A General Framework for Injecting Biophysical Priors into Protein Embeddings

Accurate ΔΔ G prediction requires integrating machine learning with biophysical insight. Existing approaches typically prioritize one while neglecting the other. We introduce an encoder-agnostic framework that injects interpretable biophysical priors into residue-level deep learning representations via cross-embedding attention. ProtBFF consistently improves performance under homology-based-clustering evaluation and enables general-purpose encoders to surpass state-of-the-art specialized models or larger models. Our results show that integrating simple, mechanistic priors into pretrained representations yields more trust-worthy predictors, offering a practical solution for broader protein engineering applications. Code github.com/Jfeldman34/ProtBFF