Abstract
Tailoring optimal treatment for individual cancer patients remains a significant challenge. To address this issue, 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 patients with lung cancer treated with tyrosine kinase inhibitors. PERCEPTION outperforms published state-of-the-art sc-based and bulk-based predictors in all clinical cohorts. PERCEPTION is accessible at https://github.com/ruppinlab/PERCEPTION. Our work, showcasing patient stratification using sc-expression profiles of their tumors, will encourage the adoption of sc-omics profiling in clinical settings, enhancing precision oncology tools based on sc-omics.
This is a preview of subscription content, access via your institution
Access options
Access Nature and 54 other Nature Portfolio journals
Get Nature+, our best-value online-access subscription
$29.99 / 30 days
cancel any time
Subscribe to this journal
Receive 12 digital issues and online access to articles
$119.00 per year
only $9.92 per issue
Buy this article
- Purchase on Springer Link
- Instant access to full article PDF
Prices may be subject to local taxes which are calculated during checkout
Similar content being viewed by others
Data availability
The entire collection of the processed datasets used in this manuscript, including preclinical models of cancer cell lines and PDCs, can be accessed in the Zenodo repository (https://zenodo.org/record/7860559)58. We collected the bulk-expression and drug response profiles generated in cancer cell lines curated from the DepMap portal (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 patients with multiple myeloma were downloaded from the original study (their supplementary Table 2; https://static-content.springer.com/esm/art%3A10.1038%2Fs41591-021-01232-w/MediaObjects/41591_2021_1232_MOESM3_ESM.xlsx); data from patients with breast cancer were downloaded from GEO (GSE158724) and data from patients with NSCLC were provided by the original study authors41.
Code availability
The scripts to replicate each step of results and plots can be accessed in a 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.
References
Tsimberidou, A. M., Fountzilas, E., Nikanjam, M. & Kurzrock, R. Review of precision cancer medicine: evolution of the treatment paradigm. Cancer Treat. Rev. 86, 102019 (2020).
Huang, K., Xiao, C., Glass, L. M. & Critchlow, C. M. Machine learning applications for therapeutic tasks with genomics data. Patterns 2, 100328 (2021).
Bhinder, B., Gilvary, C., Madhukar, N. S. & Elemento, O. Artificial intelligence in cancer research and precision medicine. Cancer Discov. 11, 900–915 (2021).
Singla, N. & Singla, S. Harnessing big data with machine learning in precision oncology. Kidney Cancer J. 18, 83–84 (2020).
Senft, D., Leiserson, M. D. M., Ruppin, E. & Ronai, Z. Precision oncology: the road ahead. Trends Mol. Med. 23, 874–898 (2017).
Tsimberidou, A. M., Fountzilas, E., Bleris, L. & Kurzrock, R. Transcriptomics and solid tumors: the next frontier in precision cancer medicine. Semin. Cancer Biol. 84, 50–59 (2022).
Siravegna, G., Marsoni, S., Siena, S. & Bardelli, A. Integrating liquid biopsies into the management of cancer. Nat. Rev. Clin. Oncol. 14, 531–548 (2017).
Heitzer, E., Haque, I. S., Roberts, C. E. S. & Speicher, M. R. Current and future perspectives of liquid biopsies in genomics-driven oncology. Nat. Rev. Genet. 20, 71–88 (2019).
Sawabata, N. Circulating tumor cells: from the laboratory to the cancer clinic. Cancers 12, 3065 (2020).
Beaubier, N. et al. Integrated genomic profiling expands clinical options for patients with cancer. Nat. Biotechnol. 37, 1351–1360 (2019).
Hayashi, A. et al. A unifying paradigm for transcriptional heterogeneity and squamous features in pancreatic ductal adenocarcinoma. Nat. Cancer 1, 59–74 (2020).
Rodon, J. et al. Genomic and transcriptomic profiling expands precision cancer medicine: the WINTHER trial. Nat. Med. 25, 751–758 (2019).
Tanioka, M. et al. Integrated analysis of RNA and DNA from the phase III trial CALGB 40601 identifies predictors of response to trastuzumab-based neoadjuvant chemotherapy in HER2-positive breast cancer. Clin. Cancer Res. 24, 5292–5304 (2018).
