A General Framework for Injecting Biophysical Priors into Protein Embeddings 1. New preprint introduces ProtBFF, an encoder-agnostic framework that enhances protein embedding models with interpretable biophysical features through cross-embedding attention. 2. The method addresses a critical flaw in SKEMPI2 benchmarking: systematic data leakage from sequence redundancy. When evaluated with proper homology-based clustering at 60% identity, existing state-of-the-art models show dramatically reduced performance. 3. ProtBFF operates as a plug-in module that reweights residue-level embeddings using five biophysical scores: interface propensity, residue burial, dihedral deviation, solvent-accessible surface area (SASA), and local distance difference test (lDDT). 4. Key innovation: rather than building specialized architectures, the framework enriches general-purpose pretrained embeddings (ESM2, ESM3, ProSST) with physical priors, enabling them to surpass dedicated binding affinity predictors. 5. Ablation studies reveal interface and burial features contribute most to accuracy, while the auxiliary ilDDT loss improves structural fidelity and generalization. 6. The 150M-parameter ESM2 with ProtBFF outperforms even 15B-parameter variants without it, demonstrating that biophysical guidance can compensate for model scale. 7. On out-of-distribution SARS-CoV-2 antibody-antigen datasets, ProtBFF-enhanced models show strong few-shot learning capability, achieving high correlation with only 10% training data. 8. The approach highlights a broader principle: pretrained representations learn rich protein features, but selective amplification through mechanistic priors unlocks task-specific performance without sacrificing generalization. 💻Code: github.com/feldmanjonathan/P… 📜Paper: biorxiv.org/content/10.64898… #ProteinDesign #MachineLearning #Biophysics #ProteinProteinInteractions #DDGPrediction #ComputationalBiology #Bioinformatics #DeepLearning