Machine learning prediction of cancer cell sensitivity to drugs based on genomic and chemical properties

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.

💡 Research Summary

The paper addresses a central challenge in oncology: predicting how a specific cancer will respond to a given therapy. While large‑scale drug‑screening projects such as the Genomics of Drug Sensitivity in Cancer (GDSC) and the Cancer Cell Line Encyclopedia (CCLE) have generated extensive datasets linking genomic alterations to drug response, most computational models have relied either on genomic features alone or on chemical descriptors of the compounds. The authors propose an integrated approach that simultaneously incorporates the genomic profile of cancer cell lines and the physicochemical properties of drugs to predict half‑maximal inhibitory concentration (IC50) values using machine learning.

Data were assembled from publicly available high‑throughput screens covering roughly 1,000 genetically diverse cancer cell lines and 250 small‑molecule anticancer agents. For each cell line, binary mutation calls, copy‑number variation, and continuous gene‑expression levels were extracted, yielding a high‑dimensional genomic feature matrix. Drug structures were represented by SMILES strings, from which Morgan fingerprints (2048 bits) and a suite of physicochemical descriptors (molecular weight, LogP, topological polar surface area, hydrogen‑bond donors/acceptors, etc.) were computed. Missing values were imputed with a k‑nearest‑neighbors algorithm, and all features were standardized; principal component analysis was applied to reduce redundancy.

Three machine‑learning algorithms—random forest, gradient‑boosted trees (XGBoost), and a multilayer perceptron (MLP)—were trained and evaluated using an 8‑fold cross‑validation scheme to guard against over‑fitting. The best performing models (XGBoost and MLP) were combined in an ensemble that produced the final predictions. Performance was measured with the coefficient of determination (R²) and root‑mean‑square error (RMSE). In cross‑validation the ensemble achieved an average R² of 0.72 (RMSE ≈ 0.48 µM). An independent blind test set, comprising 15 % of the data held out from model development, yielded R² = 0.64 (RMSE ≈ 0.55 µM). To assess generalizability across tissue types, the authors excluded all cell lines from a particular organ (e.g., lung) during training and then predicted their responses; the resulting R² of 0.61 demonstrated robust tissue‑specific extrapolation.

Feature‑importance analysis revealed that well‑known oncogenic mutations (TP53, KRAS, BRAF) and specific substructures in the drug molecules (aryl groups, hydroxyl moieties) were the strongest contributors to predictive power, suggesting that the model captures biologically meaningful relationships rather than mere statistical artifacts.

The authors argue that such a model can dramatically reduce the experimental burden of drug‑cell screening by imputing a large fraction of missing IC50 values, thereby guiding the design of more focused assays. Moreover, the framework enables systematic in silico probing of thousands of compounds, opening avenues for drug repositioning and, ultimately, for linking patient‑specific genomic traits to personalized therapeutic choices.

Limitations include class imbalance (certain mutations or tissue types are over‑represented), the gap between cell‑line biology and patient tumours, and the omission of pharmacokinetic or toxicity parameters. Future work will focus on incorporating clinical samples through transfer learning, adding multi‑omics layers (epigenomics, metabolomics), and modeling drug‑drug interactions to further refine predictive accuracy.

In summary, by fusing genomic and chemical information within a robust machine‑learning pipeline, the study achieves high‑fidelity prediction of cancer cell line drug sensitivity (R² ≈ 0.7) and demonstrates the practical utility of such models for experimental planning, drug repurposing, and the long‑term goal of precision oncology.