Vaske, O. M. et al. Comparative tumor RNA sequencing analysis for difficult-to-treat pediatric and young adult patients with cancer. JAMA Netw. Open 2, e1913968 (2019).
Wong, M. et al. Whole genome, transcriptome and methylome profiling enhances actionable target discovery in high-risk pediatric cancer. Nat. Med. 26, 1742–1753 (2020).
Lee, J. S. et al. Synthetic lethality-mediated precision oncology via the tumor transcriptome. Cell 184, 2487–2502 (2021).
Dinstag, G. et al. Clinically oriented prediction of patient response to targeted and immunotherapies from the tumor transcriptome. Med 4, 15–30.e8 (2023).
Castro, L. N. G., Tirosh, I. & Suvà, M. L. Decoding cancer biology one cell at a time. Cancer Discov. 11, 960–970 (2021).
Wensink, G. E. et al. Patient-derived organoids as a predictive biomarker for treatment response in cancer patients. npj Precis. Oncol. 5, 30 (2021).
Yao, Y. et al. Patient-derived organoids predict chemoradiation responses of locally advanced rectal cancer. Cell Stem Cell 26, 17–26 (2020).
de Witte, C. J. et al. Patient-derived ovarian cancer organoids mimic clinical response and exhibit heterogeneous inter-and intrapatient drug responses. Cell Rep. 31, 107762 (2020).
Shalek, A. K. & Benson, M. Single-cell analyses to tailor treatments. Sci. Transl. Med. 9, eaan4730 (2017).
Adam, G. et al. Machine learning approaches to drug response prediction: challenges and recent progress. npj Precis. Oncol. 4, 19 (2020).
Zhu, S. et al. Advances in single-cell RNA sequencing and its applications in cancer research. Oncotarget 8, 53763–53779 (2017).
Kim, K. T. et al. Application of single-cell RNA sequencing in optimizing a combinatorial therapeutic strategy in metastatic renal cell carcinoma. Genome Biol. 17, 80 (2016).
Suphavilai, C. et al. Predicting heterogeneity in clone-specific therapeutic vulnerabilities using single-cell transcriptomic signatures. Genome Med. 13, 189 (2021).
Fustero-Torre, C. et al. Beyondcell: targeting cancer therapeutic heterogeneity in single-cell RNA-seq data. Genome Med. 13, 187 (2021).
Ianevski, A. et al. Patient-tailored design for selective co-inhibition of leukemic cell subpopulations. Sci. Adv. 7, eab4038 (2021).
Cohen, Y. C. et al. Identification of resistance pathways and therapeutic targets in relapsed multiple myeloma patients through single-cell sequencing. Nat. Med. 27, 491–503 (2021).
Ledergor, G. et al. Single cell dissection of plasma cell heterogeneity in symptomatic and asymptomatic myeloma. Nat. Med. 24, 1867–1876 (2018).
Sade-Feldman, M. et al. Defining T cell states associated with response to checkpoint immunotherapy in melanoma. Cell 175, 998–1013 (2018).
Ghandi, M. et al. Next-generation characterization of the Cancer Cell Line Encyclopedia. Nature 569, 503–508 (2019).
Tsherniak, A. et al. Defining a cancer dependency map. Cell 170, 564–576 (2017).
Kinker, G. S. et al. Pan-cancer single-cell RNA-seq identifies recurring programs of cellular heterogeneity. Nat. Genet. 52, 1208–1218 (2020).
Plana, D., Palmer, A. C. & Sorger, P. K. Independent drug action in combination therapy: implications for precision oncology. Cancer Discov. 12, 606–624 (2022).
Yang, W. et al. Genomics of drug sensitivity in cancer (GDSC): a resource for therapeutic biomarker discovery in cancer cells. Nucl. Acids Res. 41, D955–D961 (2012).
Seashore-Ludlow, B. et al. Harnessing connectivity in a large-scale small-molecule sensitivity dataset. Cancer Discov. 5, 1210–1223 (2015).
Corsello, S. M. et al. Discovering the anticancer potential of non-oncology drugs by systematic viability profiling. Nat. Cancer 1, 235–248 (2020).
Nair, N. U. et al. A landscape of response to drug combinations in non-small-cell lung cancer. Nat. Commun. 14, 3830 (2023).
