Equivariant 3D-Conditional Diffusion Models for Molecular Linker Design

Explore an in-depth lecture on equivariant 3D-conditional diffusion models for molecular linker design. Delve into the innovative DiffLinker approach, which addresses challenges in fragment-based drug discovery by designing linkers between disconnected molecular fragments. Learn about the model's E(3)-equivariant architecture, its ability to connect multiple fragments, and its automatic determination of atom numbers and attachment points. Examine the forward diffusion process, denoising techniques, and the implementation of equivariant graph neural networks. Discover how DiffLinker outperforms existing methods in generating diverse and synthetically-accessible molecules, and see its practical applications in real-world scenarios, including target protein pocket conditioning. Gain insights into the model's performance, limitations, and potential impact on drug discovery through a comprehensive presentation and Q&A session.

- Intro
- Examples of Structure-Based Drug Design
- Problem Setup
- Existing Methods for Linker Design
- Diffusion Models
- Forward Diffusion Process
- Denoising Process
- Equivariance and 3D Conditioning
- Equivariant Graph Neural Network
- DiffLinker - Predicting the # of Atoms in the Linker
- Results Overview
- Pocket Conditioning
- Performance & Conclusion
- Q+A