axis) is correlated with the time elapsed (days from start of treatment until biopsy) (Y-axis). The points and line colors denote whether the biopsy is from patients with progressive disease or from responders. The error bars in all panels show 95% confidence interval of the fit. (B) Receiver Operating curve depicting PERCEPTION predictive power in distinguishing progressive (N=17) vs responding (N=7) patients. (C) The case of patient TH179 with multiple biopsies is presented, where the predicted viability in 14 pre (day 0) and 4 post-resistant tumors at day 331 (N=1) and day 463 (N=3) to dabrafenib are shown. Error bars show the minimum and maximum values, due to the small sample size of three data points. (D) The rate of change in abundance of top vs bottom 50% predicted resistant clones (N=21 each) with elapsed time since the start of treatment. Box plot shows median (center), 25th and 75th percentiles (bounds of box), and minima and maxima (whiskers). (E) Correlation matrix of the extent of resistance among drugs available in the trial across all patients that have acquired resistance.. The strength of the correlation (Pearson R) is provided in the respective box, represented by the size of the circle, where the color represents whether the correlation coefficient is negative or positive (red and blue, respectively). This is computed this for drugs with at least three resistant patients (# of patients=4, 4, and 3, respectively). The drugs with correlations of P <0.1 (before FDR correction) are indicated by a “*”. In both C) and D), two-sided Wilcoxon rank sum tests were performed to compare groups. (F) Correlation matrix illustrating cross-resistance between various drugs. The matrix represents the results of our analysis to identify pairs of drugs A and B, where resistance to drug A may induce cross-resistance to drug B. The cross-resistance relationship can be asymmetric. Pearson’s r test p-value denotes correlation significance.

We next analyzed the subset of patients with matched biopsies, including five patients with two biopsies each and one patient with four biopsies. Analyzing these samples in a matched manner, we find that the correlation between treatment elapsed time and the estimated extent of resistance holds true in the matched cases, and only in the patients that have acquired resistance (regression interaction P=0.003). Of particular interest is a case of a single patient (TH179), treated with dabrafenib, who had four biopsies at two different time points and developed progressive disease. The predicted viabilities to dabrafenib of the four tumor biopsies taken after 331 and 463 days of start of treatment are significantly higher than pre-treatment biopsies (Figure 4C). Furthermore, the predicted viabilities of all three biopsies from day 463 are significantly higher than the biopsy from day 331. Notably, we find that the abundance of the top 50% predicted resistant clones increases while the abundance of the bottom 50% predicted resistant clones decreases with the elapsed time since the start of treatment, as one would expect (Figure 4D, Methods). The rate of increase of abundance is significantly higher in the top 50% of the predicted resistant clones vs. the bottom 50% (Figure 4D, Methods). Taken together, these results testify that PERCEPTION can capture and quantify the emergence of treatment resistance as the disease progresses.

We next found that the features/genes used by the above models are enriched in pathways involved in cell junction organization and cell-cell communication including, extracellular matrix organization, RHO GTPase cycle, and NOTCH signaling. We also found that this signature is also enriched in the recently reported resistance mechanism for EGFR-inhibitors (EGFRi) via hypermutators driven by AXL⁴². Specifically, our prediction signature is enriched in the three resistance pathways identified in that study⁴² study AXL overexpression signature (P=3E-03), MYC overexpression (P=2.1E-02) signature, and purine synthesis (P=1.6E-04).

To prioritize candidate drugs available in this cohort whose treatment may overcome the resistance acquired, we asked if the development of resistance to a drug can induce either cross-sensitivity or cross-resistance to the other drugs⁴³. We focused on the patients (Table S8) that acquired resistance and computed the PERCEPTION response predictions for each of these drugs and the correlations between these drug sensitivity predictions across these patients (Figure 4E, Methods). PERCEPTION predictions suggest that the development of resistance to erlotinib would induce a cross-sensitivity to gemcitabine (Figure 4F, Top-Left panel, Pearson’s R= −0.94, P=0.06) and cross-resistance to dabrafenib (Figure 4F, Top-Left panel, Pearson’s R=0.91, P=0.09). A literature survey (Methods) revealed that gemcitabine treatment can overcome erlotinib resistance in cancer cell lines through downregulation of Akt⁴⁴. In patients, a combination of gemcitabine + erlotinib in pancreatic cancer in phase III trial has shown a higher overall and progression-free survival^45,46. In contrast, the addition of trametinib to erlotinib did not significantly improve survival in a phase I/II clinical trial⁴⁷. In sum, our analysis supports the possibility that erlotinib resistance may induce cross-sensitivity to gemcitabine, which may be of interest for future testing.