Griffiths, J. I. et al. Serial single-cell genomics reveals convergent subclonal evolution of resistance as patients with early-stage breast cancer progress on endocrine plus CDK4/6 therapy. Nat. Cancer 2, 658–671 (2021).
Maynard, A. et al. Therapy-induced evolution of human lung cancer revealed by single-cell RNA sequencing. Cell 182, 1232–1251 (2020).
Noronha, A. et al. AXL and error-prone DNA replication confer drug resistance and offer strategies to treat EGFR-mutant lung cancer. Cancer Discov. 12, 2666–2683 (2022).
Pluchino, K. M., Hall, M. D., Goldsborough, A. S., Callaghan, R. & Gottesman, M. M. Collateral sensitivity as a strategy against cancer multidrug resistance. Drug Resist. Updat. 15, 98–105 (2012).
Bartholomeusz, C. et al. Gemcitabine overcomes erlotinib resistance in EGFR-overexpressing cancer cells through downregulation of Akt. J. Cancer 2, 435–442 (2011).
Moore, M. J. et al. Erlotinib plus gemcitabine compared with gemcitabine alone in patients with advanced pancreatic cancer: a phase III trial of the National Cancer Institute of Canada Clinical Trials Group. J. Clin. Oncol. 25, 1960–1966 (2007).
Shin, S., Park, C. M., Kwon, H. & Lee, K.-H. Erlotinib plus gemcitabine versus gemcitabine for pancreatic cancer: real-world analysis of Korean national database. BMC Cancer 16, 443 (2016).
Luo, J. et al. Erlotinib and trametinib in patients with EGFR-mutant lung adenocarcinoma and acquired resistance to a prior tyrosine kinase inhibitor. JCO Precis. Oncol. 5, 55–64 (2021).
Mariotto, A. B. et al. Projections of the cost of cancer care in the United States: 2010–2020. J. Natl. Cancer Inst. 103, 117–128 (2011).
Svensson, V. Droplet scRNA-seq is not zero-inflated. Nat. Biotech. 38, 147–150 (2020).
Cao, Y., Kitanovski, S., Küppers, R. & Hoffmann, D. UMI or not UMI, that is the question for scRNA-seq zero-inflation. Nat. Biotechnol. 39, 158–159 (2021).
Friedman, J., Hastie, T. & Tibshirani, R. Regularization paths for generalized linear models via coordinate descent. J. Stat. Soft. 33, 1–22 (2010).
Ling, A. & Huang, R. S. Computationally predicting clinical drug combination efficacy with cancer cell line screens and independent drug action. Nat. Commun. 11, 5848 (2020).
Hao, Y. et al. Integrated analysis of multimodal single-cell data. Cell 184, 3573–3587 (2021).
Stergiopoulos, S., Getz, K. A. & Blazynski, C. Evaluating the completeness of ClinicalTrials.gov. Ther. Innov. Regul. Sci. 53, 307–317 (2019).
Lambrechts, D. et al. Phenotype molding of stromal cells in the lung tumor microenvironment. Nat. Med. 24, 1277–1289 (2018).
Song, Q. et al. Dissecting intratumoral myeloid cell plasticity by single cell RNA‐seq. Cancer Med. 8, 3072–3085 (2019).
Zilionis, R. et al. Single-cell transcriptomics of human and mouse lung cancers reveals conserved myeloid populations across individuals and species. Immunity 50, 1317–1334 (2019).
Sinha, S. Predicting patient treatment response and resistance via single-cell transcriptomics of their tumors (0.1) [Data set]. Zenodo https://doi.org/10.5281/zenodo.7860559 (2022).
Acknowledgements
This research was supported in part by the Intramural Research Program of the National Institutes of Health (NIH), National Cancer Institute (NCI), NIH grants R01CA231300 (T.G.B.), R01CA204302 (T.G.B.), R01CA211052 (T.G.B.), R01CA169338 (T.G.B.) and U54CA224081 (T.G.B.). This work used the computational resources of the NIH High-Performance Computing Biowulf cluster (http://hpc.nih.gov). We acknowledge and thank the NCI for providing financial and infrastructural support. Thanks to K. Wang, S. Rajagopal and Z. Ronai for their valuable feedback and discussion. Special thanks to J. I. Griffiths and A. H. Bild for clarifying the patient response data in reference 40 and for their helpful feedback.
