Protein Language Model Learns Dynamics from Missing NMR Peaks
bravo_abad · x · 2026-08-12
Protein dynamics research has long been limited by a lack of standardized data. Hannah Wayment-Steele and coauthors propose a clever method: using missing labels in existing NMR datasets as proxy data.
Microsecond-to-millisecond conformational exchange can broaden NMR peaks until they disappear. The team curated 9,381 proteins from the BMRB to train models predicting which residues are missing. The strongest model, Dyna-1, utilizes an intermediate representation from the multimodal protein language model ESM-3.
More from Research
- u-OPSD: Self-Distillation Without Labels or Teachers Beats GRPO on Math Reasoning — burny_tech · 2026-08-12
- Survey of 150+ Agent Memory Architectures: Self-Evolving Designs Boost Retention by 50% — blaizedsouza · 2026-08-12
- Stanford Paper Reveals Multi-Agent Flaws, Introduces Control Plane for 10x Efficiency — blaizedsouza · 2026-08-12
- New Book Reveals Imbalanced Data Truths: Data and Models Matter More Than Balancing — Al_Grigor · 2026-08-12
- Claude Solves FrontierMath Open Problem: Finds Hadamard Matrix of Order 668 — inductionheads · 2026-08-12
- AskScience AMA: Expert Discusses Transparency and Safety of AI Chain-of-Thought Reasoning — umd-science · 2026-08-12