Better understand the AlphaFold-like architecture

[kntkb]

Youtube

• [x] AlphaFold2, OpenFold, Protein Language Models and Beyond | Nazim Bouatta from Molecular Modeling and Drug Discovery (M2D2) 2022
• [x] Nazim Bosatta - Insights from retraining OpenFold from CASP meeting 2023

AlphaFold2 architecture

• [x] Nazim Bouatta | Machine learning for protein structure prediction, Part 1: Algorithm space
• [ ] Nazim Bouatta | Machine learning for protein structure prediction, Part 2: AlphaFold2 architecture
• [ ] Nazim Bouatta | Machine learning for protein structure prediction, Part 3: AlphaFold2 and OpenFold

papers

• [x] OpenFold: Retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization
• [ ] Single-sequence protein structure prediction using a language model and deep learning
• [ ] Evolutionary-scale prediction of atomic-level protein structure with a language model

[kntkb]

AlphaFold2, OpenFold, Protein Language Models and Beyond | Nazim Bouatta from Molecular Modeling and Drug Discovery (M2D2) 2022

Remaining challenges in AlphaFold2 (or potentially related methods that rely on MSA)

• Prediction of orphan and rapidly evolving proteins for which an MSA cannot be generated (or for which we have a shallow MSA)
• Analysis of allelic variations and mutations (difficult to predict the impact of single/multiple point mutations, pipeline not designed for such purposes)
• Rapid and systematic exploration of designed proteins (why??)
• Conformational changes
• Still somewhat a blackbox (does not give explanation of how and why protein folds)

Idea to overcome some of these challenges --> OpenFold

• Rethink how proteins fold (application of LLM other than MSA)
• Retrain the model for specific tasks

[kntkb]

• [ ] ColabFold - Making protein folding accessible to all via Google Colab!

[kntkb]

• [ ] How to compute co-evolutionary matrix