Author information
Authors and Affiliations
Contributions
S.S., R.V., A.A.S. and E.R. conceived the framework of the analysis. E.R. and A.A.S. mentored and guided the study. S.S. and R.V. led the analysis of the development of the models and most of the testing. A.A.S., A.V.K., R.V. and S.S. performed the analysis related to clinical trials curation and data analysis. A.A.S., S.M., S.R.D, N.U.N, M.G.J. and N.Y worked on the revisions for model validation and further testing and development of the software. W.W., D.L.K, C.M.B. and T.G.B. provided the lung cancer data and aided in its analysis. O.V.S., I.G., K.D.A., C.M.B. and C.J.T. contributed to finding relevant dosages to translate in vitro to in vivo results. S.S., R.V., A.A.S., E.R., P.J., C.H.B. and T.G.B. wrote the initial draft of the manuscript; S.S., S.M., A.A.S. and E.R. carried out the revisions.
Corresponding authors
Ethics declarations
Competing interests
S.S., R.V., A.A.S. and E.R. are inventors on a provisional patent application covering the methods in PERCEPTION. E.R. is a co-founder of Medaware, Metabomed and Pangea Biomed (divested from the latter). E.R. serves as a non-paid scientific consultant to Pangea Biomed, a company developing a precision oncology SL-based multi-omics approach, with emphasis on bulk tumor transcriptomics. T.G.B. is an advisor to Array/Pfizer, Revolution Medicines, Springworks, Jazz Pharmaceuticals, Relay Therapeutics, Rain Therapeutics and Engine Biosciences, and receives research funding from Novartis, Strategia, Kinnate and Revolution Medicines. The work in the laboratory of C.H.B. was funded in part by Amgen and Novartis. The other authors declare no competing interests.
Peer review
Peer review information
Nature Cancer thanks Federica Eduati and Tuomas Tammela for their contribution to the peer review of this work.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Extended data
Extended Data Fig. 1 Overview of PERCEPTION model’s training data and features.
A) Cancer type distribution of the 318 cell lines used during the bulk expression training of PERCEPTION (step 1). B) Similarly, showing the cancer type distribution of the 169 cell lines used during the sc-expression training of PERCEPTION (step 2) C) The performance of PERCEPTION in predicting response in unseen cell lines when built via (1) pan-cancer models: all available cell lines (N = 169) are used for training the model, (2) Cancer-type specific: trained only on cell lines of the same cancer type as those used in the testing (N = 16 melanoma cell lines, 37 lung cancer cell lines and 15 breast cancer cell lines, as we used the PERCEPTION to predict the patient’s treatment response in three clinical trial cohorts from skin, lung, and breast cancer, we compared the pan-cancer model with these three individual cancer-type models). No statistical test was performed to compare groups. Error bars indicate the standard error of the mean (SEM), reflecting data variability. D) Major classes of mechanism of action of the 133 FDA-approved drugs that were studied here. No statistical test was performed to compare between groups. E) Top pathways enriched in frequently appearing features/genes in the PERCEPTION models. This is computed using a GSEA rank test across all hallmark pathways. To assess the statistical significance of these scores, a permutation test was performed.
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.
A) The top-most panel visualizes the PERCEPTION predicted killing by nutlin-3, a canonical MDM2 antagonist and the expression of MDM2 for every single cell (each point) in the top and bottom tSNE plot, respectively. The intensity of the color denotes the extent of predicted killing in the right panel and measured MDM2 expression in the left panel. 3566 single-cells from nine p53 WT lung cancer cell lines are depicted. The tSNE clustering is performed using the expression of all the genes. B) A similar display visualizes PERCEPTION’s predicted killing and the EGFR pathway signature expression across 12,482 individual lung cancer cells. C) The four panels visualize predicted killing by four EGFR inhibitors, afatinib, icotinib, lapatinib, osimertinib, in every single cell (each point) via a tSNE plot, respectively. Here, the color of each point denotes the extent of predicted killing. In this figure, we provide data on 12,482 individual lung cancer cells. The tSNE clustering is performed using the expression profiles of all the genes. D) We present here the correlation between the predicted killing effect of nutlin-3 from the PERCEPTION prediction of each cell (x-axis) and the MDM2 gene expression in that single cell, where they are found to be strongly correlated. “MDM2 Activity” on the y-axis denotes MDM2 gene expression.