Predicting combination therapies targeting disjoint clones

We next turned to investigate PERCEPTION’s capability to identify effective combination treatments in clinics. To this end, we curated clinical trial data of various combinations tested for NSCLC with survival information to assess the predictive power of PERCEPTION models. The trials data were curated from TrialTrove (Methods). We found that PERCEPTION’s predicted improvement in response to combinations vs. the pertaining monotherapies is correlated with the survival improvement due to the combination observed in the respective clinical trials (Extended Figure 10, A–C demonstrating for multiple myelona and E-K for lung cancer, Weighted Pearson Rho=0.66, P=0.02, weighted by the number of patients in a trial). The only targeted therapy with enough unique combination trials is erlotinib and repeating this analysis for erlotinib yielded concordant results (Pearson Rho=0.76, P=0.08, Extended Figure 10H). Aside from the trials tested, among all possible combinations tested of approved drugs, the top-ranking pathways composing combinations pairs are the tyrosine kinase pathway and the tubulin polymerization pathway (Extended Figure 10I–K). This analysis was also done for multiple myeloma and results are presented in Extended Figure 10A–D. Our top-ranked combination is niraparib & ponatinib, an EGFR inhibitor, and a canonical BCR-ABL inhibitor, respectively (DKS = 0.25, Empirical P value = 1E-04). The next top combination pair with the high DKS is lapatinib and thioguanine (DKS = 0.24, Empirical P value = 1E-04), a dual HER2 and EGFR inhibitor and a purine inhibitor, respectively. Analogously, we next looked for all possible triplets of drug combinations exhaustively (Extended Figure 10B, N=13,244). Our top hits include the combination of gefitinib, icotinib and trametinib (DKS = 0.21, Empirical P value = 1E-04) and gefitinib, lapatinib and trametinib (DKS = 0.19, empirical p-value = 1E-04) (Extended Figure 10D).

Benchmarking PERCEPTION vs. state-of-the-art methods

We compared the prediction performance of PERCEPTION in the above three clinical cohorts vs. two different published predictors and four other alternatives that we implemented (Figure 5A): 1) state-of-the-model based on sc-expression (Beyondcell²⁷), 2) state-of-the-art bulk-expression based models (ATLANTIS³³), 3) usage of pseudo-bulk-RNA-seq (Pseudo-Bulk), 4) taking the mean viability across all single cells in a tumor sample (Mean viability, the strategy we used for predicting response in cell lines and PDCs, mean-response-sc), 5) Bulk-based-only PERCEPTION models that are not tuned on sc-expression, and, finally, 6) three kinds of random models created using shuffled viability labels, random gene signatures, and random coefficients (Methods). Notably, across the three cohorts as well as in each individual cohort, PERCEPTION was the best-performing model by a considerable margin (mean AUC=0.828, Figure 5B) compared to the published state-of-the-art methods. The other models studied here achieved mean AUCs as follows: (a) State of the art models: Beyondcell=0.67, ATLANTIS=0.64, (b) three bulk expression-based models that we have generated: Pseudo-bulk=0.63, mean-viability=0.663, Bulk-based-only PERCEPTION models =0.63) and finally, three randomly generated models (as expected, shuffled viability labels=0.51, random gene signature=0.55, and random coefficient model=0.53). Notably, across the three clinical cohorts studied, the mean AUC improvement of PERCEPTION over the previous best-published model, Beyondcell is considerable (0.15, P=0.002).

Figure 5: Performance of PERCEPTION vs. state-of-the-art models RNA-seq models.

A) Illustration of our overall comparison comparing the PERCEPTION model (on the right) vs. eight other models, including two state-of-the-art SC models (Beyondcell and ATLANTIS), three bulk-RNA-seq based models (pseudo-bulk, mean response across all single cells and PERCEPTION trained on only bulk-model without sc-training) and finally three random models. B) The stratification performances of differentiating responders vs. non-responders are provided on the y-axis for each model vs. the PERCEPTION in three clinical cohorts: MM (multiple myeloma, 21 responders vs. 7 non-responder), BRCA (Breast cancer, N=7 each of responder and non-responder), and lung cancer (17 progressive vs. 7 responding). The dotted blue line denotes the mean performance of PERCEPTION across the three cohorts.

How to use PERCEPTION for new a cohort or a new drug

We provide predictive pan-cancer drug models for 44 FDA-approved drugs in our source data. For a new clinical trial dataset with sc-expression that involves drugs with existing predictive models, PERCEPTION can be run using a single script (script: Running_PERCEPTION_for_new_dataset in our GitHub repository).

When the drugs involved do not have given predictive models, one can still aim to build PERCEPTION models, as follows. First, this process requires the following two inputs: a) sc-expression of cancer cells from the tumor and b) treatment information. Second, the process involves three steps: Step 1: The user should first build a bulk-expression based model for the given treatment. One can readily aim to build models for any of the 1500 drugs currently available in DepMap. We recommend the user only consider employing models that surpass the predictive threshold we used (Pearson correlation >0.3 between observed and predicted). We also recommend training such models on all cell lines available (pan-cancer model) vs training on the subset of cell lines from the pertaining specific cancer type of the patient’s cohort, as we found that pan-cancer models perform better in both patients and cell lines (Extended Figure 1C, AUC=0.75 vs. 0.88 and AUC=0.52 vs. 0.77, in lung and breast cohort; decrease of Pearson correlation of 0.38 to 0.25 in cell lines). A similar approach and guidelines should be applied for building sc-based models. step 2: The user will next cluster the cancer cells available from the patient tumor and identify each cluster mean expression in the default setting and rank normalize it. step 3: Based on the sc-models, predict patient response using the three-step heuristic approach described in previous sections. The resulting response scores are predicted to stratify patients that are more likely to respond to the given treatment, where the higher the score, the higher the likelihood of response. The code for building and testing models for new drugs is provided in Running_PERCEPTION_for_new_dataset (mode 2).

