A Constraint-Weighted Transfer Model Predicts Cross-System Generalization in Protein Fitness Landscapes

Abstract Transfer learning is widely used in protein fitness prediction but remains inconsistent: a model trained on one system may generalize well to some targets yet fail on others. The determinants of successful transfer remain poorly understood. Here, we introduce a constraint-weighted transfer model, a simple predictive framework that enables zero-shot transfer across protein systems without target-specific training. The model uses site-level sensitivity structure learned from a source system to define a weighting function over sequence variation in a target system. We systematically evaluate the model across 12 source–target transfer pairs spanning fluorescent proteins and β-lactamase systems using publicly available deep mutational scanning datasets. The model produces measurable predictive improvements over an ungated baseline in compatible systems and degrades in incompatible systems, revealing three reproducible regimes of transfer behavior. We show that transfer performance is strongly predicted by alignment of site-level sensitivity structure between source and target (Spearman r = 0.947), with additional contribution from overlap of the most sensitive sites (r = 0.752). This relationship holds across all evaluated systems, including independent validation in β-lactamase datasets. These results establish constraint alignment as a predictive criterion for cross-system generalization and position constraint-weighted transfer as a simple, interpretable baseline for zero-shot protein fitness prediction.