bioRxivJClub
Decoding the physicochemical basis of taxonomy preferences in protein design models
bioRxiv · · Preprint · Open access
Dillon, Crook + 1 more
Abstract ↗AI summary
The abstract is read at the publisher; the summary is JClub's.
Backbone-conditioned protein design scores remove 91–93% of apparent species-level variation after controlling for family and measured biophysics, versus only 15–29% for sequence-only likelihoods.
- Why it matters: Model scores are increasingly used to prioritize candidates, but the determinants of taxonomic preferences and their biophysical origins are not well understood.
- What they did: The study profiles 14 protein design models grouped by scoring context, using controls for protein family and measured biophysical properties, and compares backbone-conditioned scores to sequence-only likelihoods.
- The result: Backbone-conditioned scores favor archaeal and thermophilic proteins and reflect compactness, packing, and charge; further training ProteinMPNN on extremophile secretomes alters surface chemistry with minimal structural cost while largely preserving species rankings.
The findingWhy it mattersWhat they didThe result
- Open access