Physics informed machine learning predicts binding free energy changes in protein-RNA complexes upon mutations

Protein-RNA interactions are central to gene regulation, yet predicting mutation-induced changes in binding free energy (ΔΔG) remains difficult due to sparse, heterogeneous measurements and strong imbalance in mutation types. Here we present Pred-MutPRI, a physics-informed machine learning framework for ΔΔG prediction in protein–RNA complexes. We curated a unified dataset of experimentally measured ΔΔG values with standardized thermodynamic conventions, and introduced thermodynamic permutation (TP) to generate cycle-consistent training pairs that expand substitution-type coverage while reducing mutation-class bias. Pred-MutPRI integrates conventional structure-derived descriptors with weighted atom-level interaction networks that explicitly encode the local binding microenvironment. In addition, it incorporates two compact, information-dense features: a masked ESM-2 entropy term capturing context-dependent sequence constraint, and an AlphaFold3-derived local effective strain descriptor summarizing mutation-associated structural deformation across multiple mutant-structure predictions. With an XGBoost regressor, Pred-MutPRI achieves strong generalization on a sequence-disjoint and structurally low-overlap blind test set (PCC = 0.705) and outperforms existing predictors. The dataset and Python package are available at https://github.com/CCNU-COMPBIO/Pred-MutPRI. Pred-MutPRI is a machine learning method that predicts how mutations affect protein-RNA binding affinity. By combining thermodynamics, structural features, sequence information, and AlphaFold3-derived strain measures, it outperforms existing predictors and achieves strong accuracy on independent test sets (PCC = 0.705).