RAVEN: Frozen Random Graph Reservoirs with Physics-Informed Interaction Fingerprints for Protein–Ligand Binding Affinity Prediction

Abstract Predicting protein–ligand binding affinity directly from three-dimensional complex structures remains challenging because available structure–affinity data are limited, heterogeneous, and highly susceptible to similarity-driven data set bias. Most graph-learning approaches address this problem by optimizing molecular representations end-to-end against affinity labels, potentially coupling representation learning to data set-specific correlations. Here, we introduce RAVEN (Randomized Atomistic Views with Ensemble Neural Reservoirs), a different learning paradigm in which multiple independently initialized atomistic graph neural networks are used as fully frozen reservoirs of nonlinear structural representations. Rather than adapting the graph encoders to affinity labels, RAVEN samples multiple independently initialized structural views from the initialization-induced representation space and integrates them with deterministic, physics-informed interaction fingerprints. Heterogeneous neural and tree-based readers then exploit these representations, followed by validation-based nonnegative fusion. On a similarity-isolated PDBbind 2020R1 benchmark reconstructed using GEMS similarity criteria, RAVEN achieved a Pearson correlation of 0.8005 and an RMSE of 1.2346 pK units on the held-out test set, and a Pearson correlation of 0.8415 on the protected CASF-2016 subset. Ablation analyses further show that increasing the number of frozen graph views generally improves the predictive utility of the structure-only representation, whereas physics-informed tree-based readers exhibit distinct residual patterns that enhance heterogeneous model fusion. These results demonstrate that competitive protein–ligand affinity prediction does not require end-to-end optimization of every representation component and establish frozen randomized graph reservoirs as a modular strategy for combining structural diversity, physical priors, and complementary inductive biases in molecular prediction.