Abstract : Cancer patients often respond differently to the same treatment. Precision oncology aims at predicting which treatments will be effective on a given patient. Such predictive biomarkers of drug response typically take the form of a particular somatic mutation. However, lessons from the past indicate that these single gene-drug response associations are rare and/or often fail to achieve a significant impact in clinic. In this context, Machine Learning (ML) is emerging as a particularly promising complementary approach to precision oncology. Our results show that combining multiple gene alterations of the tumours via ML often results in better discrimination than that provided by the corresponding single-gene marker. This approach also permits assessing which type of molecular profile is most predictive of tumour response depending on treatment and cancer type. Moreover, ML multi-gene predictors generally retrieve a much higher proportion of treatment-sensitive tumours (i.e. they have a higher recall) than the corresponding single-gene marker. The latter suggest that a higher proportion of patients could benefit from precision oncology by applying this ML methodology to existing clinical pharmacogenomics data sets.

Keywords : cancer; drug sensitivity​; biomarker; machine learning; genomics