Abstract
Cell-line screens create expansive datasets for learning predictive markers of drug response, but these models do not readily translate to the clinic with its diverse contexts and limited data. In the present study, we apply a recently developed technique, few-shot machine learning, to train a versatile neural network model in cell lines that can be tuned to new contexts using few additional samples. The model quickly adapts when switching among different tissue types and in moving from cell-line models to clinical contexts, including patient-derived tumor cells and patient-derived xenografts. It can also be interpreted to identify the molecular features most important to a drug response, highlighting critical roles for RB1 and SMAD4 in the response to CDK inhibition and RNF8 and CHD4 in the response to ATM inhibition. The few-shot learning framework provides a bridge from the many samples surveyed in high-throughput screens (n-of-many) to the distinctive contexts of individual patients (n-of-one).
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 datasets generated during and/or analyzed during the current study are all public data: CCLE: https://depmap.org/portal; CERES-corrected CRISPR gene disruption scores: https://depmap.org/portal; GDSC1000 dataset: https://www.cancerrxgene.org/gdsc1000/GDSC1000_WebResources; PDTC dataset: https://figshare.com/articles/Bruna_et_al_A_biobank_of_breast_cancer_explants_with_preserved_intra-tumor_heterogeneity_to_screen_anticancer_compounds_Cell_2016/2069274; PDX dataset: https://www.nature.com/articles/nm.3954. Other miscellaneous datasets that support the findings of the present study are available at http://github.com/idekerlab/TCRP. Source data are provided with this paper.
Code availability
The software implementation of TCRP, along with all supporting code, is available at http://github.com/idekerlab/TCRP. Other supporting software is available as follows: Scikit-learn v.0.20.2: http://scikit-learn.org/stable/index.html; PyTorch 1.0: http://pytorch.org.
References
Meyers, R. M. et al. Computational correction of copy number effect improves specificity of CRISPR–Cas9 essentiality screens in cancer cells. Nat. Genet. 49, 1779–1784 (2017).
Iorio, F. et al. A landscape of pharmacogenomic interactions in cancer. Cell 166, 740–754 (2016).
Brabetz, S. et al. A biobank of patient-derived pediatric brain tumor models. Nat. Med. 24, 1752–1761 (2018).
Bruna, A. et al. A biobank of breast cancer explants with preserved intra-tumor heterogeneity to screen anticancer compounds. Cell 167, 260–274.e22 (2016).
Butler, D. Translational research: crossing the valley of death. Nature 453, 840–842 (2008).
Lieu, C. H., Tan, A.-C., Leong, S., Diamond, J. R. & Eckhardt, S. G. From bench to bedside: lessons learned in translating preclinical studies in cancer drug development. J. Natl Cancer Inst. 105, 1441–1456 (2013).
Seyhan, A. A. Lost in translation: the valley of death across preclinical and clinical divide—identification of problems and overcoming obstacles. Trans. Med. Commun. https://doi.org/10.1186/s41231-019-0050-7 (2019).
Naumov, G. N. et al. Combined vascular endothelial growth factor receptor and epidermal growth factor receptor (EGFR) blockade inhibits tumor growth in xenograft models of EGFR inhibitor resistance. Clin. Cancer Res. 15, 3484–3494 (2009).
Lee, J. S. et al. Vandetanib versus placebo in patients with advanced non-small-cell lung cancer after prior therapy with an epidermal growth factor receptor tyrosine kinase inhibitor: a randomized, double-blind phase III trial (ZEPHYR). J. Clin. Oncol. 30, 1114–1121 (2012).
Parisot, J. P., Hu, X. F., DeLuise, M. & Zalcberg, J. R. Altered expression of the IGF-1 receptor in a tamoxifen-resistant human breast cancer cell line. Br. J. Cancer 79, 693–700 (1999).
Drury, S. C. et al. Changes in breast cancer biomarkers in the IGF1R/PI3K pathway in recurrent breast cancer after tamoxifen treatment. Endocr. Relat. Cancer 18, 565–577 (2011).
Lake, B. M., Salakhutdinov, R. & Tenenbaum, J. B. Human-level concept learning through probabilistic program induction. Science 350, 1332–1338 (2015).
Santoro, A., Bartunov, S., Botvinick, M., Wierstra, D. & Lillicrap, T. Meta-learning with memory-augmented neural networks. in Proc. 33rd International Conference on Machine Learning Vol. 48 (eds Balcan, M. F. & Weinberger, K. Q.) 1842–1850 (PMLR, 2016).