Discussion

We present PERCEPTION, a first of its kind computational pipeline for systematically predicting patient response to cancer drugs at single-cell resolution. We demonstrate its application for predicting response to monotherapy and combination treatment at the level of cell lines, patient-derived primary cells, and in predicting patient response in three recent single-cell clinical cohorts, spanning multiple myeloma, breast cancer, and lung cancer. We find that incorporating the transcriptional clonal information of the tumor into the prediction process improves the overall accuracy. For a given patient, the transcriptional clone with the worst response, that is the most resistant pre-treatment clone, best explains the overall response to treatment. Performing an extensive and systematic comparison vs other expression-based models, we show that PERCEPTION achieves markedly superior performance compared to two previously published methods.

The observation that the most-resistant-clone strategy (the one used for predicting the response in clinical trials) can also stratify resistant vs. sensitive in cell-lines, albeit, with lower power than the mean-response strategy might be due to that the clinical responses are measured at much longer time scales in the patients (months) than in the cell lines (within days). Passage of more time is likely better for the selection of the most resistant clone. This underscores the importance of considering the repertoire of a given tumor’s transcriptional clones in predicting its response to therapy. Furthermore, the observation that pseudo-bulk and bulk-based models performed better vs. scRNA-seq based models in cell lines during cross-validation, might be due to the relative homogeneity of cell lines, where scRNA-seq may not offer advantages over bulk-seq, sometimes resulting in comparable or worse predictions.

Our study’s limitations include the use of homogenous 2D cell lines and sparse pre-treatment sc datasets with response labels to train our models. As data availability increases, so will our predictors’ accuracy and scope. Hence, the current demonstration of their potential value will hopefully serve to drive generation of more pre-treatment sc datasets with clinical annotations in the future. Given the $150K average yearly cost for cancer treatment in the US⁴⁸, the $15K for tumor sequencing seems justified, despite additional costs. This should be explored further, through more sc dataset collection and predictive model development. Another limitation of our study is that our model was learned over the in vitro dosages whose translation to clinical response is non-trivial and thus we chose the AUC measure, a response measure over multiple dosages (N=8), as it is more likely to lead to a more robust approach.

PERCEPTION’s predictions may be further improved by considering cancer type-specific cell lines, whenever a large number of such models become available for each cancer type. The quality of our response models depends on the quality of the sc-expression profiles available e.g., their depth, drop-out rates, etc. We deliberately did not impute the SC data given the recent reports that dropouts are limited to non-UMI-based sc-expression methods and otherwise likely reflect true biological variation^49,50. A key limitation of our pipeline is a lack of ability to predict drug effects on immune and normal cells in the tumor microenvironment, which is needed to estimate the toxicity and side effects of different combinations. A major push to future sc-based precision oncology development will come from large-scale drug screens of drugs in noncancerous cell lines, currently very scarcely available. Those will enable the construction of predictors of drug killing of non-tumor cells, using an analogous pipeline to the one presented here for tumor cells.

Finally, our results demonstrate that tracking the drug response expression in post-treatment biopsies could help follow the evolution of drug resistance at a single-cell resolution and help guide the design of future personalized combination treatments that could significantly diminish the likelihood of resistance emergence. Finally, going beyond patient stratification, we identify new combination therapies that different individual clonal clusters for multiple myeloma and lung cancer. However, we must note that these predictions require further validations.

In summary, this study is the first to demonstrate that the high resolution of information from scRNA-seq could indeed be harnessed to predict the treatment response of individual cancer patients in a systematic, data-driven manner. It is our hope that the results shown will herald many more such studies, sooner rather than later. Retrospective studies on additional clinical datasets need to be done to better assess the utility of single cell prediction approaches like PERCEPTION and its accuracy before it may be studied prospectively.

Methods

1. Data collection

We first collected the bulk-expression and drug response profiles generated in cancer cell lines curated in the DepMap³³ consortium from the Broad Institute (version 20Q1, https://depmap.org/portal/download/). The drug response is measured via area under the viability curve (AUC) across eight dosages and measures via a sequencing technique called PRISM³⁸. In total, we mined 488 cancer cell lines with both bulk-transcriptomics and drug response profiles. We next mined sc-expression of 205 cancer cell lines (280 cells per cell line) generated in a previous study³⁴ and distributed via the Broad Single-cell Portal. The metadata, identification, and clustering information were also mined from the same portal (https://singlecell.broadinstitute.org/single_cell/study/SCP542/pan-cancer-cell-line-heterogeneity#study-download). Data collection and analysis were not performed blind to the conditions of the experiments. Further information on research design is available in the Nature Research Reporting Summary linked to this article.

For the multiple myeloma data set and the breast cancer data set all human subjects data are coded data from two published papers^29,40. For the lung cancer data, we used only published data from (reference 41, Table S1). The published lung cancer data we used were obtained with informed consent from all study participants based on human subject protocols (CC13–6512 and CC17618, C. M. B. Principal Investigator) approved by an IRB at UC San Francisco and based on clinical trial NCT03433469. The details of the three clinical cohort including trial status and endpoint extraction process are provided in Table S9.

2. The PERCEPTION pipeline

A response model for a drug is built via the following two steps: 1. Learn from bulk and 2. optimize using sc-expression. In step 3, we use the models from step 2 to predict response in patients.

We first divided all the cancer cell lines into two sets - 1. Cell lines where bulk-expression is available, and sc-expression is not available (N=318) 2. Cell lines where sc-expression is available (N=170). The first set is used during learning from bulk (Step 1, expanded below) and the second in optimizing using sc-expression (Step 2).

