@article {Zhao026534, author = {Cheng Zhao and Ying Li and Zhaleh Safikhani and Benjamin Haibe-Kains and Anna Goldenberg}, title = {Using cell line and patient samples to improve predictions of patient drug response}, elocation-id = {026534}, year = {2015}, doi = {10.1101/026534}, publisher = {Cold Spring Harbor Laboratory}, abstract = {Background Recent advances in high-throughput technologies have facilitated the profiling of large panels of cancer cell lines with responses measured for thousands of drugs. The computational challenge is now to realize the potential of these data in predicting patients{\textquoteright} responses to these drugs in the clinic.Methods We address this issue by examining the spectrum of prediction models of patient response: models predicting directly from cell lines, those predicting directly from patients, and those trained on cell lines and patients at the same time. We tested 21 classification models on four drugs, that are bortezomib, erlotinib, docetaxel and epirubicin, for which clinical trial data were available.Results Our integrative models consistently outperform cell line-based predictors, indicating that there are limitations to the predictive potential of in vitro data alone. Furthermore, these integrative models achieve better predictive accuracy and require substantially fewer patients than would be the case if only patient data were available.Conclusions The integration of in vitro and ex vivo genomic data results in more accurate predictors using only a fraction of the patient information, which can help optimize the development of personalized predictors of therapy response. Altogether our results support the relevance of preclinical data for therapy prediction in clinical trials, enabling more efficient and cost-effective trial design.AUCArea under the drug dose-response curveAUROCArea Under the Receiver Operating Characteristic CurveC2PModel predicting patients{\textquoteright} drug response from in vitro (cancer cell lines) dataCDFchip definition fileCGPCancer Genome ProjectCP2PModel predicting patients{\textquoteright} drug response from the combination of in vitro (cancer cell lines) and ex vivo (patient tumors) dataIC50Drug concentration required to inhibit 50\% of the maximal cellular growth of a given cell lineNSCLCnon-small cell lung cancerP2PModel predicting patients{\textquoteright} drug response from ex vivo (patient tumors) dataPCAprincipal component analysisROCreceiver operating characteristic}, URL = {https://www.biorxiv.org/content/early/2015/09/15/026534}, eprint = {https://www.biorxiv.org/content/early/2015/09/15/026534.full.pdf}, journal = {bioRxiv} }