Extended Data Fig. 3 Evaluating PERCEPTION’s Efficacy in Unseen Lung Cancer Cell Line Screens.
A) A) Correlation Analysis: Examines the relationships across three platforms - “GDSC vs. PRISM”, “PRISM vs. PERCEPTION” (cross-validation), and “GDSC vs. PERCEPTION”. Drug response predictions at single-cell resolution were aggregated to represent overall cell line responses. B) These cross-platform correlations are provided at a drug level. Significance of correlations assessed using Pearson’s r test. C) Monotherapy Predictions by PERCEPTION: Showcases the predicted viability of monotherapies based on cell line-specific sc-expression, comparing resistant (N = 72) and sensitive (N = 84) lines using boxplots. Significance determined by one-tailed Wilcoxon rank-sum test. D) Sensitivity-Specificity Analysis: The receiver operator curve illustrates the balance between sensitivity and specificity in distinguishing between sensitive and resistant cell lines. Area under the curve (AUC) values are noted, with the dashed line representing random-model performance. E) & F) Drug Combination Response Predictions: Depict PERCEPTION’s predictions for drug combination responses in resistant (N = 28) vs. sensitive (N = 24) cell lines. G) Single-cell vs. Pseudo-bulk Level Analysis in PRISM Screens: Extends the analysis in panel A to single-cell and pseudo-bulk levels, highlighting the improved performance in pseudo-bulk data. The comparison includes predicted AUC values at both levels and experimental AUC values in PRISM for dabrafenib, AZD-7762, and trametinib, covering both testing (N = 80) and training cell lines (N = 318). H-K) Patient-Derived H&N Primary Cell Analysis: H) Prediction of Monotherapy Response: PERCEPTION’s predicted viability in resistant (n = 16) vs. sensitive (n = 16) lines. I) ROC Curve Analysis: Illustrates model’s prediction capability (sensitivity and specificity) for resistant vs. sensitive lines. AUC values are presented. J) & K) Combination Treatment Response: Similar analysis for combination treatments, comparing resistant (12) to sensitive (12) lines. All box plots show median, 25th/75th percentiles, and range.
Extended Data Fig. 4 Quality Control and Predictive Analyses in Lung Cancer Cell Line Screens.
A) Concordance between Lung Cancer and PRISM Screens: Illustrates the correlation (Rho on x-axis) and significance (y-axis) between our lung cancer screen and PRISM. Focuses on cell lines showing significantly positive correlation, as indicated by Pearson’s r test p-value. B) Predicted vs. Observed Viability Comparison: Analyzes the correlation between predicted and observed cell viability (N = 94 viability observations each, both centered and scaled). Pearson correlation and significance are noted. A best fit line with a 95% confidence interval is shown. C) Viability Prediction in Top vs. Bottom 50% Cell Lines: Compares predicted viability in resistant (N = 11, bottom 50%) versus sensitive (N = 10, top 50%) cell lines for each drug. Uses one-tailed Wilcoxon rank-sum test for statistical significance, presented for each drug. D) Combination Response Prediction in 21 Lung Cancer Cell Lines: Similar to panel B, this compares predicted versus observed combination viability (N = 49 viability observations each), with Pearson correlation and significance provided. A best fit line with a 95% confidence interval is included. E) Combination Viability Prediction in Top vs. Bottom 50% Cell Lines: Analyzes predicted combination viability (centered and scaled) for resistant (N = 11) and sensitive (N = 10) cell lines (based on observed viability) across 7 drug pairs. Uses one-tailed Wilcoxon rank-sum test for significance, presented for each combination. F) Consolidated Analysis of Monotherapies and Combinations: Integrates data from distinct drugs in panel E for combined analysis of monotherapies (N = 188) and drug combinations (N = 98). All box plots show median, 25th/75th percentiles, and range.
Extended Data Fig. 5 The predicted vs. experimental correlations obtained for individual treatments.