Step 1: Learn from bulk:

As a feature selection step, we first identified genes whose bulk-expression is correlated with drug viability profile (using the Pearson correlation). We considered the Pearson correlation Pc(d, g) between drug d and gene g as a measure of information in a gene expression profile and ranked each gene based on the strength of the correlation. While considering the top X genes, where X is a hyperparameter optimized in the next step, we built a linear regression model regularized using elastic net to predict the response to d in five-fold cross-validation, as implemented in R’s glmnet⁵¹.

Step 2: Optimize using SC-expression:

We built the above model using a Bayesian-like grid search of various possible values for X (range 10–500), where the model with the best performance using sc-expression input of 169 cell lines (left one out for testing) was chosen. We finally measured the model performance in a leave one out cross-validation (CV) using the left-out cell line, which was not used in either model building or hyperparameter optimization. Here, the model is trained on all the data except for one sample, which is held out for testing – that is, its viability is predicted by the model. The CV process is then repeated n times, with a different sample being held out each time, on which the prediction is made. After running this n times, the Pearson correlation coefficient is calculated between the predicted and the observed drug response values for all n held-out samples. Performance was measured using Pearson’s correlation between the predicted response and the actual response.

Step 3:

In the third and final step of PERCEPTION, we predict clinical response in patients using a cell lines-based model and sc-expression profiles of the patient’s tumor. We identify the major cancer cell clusters using sc-expression, compute the mean expression for each clone, and use this as input for the model to predict drug response for each clone. The overall patient response is predicted as the minimum response among all clones, as we hypothesize that the most resistant clone will determine the clinical response. Our prediction strategy was determined through trial and error in a multiple myeloma patient cohort and was fixed and applied to all other patient cohorts in the study. For a given treatment, we interpret this as the predicted response of the most resistant clone in the patient tumor determines the clinical response. We converged on this strategy via a trial-and-error approach testing five different strategies to predict a patient’s response from its individual clone-level responses. This strategy is then fixed. During the comparison of PERCEPTION performance vs state-of-the-art methods, we employed the following three types of random models: (1) shuffling the viability labels in the cell lines, by (2) randomly selected gene signatures, and finally (3) using non-predictive models of other drugs.

Description of the method and optimization formula

We employed an iterative approach using elastic net regression to identify the optimal number of genes that maximize the predictive performance of our model. By performing elastic net regression with different subsets of genes, we were able to determine the optimal combination of L1 and L2 penalty hyperparameters and gene features that contribute to the best predictive performance. The objective function for each iteration remains the same as the elastic net regression:

$${min}_{\beta }\left({\frac{1}{2N}\left||Y-X\beta |\right|}_2^2+\lambda \left({\alpha \left||\beta |\right|}_1+\frac{1-\alpha }{2}{\left||\beta |\right|}_2^2\right)\right)$$

The process involves the following steps:

1. Select a subset of genes and form the design matrix $X$ with that subset.

2. Perform elastic net regression using the objective function above, optimizing the hyperparameters $\lambda$ and $\alpha$.

3. Evaluate the performance of the model using cross-validation.

4. Repeat steps 1–3 for different numbers of genes.

5. Choose the model with the number of genes that gives the maximum predictive performance.

The final chosen model would thus have the coefficients/weights of the selected genes as parameters and would be associated with the optimal hyperparameters $\lambda$ and $\alpha$, as well as the optimal number of genes.

Data choices in Step 1 and Step 2 of PERCEPTION

The first step of PERCEPTION (model building) used bulk-RNA-seq (bRNA-seq) of 318 cell lines to build an initial set of bulk-based models based on a large set of genes as features. The second step used scRNA-seq of 169 different cell lines to further select an optimal set of predictive features, resulting in a final set of drug-specific sc-based models. This approach was designed to make sure information would not leak between two steps leading to overfitted performance, by building the models on two entirely disjoint sets of cell lines. For some cell lines, both scRNA-seq and bRNAseq are available and in these cases, only their scRNA-seq was used during the second step.

3. Evaluating PERCEPTION on three independent cell-line screens

PERCEPTION’s performance on GDSC

The pharmacological drug screens performed by the PRISM and GDSC studies are based on two independent platforms. The GDSC data were downloaded from the DepMap portal on April 15, 2020 (https://depmap.org/portal/download/). By testing the performance of PERCEPTION on these independent screening platforms, we can measure the extent to which the expression signature captured by our drug response models can be translated across the platforms. The following steps were performed to achieve this:

Step 1: Quality Check to Select Cell Lines and Drugs:

Out of the 347 cell lines in common with drug response in both GDSC and PRISM, there are 120 cell lines with sc-expression data in a previous study³⁴. We considered only the drugs shared by GDSC and PRISM that have a concordant response (Pearson Rho > 0.3 & P < 0.05), resulting in 28 drugs. Among these 28 drugs shared between GDSC and PRISM, we were able to build predictive models for 16 drugs.

Step 2: Model Building and Parameter Optimization:

For each of the drugs selected above, we ran the PERCEPTION pipeline, optimizing parameters based on the sc-expression of 90 out of 170 cell lines, and the other 80 cell lines were test data.

