TR&D 3: NETWORK GUIDED MACHINE LEARNING – PROJECT SUMMARY Powerful machine learning techniques, including recent advances in deep learning, promise to revolutionize our ability to predict biomedical outcomes and could one day replace people in tasks such as image analysis, medical diagnosis, and precision therapy. However, most machine learning methods result in “black boxes” which do not provide the mechanistic understanding needed to control and repair biological systems in industrial and medical applications. Furthermore, machine learning models trained for high accuracy in one context, such as predicting drug responses of cell lines, often transfer poorly to other contexts such as predicting drug responses of patients. How can we gain both the predictive power of machine learning and the interpretability and transferability of mechanistic models of biology? Here we explore a series of complementary and innovative approaches to this question based on integrating machine learning models with biological networks. Specifically, we aim to use networks to: [Aim 1] Guide the transfer of predictive models of drug response from model systems to patients; [Aim 2] Apply machine learning models to genotype-phenotype prediction in genome-wide association studies; and [Aim 3] Use machine learning for patient diagnosis and clinical trial selection in precision medicine applications. These aims are motivated by Driving Biomedical Projects focused on drug response prediction in cell lines and patients (Aim 1; DBPs 13,19), genome-wide association analysis of disease (Aim 2; DBPs 14-15,17), predicting patient outcomes in cancer and major depression (Aim 3; DBPs 8,16), and clinical trial design (Aim 3; DBP 18). Our methods will be made available as open source software and on prominent cloud-based biomedical data and computing environments (TPs 6,7) to support wide adoption. ​​ ​​ ​​