ProteinMPNN memorizes transmembrane folds: noise may trick it, devs speculate
sokrypton · x · 2026-11-05
- sokrypton and @DdelAlamo discuss ProteinMPNN's inner workings: they hypothesized it memorized which protein folds are transmembrane, which is why they had to retrain to exclude transmembrane proteins.
- At 0.3 noise, they guess the model no longer recognizes the fold as transmembrane.
- On implementation: either a simple bias adjustment per amino acid, or averaging activations over transmembrane/soluble proteins and modifying weights to pin those activations on (or off).
Related event: ProteinMPNN May Memorize Transmembrane Folds; Noise Disables Recognition(2 posts)→
More from Research
- Engram blog: RL training complements text memory, stays robust through millions of token edits — canaesseth · 2026-11-05
- New Book on Generative AI and Stochastic Thermodynamics by Welling Team Out This Month — mjdramstead · 2026-11-05
- New paper uses 288M job postings to measure where AI raises the value of work it can't do — robseamans · 2026-11-05
- Matthew Honnibal: LLMs can 'see' weight matrices — but only one layer at a time — honnibal · 2026-11-05
- RicursiveAI claims 16x compute efficiency gain on code pretraining vs Olmo 3 baseline — samiramanabi · 2026-11-05
- Claude drifts from 🌀 to 💖: refined AI symbols become less legible to humans (67%→44%) — mimi10v3 · 2026-11-05