Step 3: Prediction and Normalization:

For the drugs for which PERCEPTION could build models, we applied the models on the cell lines and obtained predictions for each individual cell. Monotherapy response for a given drug in a cell line was represented by the mean response of all the single cells (N=318). Since the range of PERCEPTION predicted values is typically smaller than those observed in the screens (Extended Figure 3G), we used scaled, predicted AUC scores (z-score) in further analyses.

Step 4: Testing and Performance Analysis:

The resulting response models were applied to the testing dataset, and the predicted AUC values were compared to the experimental responses from GDSC and PRISM. We computed two performance measure - AUC, stratifying top vs bottom 33% as resistance vs sensitive, and a correlation between predicted vs observed response. The former is provided in main text and the latter in Extended Figure 4. We note that the performance of the PRISM-based models in the GDSC test set is correlated with the concordance between the experimentally measured drug’s viability profiles in the two screens (Pearson R=0.49).

PERCEPTION’s performance on monotherapy and combinations

The process was carried out in distinct steps as described below:

Step 1: Quality Check to Select Cell Lines, Drugs and Data Points:

Nair’s dataset comprising monotherapy and combination response was mined from a recent study³⁹ where the response was measured via the AUC of the dosage-viability curve across eight dosages. Like our GDSC quality check, we only considered drugs with concordant response profile across Nair’s dataset and PRISM (Pearson Rho>0.3). AUC values >1 were removed as they are likely due to noise in the fitting the viability curve, due to noise and higher variability in doses that do not inhibit. This criterion yielded 14 FDA-approved drugs in 21 cell lines, and we focused on them.

Step 2: Building Models:

Standard PERCEPTION workflow was used to build a model for these 14 drugs.

Step 3: Combination Response Prediction:

We extended the prediction to the response to combinations of these 14 drugs studied in this screen (Table S5). A combination response in a cell line was predicted by adopting the IDA model across all the single cells from that cell line⁵²; i.e., the predicted combination response of N drugs is the effect of the single most effective drug in the combination. Performance was measured using ROC-AUC. Throughout our work, the combination response was predicted using the IDA principle.

Step 4: Performance Measurement:

Like above, we converted the continuous measures of viability to sensitive vs resistant labels. Using these labels, we computed the stratification AUC for monotherapy and combination response prediction.

PERCEPTION’s performance on head and neck cancer cell lines

The approach for this prediction was undertaken through specific stages:

Step 1. Data Collection and Initial Analysis:

We obtained the single-cell expression data for five head and neck squamous cell cancer (HNSC) patient-derived cell lines, along with their treatment response for eight drugs and combination therapy at two different dosages²⁶. An initial assessment revealed that PERCEPTION was unable to build drug response models with a Spearman correlation greater than 0.3 between predicted and experimental viability using PRISM screens. Therefore, we introduced changes to the PERCEPTION pipeline.

Step 2. Modification of the PERCEPTION Pipeline – Building Models from GDSC Screens:

We turned to GDSC screens to build models for drug response, utilizing data from more than approximately 800 cell lines specific to these drugs.

Step 3. Building Models:

We considered only the top 3000 highly expressed genes (with fewer dropouts in the HNSC dataset) in common between the bulk expression and PDC datasets to ensure a focused analysis on relevant genes. We then built a PERCEPTION model using these 3000 genes and GDSC response profiles. The monotherapy and combination responses were calculated following the same methodology used in the GDSC and Nair’s dataset cases.

Step 4. Performance Measurement:

Like the analyses above, we calculated both stratification AUC and correlations for assessing the performance.

4. Predicting combinations response in multiple myeloma patients

Response labels, sc-expression of patients’ tumors (MARS-seq), clustering annotation and mean cluster expression for the multiple myeloma data were mined from the original publication²⁹. No statistical methods were used to pre-determine sample sizes; we used all available samples. We only used the cells annotated as malignant. Predicting the combination response of a patient can be divided into a two-step process: Step 1. Predict the combination response of each clone in that tumor, Step 2. Predict the patient’s response from the clone-level combination response. To this end, we first tried to build PERCEPTION response models for the four treatments used in the combination therapy. We succeeded in building PERCEPTION response models for two of the four drugs in the trial that are predictive in cell lines (carfilzomib and lenalidomide, further details in Supplementary Note 1) and used them to predict the treatment response in patients. We first predicted the combination response for each transcriptional cluster (or simply referred to here as a “clone”). To this end, we predicted the response for each of the two drugs separately and computed the killing using the IDA principle; i.e., the predicted combination response of N drugs is simply the effect of the single most effective drug in the combination⁵². To overcome the challenge of the discrepancy of dosages used in the clinic vs. pre-clinical testing where our models are built, we z-scale our predicted response profile of a drug across clones, where this z-score predicted response represents the relative response of a clone compared to all the other available in the cohort.

In step 2, we use this clone-level combination killing profile in a patient to predict the overall patient’s response. We considered the predicted response of the least responsive clone found in each patient as that patient’s response. This is based on the notion that it would be selected by the treatment and eventually dominate the overall tumor. Performance was measured using ROC-AUC. For our model building control, we built random models using either shuffled labels, randomized features in the regression model, or a non-predictive model of another drug in the screen for 1000 times and computed the number of times that the stratification power denoted by AUC is higher than our original model. This proportion is provided as an empirical P-value.

