Overview
Explore a comprehensive 57-minute talk on DiffDock, a novel approach to molecular docking using diffusion generative models. Learn about the challenges in predicting small molecule ligand binding to proteins and how DiffDock addresses these issues. Discover the innovative use of diffusion processes on non-Euclidean manifolds, the product space mapping of degrees of freedom, and the significant improvements in accuracy compared to traditional and deep learning methods. Gain insights into the fast inference times and confidence estimates provided by DiffDock. Delve into topics such as the basics of molecular docking, problems with existing deep learning models, diffusion generative models, product space diffusion, score models, and the overall workflow. Examine the results, reverse diffusion process, and confidence score quality through detailed explanations and Q&A sessions.
Syllabus
- Intro
- What is Molecular Docking?
- Problem with Deep Learning Models for Docking
- Diffusion Generative Models for docking
- Product Space Diffusion
- Score Model
- Workflow Overview
- Q+A
- Results
- Reverse Diffusion Process and Confidence Score Quality
- Q+A
Taught by
Valence Labs