Decoding the physicochemical basis of taxonomy preferences in protein design models

Recent improvements in machine-learning protein design have enabled models to generate more promising candidate sequences than can be experimentally tested, placing increasing reliance on scores whose determinants remain poorly understood. Here, we profile 14 models grouped by scoring context to disentangle taxonomic and biophysical contributions. Controlling for protein family and measured biophysical properties removes 91 to 93% of apparent species-level variation in backbone-conditioned scores, compared with only 15 to 29% in sequence-only likelihoods. Species rankings also diverge: backbone-conditioned models favour archaeal and thermophilic proteins, whereas sequence-only models favour eukaryotic proteins. Backbone-conditioned scores instead reflect compactness, packing and charge, with these biases propagating into redesigned sequences. Further training ProteinMPNN on extremophile secretomes alters designed surface chemistry with minimal structural cost but largely preserves species rankings. Together, these findings show how scoring context and training data shape what protein design models favour, with implications for which sequences are generated and prioritised for experimental testing.