Testing Prediction Strategies for Multiple Myeloma

Five different strategies were designed to translate clone-level killing to an overall clinical response prediction. These strategies were: a) Weighted Average Response: Calculating the average response across all the clones, with each clone’s response weighted by its abundance in the tumor, b) Unweighted Average Response: Taking a simple average of the response across all the clones, c) Most-Sensitive Clone Response: Using the response of the clone with the highest predicted response, d) Unweighted Most-Resistant Clone Response: Utilizing the response of the most-resistant clone (the clone with the least response), without weighting, e) Most-Resistant Clone Response: Choosing the response of the most-resistant clone but weighted by its abundance proportion. The performance of these strategies was tested in a cohort comprising responders (N=7) and non-responders (N=21) to the treatment. The accuracy of each strategy was measured using the Area Under the Receiver Operating Characteristic curve (AUC).

5. Predicting combinations response in breast cancer clinical trial analysis

The pre-filtered 10X-based single-cell RNAseq count data and the cell type annotations of the 65 breast cancer samples (34 patients) were downloaded from GEO (GSE158724). No statistical methods were used to pre-determine sample sizes; we used all available samples. These patients have samples collected at different time points during their treatment: 1) collected at the time of screening (S), 2) on day 14 (M) and 3) on day 180 at the end of the trial (E). In our analysis, we considered only the cells annotated as tumor cells. As defined in the primary publication of the dataset⁴⁰, we applied Seurat (v.4.0.5)⁵³. We filtered out samples with fewer than 100 cells. The 38 transcription clusters identified in all 65 samples post-filtering and data processing ares presented in Extended Figure 7A–B. We used the reciprocal principal-component analysis integration workflow to integrate the tumor cells from the remaining samples⁵³. The data were normalized using the SCTransform function and the top 5000 variably expressed genes and the first 50 principal components (PCs) were used in the anchor-based integration step. The first fifty PCs and a k.param value of 20 were used to identify neighbors and the resolution was set to 0.8 to find distinct clusters. We identified 36 different clones, of which only 16 clones were found in the pretreated samples from patients in Arms B and C. The sc-expression of 16 clones was considered in the drug response prediction analysis. The patient response information was obtained from Table S12 in the original publication⁴⁰.

We used patients with paired samples at time points S and E to study the change in response post-treatment. Extended Figure 7B, shows the clonal distribution in each sample processed, all sub-clones which represent <5% of the cells in the sample are excluded in our analysis. The default PERCEPTION pipeline was used to build drug response models except for a single change. The top ~2500 highly expressed genes (ranked by the total number of non-zeroes across all the cancer cells) in the breast cancer dataset that are in common with the cancer cell line bulk expression data were used in the pipeline. The resulting models were used to predict response at the patient level in a similar manner to what we did for the multiple myeloma data. The controls for the model building were also tested for the breast cancer data, like the testing we did for the multiple myeloma data. We note that the number of clusters identified using the standard Seurat pipeline slightly changed when the initial seed for random number generation was changed. These changes did influence the performance by up to 1 in the first significant digit (AUC varied from 0.70 to 0.83 when the seed was changed), but the overall inferences were consistent.

6. Response models to distinguish responders vs. non-responders

We built bulk-based drug response models to compare their performance vs. PERCEPTION models in stratifying responders from non-responders in the two clinical trials. To build drug response models based on bulk expression data, we considered all ~500 cell lines with bulk expression and PRISM-based drug response. For each drug, we randomly divided the data into training (1/3^rd of the cell lines) and test set (2/3^rd of the cell lines). As a feature selection step, we first identified genes whose bulk expression is correlated with the drug viability profile (Pearson R) in the training set. We considered the correlation for each gene as a measure of information in a gene expression profile and ranked each gene based on the strength of the correlation. While considering the top 100 genes, we built a linear regression model regularized using an elastic net to predict the response to in leave one out cross-validation, as implemented in R’s glmnet⁵¹. The resulting model performance was validated on the testing dataset.

To build state-of-the-art bulk-based drug response models as defined in a previous study³³, we generated random-forest-based models in a similar framework as defined above. To make sure that the gene features used in the resulting model predictors are detected to be expressed in the patient sc-dataset, we consider genes that overlap in both the cell line bulk expression data and patient sc-dataset to build the models. For each drug, we repeated the above model-building steps 100 times and presented the mean and standard error of their performances in stratifying responders from non-responders in their respective clinical trials.

7. Predicting resistance to tyrosine kinase inhibitors in NSCLC

The sc-expression profiles of 39 biopsies from 25 NSCLC patients were provided by the original study authors⁴¹. The clinical annotations used for this analysis were mined from the original publication⁴¹, specifically Table S1. Like in previous sections, we focused only on the subset of single cells labeled in the publication as malignant. Seurat clustering was performed with the resolution = 0.8, dims = 10, number of features = 2000, scale.factor = 10000, log normalization method with minimum cells in a cluster required to be > 3 and minimum features required to be > 200, to identify a total of 16 clones (Extended Figure 8A). The expression of each transcriptional cluster/clone in a patient is the averaged expression across all the single cells associated with that cluster in that given patient and a rank normalization is performed. We successfully built drug response models for dabrafenib, erlotinib, gemcitabine, osimertinib, and trametinib. The response observed in the most resistant clone of a patient is considered as the PERCEPTION’s predicted response. We primarily studied the development of drug resistance in the trial. To this end, we defined a term called “Extent of Resistance” of a drug, which is the difference between a drug’s predicted viability from PERCEPTION and the predicted baseline viability. The predicted baseline viability is defined as the average predicted viability of the respective treatments in all treatment-naive samples. This difference in response from the naive state denotes the extent of resistance and is thus named accordingly. We computed both Spearman and Pearson correlations to identify robust correlations.