Each scatter plot compares the experimentally observed cell viability (x-axis; at median IC50 concentration) to the predicted viability (y-axis; rescaled AUC value) for the four drugs docetaxel, epothilone-b, gefitinib, and vorinostat (top four) and the pairwise combinations among {docetaxel, epothilone-b, gefitinib} (bottom three). Each dot represents the response of patient-derived cell lines (N = 5, color coded) for the drugs they were screened with. The Spearman rank correlation (cor) is provided at the bottom of each plot. These plots are provided for the following treatment concentrations - A) median IC50 B) one-third of median IC50. The error bands in all panels of this figure show 95% confidence interval of the fit.
Extended Data Fig. 6 Correlation of Predicted and Observed Viability in Monotherapies and Combination Treatments in Cell Lines.
Each scatter plot compares experimental cell viability (N = 20, x-axis; scaled per drug treatment) with predicted viability (N = 20, y-axis; rescaled AUC value). Points represent patient-derived cell line responses, color-coded by line and shape-coded by drug. Pearson correlation (R) is noted in each plot’s lower right corner. All panels feature error bands showing the 95% confidence interval of the fit. A) Monotherapy Response at Median IC50: Relation between monotherapy response and experimental response (N = 20 each). B) Combination Therapy Response at Median IC50: Similar analysis for combination therapy (N = 15 each). C) Monotherapy Response at 3x Median IC50: Examines monotherapy response at higher concentration (N = 20 each). D) Combination Therapy at 3x Median IC50: Analyzes combination therapy response at increased concentration (N = 15 each). E-G) Monotherapy and Combination Response Prediction in Lung Cancer Cell Lines: E) UMAP Clustering: Represents 53,514 cells from 199 cell lines (~300 cells/line) using sc-expression, identifying 29 clusters with cells from four unique sub-clones. F-G) Predicted Viability Based on Most-Resistant Clone: Viability predictions for 21 lung cancer cell lines (N = 11 resistant & 10 sensitive cell ines), considering the most resistant clone. Statistical significance assessed with two-sided Wilcoxon rank-sum test. H-I) Monotherapy and Combination Response Prediction in Patient-Derived HNSC Primary Cells (N = 5): H) Monotherapy Response Based on Most-Resistant Clone: Presents PERCEPTION predicted viability and resistance vs. sensitivity stratification (N = 2 resistant & 3 sensitive). Includes drugs docetaxel, epothilone-b, gefitinib, and vorinostat. I) Combination Response: Similar analysis for combination treatments. Both panels include a left-side plot for predicted viability in resistant (N = 2) vs. sensitive (N = 3) lines and a right-side ROC plot showing prediction power (sensitivity and specificity). AUC values are provided, with the dashed line indicating random-model performance. Statistical analysis performed with two-sided Wilcoxon rank-sum test. All box plots depict median, 25th/75th percentiles, and range.
Extended Data Fig. 7 Comparing PERCEPTION with Existing Bulk Response Models in a Breast Cancer Clinical Trial.
A) tSNE Transcriptional Clustering: Displays 36 transcriptional tumor clusters identified in the trial, integrating cells from 34 patients at three time points. Clusters, color-coded and defined in the legend, were derived using Seurat package. B) Malignant Sub-Clone Abundance: Shows the distribution of malignant sub-clones (y-axis) in breast cancer samples (x-axis), based on sc-expression. Different sub-clones are color-coded in the legend. Sample labels on the x-axis indicate patient id and time point of collection (“_S” - day 0, “_M” - day 14, “_E” - day 180). C) Pre-Treatment Clone-Level Response in Arms B and C: Predicted ribociclib viability (y-axis) versus various clones in pre-treatment samples (x-axis). Response status is displayed at the top of each column, with sample names below. Dot sizes represent the proportion of each cluster/clone, with a color scale indicating predicted viability (dark blue for low, yellow for high). D-E) Stratification Power of PERCEPTION vs Published Models: D) Bulk Expression-Based Models: Compares PERCEPTION with models trained only on bulk expression (N = 7 responders and 7 non-responders). E) Models Not Tuned on sc-Expression: PERCEPTION compared against models without sc-expression tuning (N = 7 responders and 7 non-responders). Both panels include deterministic model generation (seed=1) for training and test sets. Left-side plots present PERCEPTION predicted viability in responders vs. non-responders. Right-side ROC plots depict prediction power (sensitivity and specificity), with AUC values near the lower right corner. The dashed diagonal line indicates performance of a random model. Statistical significance assessed using two-sided Wilcoxon rank-sum test. F) Stratification Using Average sc-Viability: Stratifies responders (N = 7) vs. non-responders (N = 7) in combination therapy arms using average sc-viability in the FELINE trial. Statistical significance evaluated by two-sided Wilcoxon rank-sum test. Box plots show median, 25th/75th percentiles, and range.
