IDEA #9RIVD0 MultiDCP: High Throughput Predictive Modeling of Chemical Phenomics

Chemical phenomics, particularly dose-dependent chemical-induced transcriptomics, proteomics and drug-response curve, provides an efficient approach to mechanism-driven phenotype-based drug discovery. However, state-of-the-art machine learning methods of chemical phenomics are less successful in screening compounds for novel cells and new patients because only limited number of cell lines have chemical treatment data. We designed a new multi-task deep learning model, MultiDC, to expand the scope of chemical phenomics to unseen cells and individual patients by incorporating a knowledge-driven autoencoder and integrating labeled and unlabeled omics data. MultiDC significantly outperforms the state-of-the-arts, for the first time predicts dose-dependent chemical proteomics and drug sensitivity for novel cells and patients, and demonstrates a stronger predictive power than noisy experimental data for downstream tasks.