Presented at UWO BioSeminar, Western University.
This talk traces the evolution of computational protein structure prediction—from the MSA-based transformer architecture of AlphaFold2 to the all-atom diffusion model in AlphaFold3—and discusses what these advances mean for AI-driven drug discovery.
Topics Covered
- The protein folding problem — why predicting 3D structure from sequence is hard
- AlphaFold2 architecture — Evoformer, triangle multiplicative updates, and structure module
- AlphaFold3 — unified all-atom prediction with a diffusion-based decoder for proteins, DNA, RNA, and small molecules
- ESM family (Meta AI) — protein language models trained on evolutionary sequences
- RFdiffusion & ProteinMPNN — generative models for de novo protein design
- Connections to my research — how structure-grounded LLM reasoning builds on these representations
- Open challenges — conformational ensembles, intrinsically disordered regions, and co-folding