8. Literature survey of cross-resistance and cross-sensitivity

To search for evidence available in published papers for a cross-resistant or cross-sensitive drug pair, we used the search term “drug X AND drug Y” e.g., erlotinib AND gemcitabine, in the PubMed search portal https://pubmed.ncbi.nlm.nih.gov/ on December 26, 2021. The resulting clinical trials in the first fifty matches, sorted by best match, were manually curated for outcomes. For pre-clinical evidence for or against, non-clinical studies testing the combinations were manually curated.

9. Change of abundance vs predicted resistance of a clone

We first computed and ranked all clones with at least two data points at different time points by their mean predicted resistance across all samples they are present in. For each clone, we next computed the rate of change of abundance (slope) of the best-fit line of abundance vs biopsy time from the start of treatment. Finally, we compared this “rate of change of abundance” vs “mean predicted resistance” of each clone.

10. Comparing PERCEPTION’s performance on three clinical cohorts

We identified relevant, competing state-of-the-art single-cell methods for benchmarking PERCEPTION by searching PubMed using keyword “single-cell expression prediction”. This search yielded only Beyondcell, a state-of-the-model based on single-cell expression (Fustero-Torre, et al. 2021). Additionally, we tested a state-of-the-art bulk (ATLANTIS) and four alternative methods. The implementation of Beyondcell was downloaded from https://github.com/cnio-bu/beyondcell and the default signatures provided by Beyondcell’s authors were used. The random forest-based ATLANTIS was downloaded from https://github.com/cancerdatasci/atlantis/releases and used in the default setting. Benchmarking PERCEPTION against these tools to predict patient response, we calculated the area under the receiver operating characteristic curve (AUC) for each model in the three clinical cohorts (multiple myeloma, breast cancer and lung cancer). We then calculated the mean AUC across the three cohorts for each model to determine the overall performance.

11. Testing the most-resistant clone strategy in cell lines

We tested the performance in cell lines for the most resistant clone strategy to stratify resistant vs sensitive cell lines. To this end, we first clustered the 200 cell lines via Seurat using uniform parameters used across the study noting 29 clusters and four clusters per cell line (Extended Figure 9F). We repeated this process for the head and neck patient-derived cell lines. The transcriptional cluster/clonal information was obtained from original publication. We analyzed sc-expression of primary cells derived from five different patients treated with eight different drugs at two concentrations (Table S6), including both monotherapy and combination therapies⁹. We could build PERCEPTION response models for 4 out of the 8 drugs tested (docetaxel, epothilone-b, gefitinib, and vorinostat; Pearson R > 0.25). Resistant vs. sensitive cell were the top vs. bottom 40% cell lines ranked by viability. Our predictions were performed for 2 dosages x 4 monotherapies x 5 cell lines. The predicted viability over the 20 (monotherapy, cell line) pairs, comprising four monotherapies x 5 cell lines, is correlated with the observed viability, and individual drug-level correlations are provided in Extended Figure 5. We plotted the predicted vs. experimental correlations obtained for all data points and drug levels are provided in Extended Figures 5 and 6.

12. Drug combinations targeting multiple myeloma clones

To predict combinations for multiple myeloma patients that target multiple clones in the tumor disjointly and thus, have a low likelihood for resistance emergence, we began with all combinations of two drugs with predictive PERCEPTION model (N=44 drugs) and ranked every pair by a score denoting the extent of their disjoint killing, termed its Disjoint Killing Score (DKS). This score quantifies the increase in predicted killing compared to the expected (better killing of the two monotherapies) of a drug combination. Out of the 946 possible combinations scanned, 842 pairs show no improvement over the expected (Disjoint Killing Score=0). Analogously, we next looked for all possible triplets of drug combinations exhaustively (N=13,244). Once validated, this design can be utilized for designing optimal combinations targeting multiple clones in a patient. Applying this approach in lung cancer, we ranked every pair (N=946) by the DKS computed across the four different sc lung cancer cohorts. Out of the 946 possible combinations evaluated, 915 pairs show no improvement over the monotherapy treatment (DKS=0). The remaining combinations, with Disjoint Killing Score > 0, are shown in Extended Figure 10A, C. We also computed therapy types that are more likely to have high DKS.

13. Using TrialTrove to test PERCEPTION on predicting response

We reasoned that it would be possible to curate clinical trial data to assess the predictive power of PERCEPTION models in a clinical setting. We used a licensed database TrialTrove which has more trials and more detailed and structured information ClinicalTrials.gov, better facilitating our data extraction⁵⁴. We assembled a collection of clinical trials data of combination therapy, using software we have built to parse the TrialTrove database. No statistical methods were used to pre-determine sample sizes.

