New paper on machine learning for cancer pharmacogenomics

Mutations in oncogenes influence the response of a patient’s tumour to drug treatments. Therefore, understanding the impact of these mutations on cancer cell sensitivity to drugs has the potential to make therapy safer and more effective by determining which drugs will be more appropriate for a given patient. The recent availability of the most comprehensive drug screening on cancer cell lines to date, measuring the sensitivity of about 700 human cancer cell lines to 120 drugs, has enabled a modelling approach to this problem.

In collaboration with groups at EBI, Sanger Institute and Harvard Medical School, we have investigated the prediction of cancer cell sensitivity to drugs using these data. In this study, an innovative multi-drug approach integrating genomic properties from cell lines and chemical information from drugs was proposed, which has led to powerful machine learning models. Results on blind tests showed that it is possible to impute missing IC50 values with remarkable accuracy and also to predict the IC50 of cell lines that have not been screened yet by exploiting a small part of the available experimental data. More importantly, this model was able to recapitulate a large fraction of the significant drug-oncogene associations, which demonstrates its suitability for the discovery of new predictive biomarkers.

This study has just been published at PLOS ONE: