Major Advances in Biomolecular AI: Megascale Dataset and Open Protein Models
Key Questions
What is the MGnify Stability Dataset released by Rocklin Lab?
The dataset contains 1.8 million protein stability measurements, including negative data, and was developed with support from the OpenFold Consortium. It focuses on folding stability to support biomolecular AI research and drug discovery.
How does this dataset improve machine learning models for proteins?
By providing large-scale data with negative examples, it enables better training of ML models. Demonstrations with SaProtΔG and ESM3ΔG show effective real-world transfer performance.
Why is this release significant for biomolecular AI?
It represents a major open resource that addresses data limitations in protein stability prediction. This advancement supports improved model development for applications in drug discovery and related fields.
Rocklin Lab released 1.8M protein stability measurements (Megascale) with negative data, enabling better ML models (SaProtΔG, ESM3ΔG). CZ Biohub released open protein models ESMC and ESMFold2, trained on billions of sequences. These resources are major contributions to biomolecular AI and drug discovery.