Diffusion Based Distributional Modeling of Conformers, Blind Docking and Proteins

Explore a comprehensive lecture on diffusion-based distributional modeling in molecular systems. Delve into the realms of conformer generation, blind docking, and protein modeling with MIT's Tommi Jaakkola. Discover innovative approaches to 3D conformer realization, including torsional diffusion and deep learning methods. Examine the concept of blind docking as a generative model and learn about the DiffDock performance with ESM folded structures. Investigate 3D motif scaffolding challenges and solutions, including conditioning via Sequential Monte Carlo. Gain insights into the integration of protein folding and design, and understand the Poisson flow inspired by electrostatics. This 54-minute presentation, part of IPAM's Learning and Emergence in Molecular Systems Workshop, offers a deep dive into cutting-edge techniques in computational biology and molecular modeling.

Intro
(1) Realizing likely 3D conformers
(1) Torsional diffusion for conformer generation
Search-based methods
Deep learning approaches
Rethinking blind docking as generative modeling
A case for generative docking
Generative pose prediction
Technical note: forward diffusion
De-noising (score) model
DiffDock: performance with ESM folded structures
3D motif scaffolding
(3) Backbone scaffolding challenge
(3) Conditioning via Sequential Monte Carlo
(3) Motif-scaffolding case-studies
(3) Integrating protein folding & design
Poisson flow - inspired by electrostatics