Implicit modeling of the conformational landscape and sequence allows scoring and generation of stable proteins

Generative protein modeling provides advanced tools for designing diverse protein sequences and structures. However, accurately modeling the conformational landscape and designing sequences--ensuring that the designed sequence folds into the target structure as its most stable structure--remains a critical challenge. In this study, we present a systematic analysis of jointly optimizing P(structure|sequence) and P(sequence|structure), which enables us to find optimal solutions for modeling the conformational landscape. We support this approach with experimental evidence that joint optimization is superior for (1) designing stable proteins using a joint model (TrROS (TrRosetta) and TrMRF) (2) achieving high accuracy in stability prediction when jointly modeling (half-masked ESMFold pLDDT+ ESM2 Pseudo-likelihood). We further investigate features of sequences generated from the joint model and find that they exhibit higher frequencies of hydrophilic interactions, which may help maintain both secondary structure registry and pairing.