Dai, W., Yang, Q., Xue, G.-R. & Yu, Y. Boosting for transfer learning. in Proc. 24th International Conference on Machine Learning 193–200 (Association for Computing Machinery, 2007).
Blitzer, J., McDonald, R. & Pereira, F. Domain adaptation with structural correspondence learning. in Proc. 2006 Conference on Empirical Methods in Natural Language Processing 120–128 (EMNLP, 2006).
Argyriou, A., Evgeniou, T. & Pontil, M. Multi-task feature learning. in Advances in Neural Information Processing Systems Vol. 19 (eds Schölkopf, B. et al.) 41–48 (MIT Press, 2007).
Lake, B. M., Salakhutdinov, R. & Tenenbaum, J. B. The Omniglot challenge: a 3-year progress report. Curr. Opin. Behav. Sci. 29, 97–104 (2019).
Altae-Tran, H., Ramsundar, B., Pappu, A. S. & Pande, V. Low data drug discovery with one-shot learning. ACS Cent. Sci. 3, 283–293 (2017).
Medela, A. et al. Few shot learning in histopathological images: reducing the need of labeled data on biological datasets. in Proc. 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI, 2019); https://doi.org/10.1109/isbi.2019.8759182
Snell, J. et al. Prototypical Networks for Few-shot Learning. in Advances in Neural Information Processing Systems 4077–4087 (Curran Associates, 2017); https://proceedings.neurips.cc/paper/2017/hash/cb8da6767461f2812ae4290eac7cbc42-Abstract.html
Vinyals, O., Blundell, C., Lillicrap, T., Kavukcuoglu, K. & Wierstra, D. Matching networks for one shot learning. in Advances in Neural Information Processing Systems Vol. 29 (eds Lee, D. D. et al.) 3630–3638 (Curran Associates, 2016).
Finn, C., Abbeel, P. & Levine, S. Model-agnostic meta-learning for fast adaptation of deep networks. Proceedings of the 34th International Conference on Machine Learning 70, 1126–1135 (2017).
Preuer, K. et al. DeepSynergy: predicting anti-cancer drug synergy with Deep Learning. Bioinformatics 34, 1538–1546 (2018).
Yu, D.-D., Guo, S.-W., Jing, Y.-Y., Dong, Y.-L. & Wei, L.-X. A review on hepatocyte nuclear factor-1beta and tumor. Cell Biosci. 5, 58 (2015).
Gao, H. et al. High-throughput screening using patient-derived tumor xenografts to predict clinical trial drug response. Nat. Med. 21, 1318–1325 (2015).
Lipton, Z. C. The mythos of model interpretability. ACM Queue https://doi.org/10.1145/3236386.3241340 (2018).
Ma, J. et al. Using deep learning to model the hierarchical structure and function of a cell. Nat. Methods https://doi.org/10.1038/nmeth.4627 (2018).
Liu, F. & Matsuura, I. Inhibition of Smad antiproliferative function by CDK phosphorylation. Cell Cycle 4, 63–66 (2005).
Zhao, M., Mishra, L. & Deng, C.-X. The role of TGF-β/SMAD4 signaling in cancer. Int. J. Biol. Sci. 14, 111–123 (2018).
Zhang, F., Bick, G., Park, J.-Y. & Andreassen, P. R. MDC1 and RNF8 function in a pathway that directs BRCA1-dependent localization of PALB2 required for homologous recombination. J. Cell Sci. 125, 6049–6057 (2012).
Lu, C.-S. et al. The RING finger protein RNF8 ubiquitinates Nbs1 to promote DNA double-strand break repair by homologous recombination. J. Biol. Chem. 287, 43984–43994 (2012).
Kobayashi, S. et al. Rad18 and Rnf8 facilitate homologous recombination by two distinct mechanisms, promoting Rad51 focus formation and suppressing the toxic effect of nonhomologous end joining. Oncogene 34, 4403–4411 (2015).
Smith, R., Sellou, H., Chapuis, C., Huet, S. & Timinszky, G. CHD3 and CHD4 recruitment and chromatin remodeling activity at DNA breaks is promoted by early poly(ADP-ribose)-dependent chromatin relaxation. Nucleic Acids Res. 46, 6087–6098 (2018).
Larsen, D. H. et al. The chromatin-remodeling factor CHD4 coordinates signaling and repair after DNA damage. J. Cell Biol. 190, 731–740 (2010).
