Background: Clinical responses to anti-cancer therapies often only benefit a defined subset of patients. Predicting the best treatment strategy hinges on our ability to effectively translate genomic data into actionable information on drug responses.
Results: To achieve this goal, we compiled a comprehensive collection of baseline cancer genome data and drug response information derived from a large panel of cancer cell lines. This data set was applied to identify the signature genes relevant to drug sensitivity and their resistance by integrating CNVs and the gene expression of cell lines with in vitro drug responses. We presented an efficient in-silico pipeline for integrating heterogeneous cell line data sources with the simultaneous modeling of drug response values across all the drugs and cell lines. Potential signature genes correlated with drug response (sensitive or resistant) in different cancer types were identified. Using signature genes, our collaborative filtering-based drug response prediction model outperformed the 44 algorithms submitted to the DREAM competition on breast cancer cells. The functions of the identified drug response related signature genes were carefully analyzed at the pathway level and the synthetic lethality level. Furthermore, we validated these signature genes by applying them to the classification of the different subtypes of the TCGA tumor samples, and further uncovered their in vivo implications using clinical patient data.
Conclusions: Our work may have promise in translating genomic data into customized marker genes relevant to the response of specific drugs for a specific cancer type of individual patients.
Keywords: cancer cell line; drug resistance; drug response; drug sensitivity; personalized treatment.