Explore a comprehensive lecture on chemically transferable generative backmapping of coarse-grained proteins presented by Soojung Yang from Valence Labs. Delve into the challenges of backmapping in molecular simulations and discover innovative machine learning approaches to address transferability and reliability issues. Learn about the combined set of innovations, including representation based on internal coordinates, equivariant encoder/prior, custom loss functions, and expert curation of high-quality protein data. Gain insights into the conditional VAE framework, protein ensemble database, equivariant message passing, backbone and sidechain reconstruction, and various ablation studies. The lecture concludes with a discussion on diversity metrics and a Q&A session, providing a thorough understanding of this cutting-edge research in computational biology and drug discovery.
Overview
Syllabus
- Intro
- Glossary
- Protein Conformation and Energy Landscape
- Seminal Prior Work
- Conditional VAE Framework of Generative Backmapping
- Can Backmapping be Made Transferable?
- Proof of Concept Study
- Data: Protein Ensemble Database
- Encoder and Prior: Equivariant Message Passing
- Decoder: Backbone Reconstruction
- Encoder: Sidechain Reconstruction
- Learning Objectives
- Ablation Study 1: Transferability
- Ablation Study 2: Equivariance of the Encoder
- Ablation Study 3: Decoder Degrees of Freedom
- Diversity Metrics
- Q+A
Taught by
Valence Labs
