AlphaFold2 Structure Prediction

In this session, we explore the fundamentals of AlphaFold 2 and why it revolutionized protein structure prediction. We begin with its breakthrough performance in CASP14, where it dramatically outperformed competing methods, and then dive into the core architecture—highlighting key concepts like multiple sequence alignments (MSAs), the Evoformer, and recycling. Through an interactive Q&A, we unpack practical considerations such as when MSAs help (or hurt), limitations in predicting point mutations, and how sampling and dropout influence model diversity. We also discuss real-world workflows, parameter tuning, and extensions like AlphaFold Multimer and database-scale predictions. The session concludes with a hands-on activity where you’ll experiment with different AlphaFold parameters to understand how they impact structure prediction outcomes. Video on the Alphafold architecture: https://www.youtube.com/watch?v=P_fHJIYENdI Video on transformers: https://www.youtube.com/watch?v=wjZofJX0v4M More info about this Bootcamp can be found at https://rosettamlbootcamp2025.github.io/ Instructors: Ian Anderson (UC Davis) & Ben Orr (UCSF) Chapters 00:00 – Introduction to AlphaFold 2 and activity overview 00:24 – CASP14 and breakthrough performance 01:39 – Accuracy metrics and why AlphaFold changed the field 02:02 – OpenFold, PyTorch vs JAX, and implementation context 03:27 – Core architecture: MSA + Evoformer explained 06:06 – Attention mechanisms and co-evolution signals 07:23 – Recycling: why multiple passes improve predictions 10:49 – When (not) to use MSAs and de novo proteins 13:32 – Limitations: point mutations and model sensitivity 19:22 – Sampling, dropout, and improving difficult predictions 21:17 – Training data and structural databases 26:02 – SE(3) invariance and coordinate independence 28:41 – Extensions: multimer, AlphaFold database, and variants 30:23 – Advanced workflows and parameter tuning 36:26 – Key parameters: recycling, MSA settings, model choices 42:46 – Activity walkthrough and experimentation setup