AI Biology Breakthrough: ESMFold2 & Protein World Models
Key Questions
What has BioHub released in the ESMFold2 announcement?
BioHub (Zuckerberg) has released ESMFold2 and ESMC under an MIT license, along with a 6.8B protein atlas. These tools apply large-scale transformer models to protein science without specialized inductive biases.
How does ESMFold2 perform compared to specialized models like AlphaFold3?
Vanilla BERT-like transformers in ESMFold2 outperform specialized models such as AlphaFold3 on antibody interactions by scaling data and compute. Inference-time scaling has also shown effectiveness on five cancer and immunology targets.
What are the potential implications of this AI biology release?
The open release signals major progress in AI-driven biology and open science. It could accelerate drug discovery and protein engineering by demonstrating that general scaling approaches can surpass domain-specific designs.
BioHub (Zuckerberg) releases ESMFold2 and ESMC under MIT license. The 'Bitter Lesson' applied to proteins—vanilla BERT-like transformers beat specialized models like AlphaFold3 on antibody interactions by scaling data and compute, not inductive biases. Inference-time scaling works for five cancer/immunology targets. Also releasing 6.8B protein atlas. This is a major signal for AI-driven biology and open science, potentially accelerating drug discovery and protein engineering.