Yang Shen - Forward Prediction and Reverse Design for Molecular Discovery скачать в хорошем качестве

Yang Shen - Forward Prediction and Reverse Design for Molecular Discovery

Presented on 3-11-2021 Abstract This talk focuses on forward prediction and inverse design for protein-ligand interactions at both the molecular and the system levels. The first part will be centered around forward prediction of protein-ligand affinities at the molecular level. Building upon our semi-supervised DeepAffinity with attention mechanisms, we have recently progressed in both model interpretability and protein embedding. 1) Upon answering how to define and assess model interpretability in this context, we have enhanced model interpretability through structure-aware regularization and supervision of joint attentions; and have even built interpretability intrinsically into model architectures. 2) For protein embedding, we regard proteins as multi-modal data including languages (sequences) and graphs (sequence-predicted contact maps). We show that cross-modality protein embedding can improve model generalizability; and our recent methods for self-supervised graph representation learning brought further improvements. The second part will be centered around inverse design of ligand combinations for a disease module (at the system level of protein-protein interaction subnetworks). We have developed a hierarchical variational graph encoder to embed diseases through disease-disease interactions, disease-gene associations, and gene-gene interactions. With a network-principled loss besides chemical and adversarial losses, we have introduced a reinforcement learning agent to generate ligand combinations as graph sets for a target disease. Numerical results show that the principle-respecting designs are predicted to have lower toxicity from drug-drug interactions and reproduce FDA-approved polypharmacology strategies in case studies. If time permits, I will also briefly introduce our latest efforts in protein docking and protein design. Bio Dr. Yang Shen is an assistant professor in the Department of Electrical and Computer Engineering at Texas A&M University (TAMU). His research interests are in optimization and learning algorithms for modeling biological molecules, systems, and data. In particular, ongoing projects include prediction and design of protein interactions, mechanistic prediction of protein mutational effects, and resistance-overcoming drug design. Dr. Shen received his B.E. from the University of Science and Technology of China, Ph.D. in Systems Engineering from Boston University, and postdoctoral training in Biological Engineering and Computer Science at the Massachusetts Institute of Technology. Prior to joining TAMU, he had been a research assistant professor at Toyota Technological Institute at Chicago. He is a recipient of the MIRA award for early stage investigators from the National Institute of General Medical Sciences (2017) and the CAREER award from the National Science Foundation (2020). Lab website: https://shen-lab.github.io/ Seminar Schedule: https://docs.google.com/document/d/1G...