AlphaFold3 MSA retraining probes whether it learns inverse covariance structure

anshulkundaje · x · 2026-07-22

Researchers retrained the MSA module of AlphaFold3 from scratch on an unsupervised objective to probe what it is learning.

• The goal is to test the claim that AlphaFold3 learns to invert the covariance matrix.
• The post says they essentially retrained just the MSA part and used the setup to see whether the model can recover unsupervised contact prediction behavior.
• The attached diagram shows the AlphaFold3-style MSA and pair representations, triangle updates, and the final unsupervised contact prediction head.

Original post →

More from Research

• Datology webinar: Curation Studio matched top data baselines with under 20% compute — josh_wills · 2026-10-27
• KANLab: an interactive browser playground for Kolmogorov-Arnold Networks — Lukas Niehaus · 2026-10-27
• CMU's LLM2Cert synthesizes stability certificates verified via SMT and Lean — CarnegieMellonU · 2026-10-27
• NVIDIA open-sources FLORA, a robotics data engine ~125x cheaper than GPT-6 Astra — nvidia · 2026-10-27
• RSNA 2026 Deep Learning Labs return: eight hands-on AI classes over four days in Chicago — PRouzrokh · 2026-10-27
• Steering two orthogonal concept neurons makes a model compose "gnome" — Sauers_ · 2026-10-27

Research channel →