Deep Learning in Structural Biology and Protein Design: How, Where, and Why
summary
Explains how AlphaFold2 combines MSAs and an EvoFormer with recycling iterative refinement to predict all-atom coordinates and outputs a confidence score well calibrated; shows that evolutionary MSAs remain the dominant input and that search for related sequences is the main time cost. Notes DeepMind’s release of 200 million predicted structures versus ~100k PDB entries (~0.1% coverage of known sequences), describes diffusion and inverse‑folding generative models, reports perplexity improvements (perplexity 10 to 4 with structure) and uses in speeding cryo‑EM interpretation and batched design libraries.
AI-written summary · about
chapters (17)
- 0:00Introduction to speaker
- 0:38Deep learning in biology
- 2:20Biological systems overview
- 7:24The role of proteins
- 10:38Protein structures
- 14:26Protein data bank
- 16:53Structure prediction
- 24:35AlphaFold applications
- 28:12Data scale and types
- 30:20Generative modeling
- 34:41Protein language models
- 38:57Sequence design
- 45:20Evolutionary data
- 49:13Traditional modeling
- 53:27AlphaFold refinement
- 59:19Future challenges
- 1:06:14Q&A and conclusion
related talks
53mHow machine learning is influencing protein engineering
04mHow Machine Learning tools at Google made AlphaFold possible
29mProtein folding and what it means for drug discovery
36mAlgorithmic Alchemy: Transforming Drug Discovery with Open AI
46mMachine Learning-based Design of Proteins and Small Molecules
14mAI Is Built to Be Applied
14mAI for Science
06mTools vs Products: Lessons learned from a niche industry