Abstract: A quantitative understanding of protein-ligand binding kinetics, particularly the dissociation rate constant (k_off), is fundamental to characterizing the dynamics of molecular recognition and optimizing in vivo pharmacodynamics. However, predicting k_off from molecular data remains a central challenge due to complex unbinding transition states and persistent data scarcity. Here, we confront these challenges by first curating KinetX, the largest comprehensive benchmark dataset of experimental k_off values to date, comprising 5624 measurements across 256 distinct protein targets. We then introduce BiCoA-Net, an interpretable deep learning framework designed to model reciprocal physicochemical interactions from sequence representations. The model employs a bidirectional co-attention mechanism and an adaptive gated fusion layer to learn these complex, non-linear interactions. BiCoA-Net achieves state-of-the-art accuracy and shows strong generalization in cold-start scenarios, where protein targets are novel. Furthermore, the model’s interpretability features identify key physicochemical descriptors and reveal bimodal interaction strategies. Validation on four held-out, clinically relevant targets confirms its practical utility, demonstrating robust compound ranking and significant enrichment in virtual screening tasks. BiCoA-Net and KinetX provide a robust, generalizable, and interpretable framework for the computational prediction and optimization of drug-target binding kinetics.

Key words: Protein-ligand binding kinetics, Dissociation rate constant, Deep learning, Bidirectional co-attention networks, Interpretable AI, Rational drug design