Extended Data Fig. 8 Pre-processing and predicting clone level response in lung cancer patient cohort.
(A) A UMAP of 3671 malignant cells derived from 25 patients with 26,485 genes are clustered using Seurat considering the first 10 axes with the most variance. Each clone (a transcriptional cluster) output is annotated using a color where the legend is provided on the right. (B) The proportions of these clones (y-axis) are provided in each patient (x-axis) faceted by the time point at which these biopsies are collected. (C-F) Predicted viability of the four tyrosine kinase inhibitors: erlotinib, dabrafenib, osimertinib, and trametinib, in respective order, is provided at a clonal level for each patient where response status is provided at the bottom of each facet.
Extended Data Fig. 9 Correlation between the elapsed treatment time and estimated resistance holds true across different conditions.
In A-D), The extent of resistance to a treatment from the baseline (x-axis) is correlated with the treatment elapsed time (Number of days from the start of the treatment before the biopsy was taken) (y-axis). (A) The points and line colors denote the treatment administered to the patients listed by the right legend. B) Color denotes prior treatment. C) Color denotes the patient’s ID. D) Color denotes whether the disease is metastatic or primary at the time of biopsy. E) Extent of Resistance was calculated using bulk-expression of the tumor, where the increase with “Treatment Elapsed time” is positive, however, insignificant, and weaker than when the patient response is taken as the most-resistant clone available response. The error bands in all panels of this figure show 95% confidence interval of the fit.
Extended Data Fig. 10 Identifying Optimal Drug Combinations for Multiple Myeloma and Lung Cancer Patients.
A) Median Disjoint Killing Score (DKS) in Myeloma: For 94 drug pairs with positive DKS, the median DKS (y-axis) is plotted against each pair (x-axis). Color intensity denotes the proportion of patients (N = 12) with DKS > 0, with the top pairs labeled. Legend for color intensity is at the top. B) DKS for Triplets: Similar analysis for drug triplets. C) Clone-Level Disjoint Killing for Top Pairs: Viability profiles of clones for top pairs from C are shown for each patient (facet), with color intensity indicating post-treatment viability of each clone (x-axis) for a given drug (y-axis). Legend on the right. D) Clone-Level Disjoint Killing for Triplets: Analogous to C, but for drug triplets (N = 86, Triplets with DKS > 0). E-L) Analysis in Lung Cancer: E) Correlation in Clinical Trials: Examines the correlation between response difference of combination vs monotherapy (x-axis) and observed survival difference in combination vs single-treatment arms. Dot size represents patient numbers, with a best-fit line shown. Legend for dot sizes and error bands showing 95% confidence interval are at the top. Weighted Pearson’s r test p-value denotes correlation significance. F-H) Repeated for progression-free survival, overall survival, and erlotinib combinations. I) DKS for Lung Cancer Drug Combinations: Median DKS (y-axis) for 31 positive pairs plotted against each pair (x-axis). Color intensity shows proportion of patients with positive DKS, top pairs labeled, legend at the top. J-K) Disjoint Killing by Drug Class and Mechanism: Compares DKS (log10 value on y-axis) by general drug classes (N = 3 chemo+chemo, 7 chemo+targeted, 5 targeted+ targeted) (J) and mechanisms of action (N = 3 each MOA) (K). Evaluated by two-sided Wilcoxon rank-sum test. Box plots show median, 25th/75th percentiles, and range. L) Clone-Level Response in Lung Cancer: Shows post-treatment viability for top effective combinations, one facet per patient. Color intensity indicates clone viability (x-axis) for each drug (y-axis), for the top three patients ranked by highest DKS score per drug.
Supplementary information
Supplementary Tables
Supplementary Tables 1–9.
Rights and permissions
About this article
Cite this article
Sinha, S., Vegesna, R., Mukherjee, S. et al. PERCEPTION predicts patient response and resistance to treatment using single-cell transcriptomics of their tumors. Nat Cancer (2024). https://doi.org/10.1038/s43018-024-00756-7
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s43018-024-00756-7
