Institute of Information Science, Academia Sinica

Targeted and chemo-therapies are ubiquitous in cancer treatment. Identification of cancer patients who will respond to these therapies remains challenging for precision medicine. As large drug response datasets have been generated for cell lines and patient drug response data are limited, methods to efficiently transfer predictors trained by cell line datasets to human setting will be important for clinical practice. We propose feature selection procedures to combine with a logit model (LogitDA) and with K-nearest neighbor (KNNDA); these procedures in fact can be combined with any classifier, thus are versatile. We selected genes which have similar conditional distributions across the source and target domains, prioritized genes by their differential expression between sensitive versus resistant cell lines, and ranked these genes by their power to separate sensitive from resistant cell lines. Next, we trained a logit model and KNN using expression data of the top-ranked p (≤ 1000) genes of cell lines in GDSC. These predictors are shown to significantly outperform the baseline logit model and a deep-learning based predictor, in terms of high prediction AUC. In particular, they achieve 0.70-1.00 AUC for seven of the ten drugs tested, using merely a few hundred genes, thus are interpretable. We further adjust the prediction probability cutoff for LogitDA to yield high prediction accuracy for datasets which do not satisfy the assumption. In particular, LogitDA achieves prediction accuracy 0.70-0.93 for seven of the ten drugs, e.g. Erlotinib and Cetuximab. Finally, pathways relevant to anti-cancer therapies and metabolism are uncovered for Erlotinib and Cetuximab.

Prof. Shieh’s team has worked on computational approaches to reveal prognostic/prediction biomarkers for various cancers. In the past three years, her team has focused prediction drug response of cancer patients to targeted and chemotherapies, and immunotherapies. They are developing machine-learning methods using multi-omics data.