Source-linked AI summary
Machine learning prediction of cancer cell sensitivity to drugs based on genomic and chemical properties
Michael P. Menden, Francesco Iorio, Mathew Garnett, Ultan McDermott, Cyril Benes, Pedro J. Ballester, Julio Saez-Rodriguez
TL;DR
Predicting cancer-cell drug sensitivity from genomic and chemical information could support personalized therapy, but existing approaches have treated these sources separately. This paper integrates both in machine-learning models, achieving R2 values of 0.72 in cross-validation, 0.64 in an independent blind test, and 0.61 across tissues.
Problem
Personalized cancer therapy needs robust tools connecting tumour genomic properties with drug sensitivity.
Method
The authors build multi-drug machine-learning models using cancer-cell genomic features and drug chemical information derived from molecular structures.
Results
Multi-drug models imputed missing IC50 values and identified genomic events associated with altered drug sensitivity.
Takeaways & Limitations
The framework can support in silico drug-screening design and genomic analysis of altered drug sensitivity.
Takeaways & Limitations
Predictive ability for individual IC50 values remains limited and could be improved using alternative response metrics.
Abstract
from arXiv · showhide
Predicting the response of a specific cancer to a therapy is a major goal in modern oncology that should ultimately lead to a personalised treatment. High-throughput screenings of potentially active compounds against a panel of genomically heterogeneous cancer cell lines have unveiled multiple relationships between genomic alterations and drug responses. Various computational approaches have been proposed to predict sensitivity based on genomic features, while others have used the chemical properties of the drugs to ascertain their effect. In an effort to integrate these complementary approaches, we developed machine learning models to predict the response of cancer cell lines to drug treatment, quantified through IC50 values, based on both the genomic features of the cell lines and the chemical properties of the considered drugs. Models predicted IC50 values in a 8-fold cross-validation and an independent blind test with coefficient of determination R2 of 0.72 and 0.64 respectively. Furthermore, models were able to predict with comparable accuracy (R2 of 0.61) IC50s of cell lines from a tissue not used in the training stage. Our in silico models can be used to optimise the experimental design of drug-cell screenings by estimating a large proportion of missing IC50 values rather than experimentally measure them. The implications of our results go beyond virtual drug screening design: potentially thousands of drugs could be probed in silico to systematically test their potential efficacy as anti-tumour agents based on their structure, thus providing a computational framework to identify new drug repositioning opportunities as well as ultimately be useful for personalized medicine by linking the genomic traits of patients to drug sensitivity.
2 Cancer Genome Project, Wellcome Trust Sanger Institute, Wellcome Trust Genome Campus -- · 3 Center for Molecular Therapeutics, Massachusetts General Hospital Cancer Center and Harvard
The study developed integrative machine-learning models that combine cancer-cell genomic features with drug chemical properties to predict IC50 values. These models accurately imputed missing responses, generalized to unseen cell lines and tissues, and supported drug-screening design and repositioning analyses.
- Introduction: High-throughput drug screens reveal genomic determinants of sensitivity but produce incomplete IC50 matrices whose experimental completion is expensive and laborious.IC50 denotes the half maximal (50%) inhibitory concentration with respect to cell viability.
- Results: 0.72 R2 was achieved in 8-fold cross-validation, alongside Rp of 0.85 and RMSE of 0.83 across all 111 drugs.Random forests achieved Rp of 0.85, R2 of 0.72 and RMSE of 0.84.
- Results: 0.64 R2 was obtained on a blind test of 13,565 newly generated IC50 values, with Rp of 0.79 and RMSE of 0.97.The blind-test matrix was updated by ~18%, and the new IC50s were used exclusively as the blind test set.
- Results: 0.61 R2 was achieved when all 106 lung cancer cell lines were excluded from training, with Rp of 0.79 and RMSE of 0.99.A stricter cell-line-held-out validation achieved Rp of 0.82, R2 of 0.68 and RMSE of 0.89 for completely new cell lines.
- Discussion: Multi-drug models had significantly better Pearson correlation and R2 than single-drug models while maintaining similar RMSE, covering a broader dynamic range with similar precision.Chemical descriptors provided two orders of magnitude more training data than each single-drug model.
- Discussion: The models can estimate responses for unscreened cell lines, balance screening coverage against accuracy, evaluate novel chemical libraries, identify repurposing candidates, and support patient-tailored therapeutic strategies.Individual-value prediction remains limited, and future models may benefit from transcriptional, phosphoproteomic, epigenetic, treatment-response, and network features.
Figures:
The figures show a machine-learning workflow combining genomic and chemical drug features to predict log(IC50), with strong performance in cross-validation, blind testing, sparse-matrix completion, and prediction across tissues.
- IC50 prediction workflow: The workflow combines mutation states of 77 oncogenes with PaDEL-generated drug features from SMILES and predicts continuous IC50 values using neural networks and random forests.The method outputs the base-10 logarithm of IC50 in µM.
- Matrix completion: 20% of the whole cell-to-drug matrix achieved Rp of 0.84 instead of 0.85 and R2 of 0.70 instead of 0.72, indicating similar prediction accuracy with sparse experimental data.The published dataset was approximately 58% complete because of logistical constraints.
- Blind test: 0.97 RMSE, 0.64 R2, and 0.79 Rp were obtained on the blind test, compared with 0.83 RMSE, 0.72 R2, and 0.85 Rp in cross-validation.The blind test contained 13,565 novel IC50 values, approximately 18% additional data points excluded from training.
- Prediction across tissues: For lung cell lines excluded from training, RMSE increased from 0.83 to 0.99, R2 declined from 0.72 to 0.61, and Rp decreased from 0.85 to 0.79.Lung cell lines represented approximately 17% of the data.