Learning Protein Fitness Landscapes with Multimodal Stability Priors

Predicting how mutations change protein fitness is central to protein engineering and variant interpretation, yet most experimentally measured fitness landscapes contain only limited labeled variants. We present PsiFit, a stability-informed framework that adapts protein language models for low-$N$ fitness prediction by injecting mutation-induced stability changes predicted by a multimodal sequence-structure foundation model. By integrating biophysical stability priors into contrastive fine-tuning, PsiFit aims to improve data efficiency, reduce overfitting, and provide a general strategy for learning protein fitness landscapes from sparse assays.