Prahallad, A. et al. Unresponsiveness of colon cancer to BRAF(V600E) inhibition through feedback activation of EGFR. Nature 483, 100–103 (2012).
Young, L. C. et al. SHOC2–MRAS–PP1 complex positively regulates RAF activity and contributes to Noonan syndrome pathogenesis. Proc. Natl Acad. Sci. USA 115, E10576–E10585 (2018).
Tzivion, G., Luo, Z. & Avruch, J. A dimeric 14-3-3 protein is an essential cofactor for Raf kinase activity. Nature 394, 88–92 (1998).
Schwartz, L. H. et al. RECIST 1.1—update and clarification: from the RECIST committee. Eur. J. Cancer 62, 132–137 (2016).
Yu, K. et al. Comprehensive transcriptomic analysis of cell lines as models of primary tumors across 22 tumor types. Nat. Commun. https://doi.org/10.1038/s41467-019-11415-2 (2019).
Ghandi, M. et al. Next-generation characterization of the Cancer Cell Line Encyclopedia. Nature 569, 503–508 (2019).
Li, T. et al. A scored human protein-protein interaction network to catalyze genomic interpretation. Nat. Methods 14, 61–64 (2017).
Cerami, E. G. et al. Pathway Commons, a web resource for biological pathway data. Nucleic Acids Res. 39, D685–D690 (2011).
Giurgiu, M. et al. CORUM: the comprehensive resource of mammalian protein complexes—2019. Nucleic Acids Res. 47, D559–D563 (2019).
Meyers, R. M. et al. Computational correction of copy-number effect improves specificity of CRISPR–Cas9 essentiality screens in cancer cells. Nat. Genet. 49, 1779–1784 (2017).
Kim, Y., Bismeijer, T., Zwart, W., Wessels, L. F. A. & Vis, D. J. Genomic data integration by WON-PARAFAC identifies interpretable factors for predicting drug-sensitivity in vivo. Nat. Commun. 10, 5034 (2019).
Harakalova, M. et al. Multiplexed array-based and in-solution genomic enrichment for flexible and cost-effective targeted next-generation sequencing. Nat. Protoc. 6, 1870–1886 (2011).
Glorot, X., Bordes, A. & Bengio, Y. Deep sparse rectifier neural networks. in Proc. Fourteenth International Conference on Artificial Intelligence and Statistics 15, 315–323 (2011).
Kingma, D. & Ba, J. Adam: a method for stochastic optimization. Preprint at arXiv https://arxiv.org/abs/1412.6980 (2014).
Baumann, D. & Baumann, K. Reliable estimation of prediction errors for QSAR models under model uncertainty using double cross-validation. J. Cheminform. 6, 47 (2014).
Ribeiro, M. T., Singh, S. & Guestrin, C. Why should I trust you?: explaining the predictions of any classifier. in Proc. 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 1135–1144 (2016).
Zou, H. & Hastie, T. Regularization and variable selection via the elastic net. J. R. Stat. Soc. Series B Stat. Methodol. 67, 301–320 (2005).
Tibshirani, R. Regression shrinkage and selection via the lasso. J. R. Stat. Soc. Series B Stat. Methodol. 58, 267–288 (1996).
Binder, A., Montavon, G., Lapuschkin, S., Müller, K.-R. & Samek, W. Layer-wise relevance propagation for neural networks with local renormalization layers. in Artificial Neural Networks and Machine Learning—ICANN 2016 (eds Villa, A. et al.) 63–71 (Springer, 2016).
Simonyan, K., Vedaldi, A. & Zisserman, A. Deep inside convolutional networks: visualising image classification models and saliency maps. International Conference on Learning Representations https://openreview.net/forum?id=cO4ycnpqxKcS9 (2014).
Acknowledgements
We thank the following for their support for the present study: the National Cancer Institute for grants (nos. U54CA209891 to T.I., R01CA204173 to C.B. and K22CA234406 to J.S.), the National Institute of General Medical Sciences for a grant (no. P41GM103504 to T.I.) and the National Human Genome Research Institute for a grant (no. R01HG009979 to T.I.). R.S. was supported by a research grant from the Israel Science Foundation (grant no. 715/18). J.P. was supported by a grant from the National Science Foundation (grant no. 1652815). L.W. and S.M. were supported by the ZonMw TOP grant COMPUTE CANCER (40-00812-98-16012). J.S. was supported by the Cancer Prevention and Research Institute of Texas (CPRIT RR180035).
