Protein LM embeddings + similarity-based label augmentation improve pKa prediction
bravo_abad · x · 2026-10-27
Predicting amino acid protonation (pKa) matters for understanding enzyme activity, but experimental measurements are scarce. Shen et al. trained a small neural network on ESM protein-language-model embeddings to predict pKa directly from sequence.
The data-augmentation trick is worth borrowing: starting from an amino acid with a measured pKa, they search protein databases for other residues with highly similar embeddings and assign them the measured value as a synthetic label, assuming representation similarity implies similar biochemical environments. The RMSE chart shows GAINES augmentation at a high similarity threshold (0.8–1.0) clearly cuts error versus no augmentation or naive duplication.
More from Research
- Claimed proof would solve 3-SAT in sub-exponential time, breaking all NP-complete problems — IgorCarron · 2026-10-27
- Gowers: OpenAI's math drop includes Kakeya proof; AGMAI urges release of Yang-Mills outputs — calabi_and_yau · 2026-10-27
- One activated neuron makes Qwen3-8B roleplay a dying star — tszzl · 2026-10-27
- New eLife study makes whole-brain mean-field models more biologically realistic — AnnaCiaunica · 2026-10-27
- GenXR-Bench: AI coding agents fail at XR apps, best model scores just 50.9% — rohanpaul_ai · 2026-10-27
- EmbodiedLab Benchmarks Robots That Must Act to Know in Simulated Labs — HexaFuture · 2026-10-27