#30 · Structural bio & protein design

Curated resources and sequence models probe protein function and predictor limits

The common thread is sequence-based prediction applied at scale to protein classes with limited experimental annotation. AFP-R pairs AFP-DB, a curated set of experimentally validated antifreeze proteins, with an ESM2-based sequence predictor. The thermodynamic framework maps elastic recoil mechanisms across the human proteome from sequence, extended with protein language model embeddings, while RegimeFormer and its RegimeAtlas index 202M protein sequences to predict mutation-response regimes. A distinct diagnostic result reports that AlphaFold2/3 and PUNCH2 show pLDDT-disorder discordance on de novo and random proteins, driven by a compositionally identifiable subset. This is the most specific new finding: pLDDT-disorder discordance on orphan proteins is linked to a compositionally identifiable subset. The thread is not uniform across items, since the discordance study evaluates existing predictors rather than proposing a resource.