Author information
Authors and Affiliations
Contributions
J.M. and T.I. designed the study and developed the conceptual ideas. J.M. and Y.L. implemented the main algorithms. J.M. and S.H.F. collected all the input sources. J.M., S.M., L.F.A.W. and M.H. developed the strategy for alignment of in vitro and in vivo drug responses. J.M., C.J.B. and T.I. interpreted the results. J.M., S.H.F., R.S., C.J.B., J.P., J.P.S. and T.I. wrote the manuscript.
Corresponding author
Ethics declarations
Competing interests
T.I. is co-founder of Data4Cure, Inc., is on the Scientific Advisory Board and has an equity interest. T.I. is on the Scientific Advisory Board of Ideaya BioSciences, Inc. and has an equity interest. The terms of these arrangements have been reviewed and approved by the University of California San Diego in accordance with its conflict-of-interest policies. L.W. received project funding from Genmab BV. The other authors declare no competing interests.
Additional information
Peer review information Nature Cancer thanks Cyril Benes, Roland Eils and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Extended data
Extended Data Fig. 1 Analysis of fitness versus predictive performance for the panel of gene knockouts in our study.
a, Distribution of relative growth values after CRISPR gene knockout, median for all n = 341 cell lines. Blue: pooling knockouts of all n = 17670 genes; Pink: pooling n = 469 knockouts of genes selected in our study. Fitness is corrected by the Copy Number Variation by the CERES algorithm. b, For each knockout of a selected gene, predictive performance (y axis) is computed as the Pearson correlation between predicted and actual growth measurements over all n = 341 cell lines. This performance is displayed as a function of the median growth fitness of that knockout (x axis). Growth fitness is binned according to percentiles, for example the first bin (0-10%) represents the top 10% of selected genes with the strongest median effects on growth. The distribution of predictive performance for each bin is shown with a violin plot. Error bars represent 95% confidence interval.
Extended Data Fig. 2
Training accuracy of TCRP and other baseline models for all challenges.
Extended Data Fig. 3 Alternative calculation of model performance using Spearman correlation.
While Pearson correlation is used to calculate model performance in the main text, this supplemental figure provides equivalent performance calculations using the non-parametric rank-based Spearman correlation. a, Related to Fig. 3b on n = 83 PDTC models. b, Related to Fig. 4a on n = 228 PDX models.
Extended Data Fig. 4 Comparison of transferability of different machine learning models to patient-derived xenografts.
Predictive models were pre-trained using responses of cancer cell lines to perturbations with drugs, one model per drug. Few-shot learning was then performed on 0-10 PDX breast tumor samples exposed to that drug (x-axis), and model accuracy (y-axis) was measured by a, Pearson correlation or b, Spearman correlation on the remaining held-out PDX samples. Results averaged across five drugs (see main text). This experiment considers n = 228 PDX models.
Supplementary information
Source data
Source Data Fig. 2
Raw data file.
Source Data Fig. 3
Raw data file.
Source Data Fig. 4
Raw data file.
Source Data Fig. 5
Raw data file.
Source Data Fig. 6
Raw data file.
Source Data Extended Data Fig. 1
Raw data file.
Source Data Extended Data Fig. 2
Raw data file.
Source Data Extended Data Fig. 3
Raw data file.
Source Data Extended Data Fig. 4
Raw data file.
Rights and permissions
About this article
Cite this article
Ma, J., Fong, S.H., Luo, Y. et al. Few-shot learning creates predictive models of drug response that translate from high-throughput screens to individual patients. Nat Cancer 2, 233–244 (2021). https://doi.org/10.1038/s43018-020-00169-2
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1038/s43018-020-00169-2
This article is cited by
-
Artificial intelligence in the risk prediction models of cardiovascular disease and development of an independent validation screening tool: a systematic review
BMC Medicine (2024)
-
Meta-learning-based Inductive logistic matrix completion for prediction of kinase inhibitors
Journal of Cheminformatics (2024)
-
Few-shot classification of ultrasound breast cancer images using meta-learning algorithms
Neural Computing and Applications (2024)
-
A Modified Neighborhood Hypothesis Test for Population Mean in Functional Data
Journal of Agricultural, Biological and Environmental Statistics (2024)
-
Patient-derived xenograft models in cancer therapy: technologies and applications
Signal Transduction and Targeted Therapy (2023)