To test our model, our general approach was to identify combination, multi-arm trials in which one patient arm was administered two drugs, A+B, and another patient arm was administered only drug A. Specifically, we mined trials meeting two criteria. 1. Uniform and consistent trial efficacy labels for either median progression-free survival or median-overall survival 2. Having at least two arms with treatment design of drug A vs drug A + drug B, and 3. Targeted therapy treatments (drugs A and B) with predictive PERCEPTION models. Among several cancer types we investigated, NSCLC was the cancer type for which we could find sufficient and the most abundant homogeneous data for N > 10 trials, partly because the median survival data are more readily available for cancers such as NSCLC with poor survival, hence, we focused on NSCLC in this subsection. An additional filter we applied is that we must be able to build a PERCEPTION model for both drugs A and B. We next predicted the improvement in response to such combinations over whichever monotherapy was tested in the trial, computed as the response difference between the combination and monotherapy (survival improvement due to combination), in patients from four NSCLC cohorts with sc-expression^55–57, serving as representative samples of sc tumor data of NSCLC patients (Total patients=18).

We started with a repository of 66,116 oncology trials in phases beyond Phase I. To identify combination trials, we used a python implementation of a modified form of a query suggested by a TrialTrove curator. One of the three fields {Trial Title, Trial Objective, Treatment Plan} should contain any one of the seven strings: “combination”, “both drugs”, “with or without”, “combined”, “plus”, “with and without”, “alone or with”, “concurrent”. In addition, the Trial Keywords must not contain the string “single-arm”. To narrow down our list of applicable trials to only those with results of interest, we required (via another python program) that the Trial Results field must contain any one of 93 strings such as: ORR, OS, PFS, response rate, overall survival, progression-free survival, disease control rate, etc. To identify trials that used drugs that PERCEPTION can model, we processed the fields named Primary Tested Drug and Other Tested Drug to require that together these two fields must contain at least two drugs that can be modeled by PERCEPTION and at least one drug that is a targeted therapy rather than chemotherapy. The Primary Tested Drug and Other Tested Drug fields are already normalized for synonyms. The overall goal of the three filters (combination trials, results available, PERCEPTION-suitable drugs used) was to eliminate false negatives, trials that would not be useful in testing PERCEPTION-built models. Trials that survived these three filters were then curated manually to obtain accurate arm information and results.

We measured our performance by computing a correlation between PERCEPTION’s predicted improvement to combination vs. survival improvement due to the combination observed in the respective clinical trials. Separate analyses for overall and progression-free survival were also done. However, we note the small cohorts available for these two analyses. Repeating this analysis in a drug-specific manner, we focused on trials of different drug combinations (N=6) with erlotinib, the only targeted therapy with enough unique combination trials.

Extended Data

Extended Data Fig. 1:

Overview of PERCEPTION model’s training data and features.

Extended Data Fig. 2:

Visualization of PERCEPTION’s ability to predict viability at four recent EGFR inhibitors vs the EGFR pathway activity at single-cell resolution.

Extended Data Fig. 3:

Evaluating PERCEPTION’s Efficacy in Unseen Lung Cancer Cell Line Screens.

Extended Data Fig. 4:

Quality Control and Predictive Analyses in Lung Cancer Cell Line Screens.

Extended Data Fig. 5:

The predicted vs. experimental correlations obtained for individual treatments.

Extended Data Fig. 6:

Correlation of Predicted and Observed Viability in Monotherapies and Combination Treatments in Cell Lines.

Extended Data Fig. 7:

Comparing PERCEPTION with Existing Bulk Response Models in a Breast Cancer Clinical Trial.

Extended Data Fig. 8:

Pre-processing and predicting clone level response in lung cancer patient cohort.

Extended Data Fig. 9:

Correlation between the elapsed treatment time and estimated resistance holds true across different conditions.

Extended Data Fig. 10:

Identifying Optimal Drug Combinations for Multiple Myeloma and Lung Cancer Patients.

Supplementary Material

Supplementary Tables 1-9

Table S1. List of cell lines used in PERCEPTION pipeline and data available for them.

Table S2. List of FDA approved cancer drugs used in the study.

Table S3: Drugs with AUC response in both PRISM and GDSC screens.

Table S4. Correlation of drug response between PRISM, GDSC and PERCEPTION in the 80 test cell lines.

Table S5: PERCEPTION’s response using SC in lung cancer cell lines.

Table S6. Mono- and combination therapies tested on HNSC patient derived cell lines.

Table S7. Clinical information of patients from breast cancer clinical trial.

Table S8. Predicted extent of resistance and demographics of patients from lung cancer cohort.

Data availability

The entire collection of the processed datasets used in this manuscript, including pre-clinical models of cancer cell lines and PDCs, can be accessed via the following Zenodo repository (https://zenodo.org/record/7860559). We collected the bulk-expression and drug response profiles generated in cancer cell lines curated from https://depmap.org/portal/download/ (version 20Q1). The sc-expression of 205 cancer cell lines was generated in a previous study³⁴ and was downloaded from: https://singlecell.broadinstitute.org/single_cell/study/SCP542/pan-cancer-cell-line-heterogeneity#study-download. The sc-expression profiles of multiple myeloma patients were downloaded from the original study Supplementary Table 2 (https://static-content.springer.com/esm/art%3A10.1038%2Fs41591-021-01232-w/MediaObjects/41591_2021_1232_MOESM3_ESM.xlsx), of breast cancer were downloaded from GEO (GSE158724), and, of NSCLC patients were provided by the original study authors41.

Code availability

The study’s scripts to replicate each step of results and plots can be accessed via the following GitHub repository (https://github.com/ruppinlab/SCPO_submission). We used open-source R versions 4.0 through 4.2 to generate the figures. Wherever required, commercially available Adobe Illustrator was used to create the figure grids.