Quantum and Classical Graph Convolutional Neural Networks for Protein Ligand Dissociation Constant Prediction

Abstract

How long a drug stays bound to its target - the residence time - is now recognized as a stronger in vivo efficacy driver than binding affinity alone. Yet current machine learning (ML) models for dissociation kinetics (koff) ignore two critical sources of structure information: (1) the change in protein-ligand geometry across time, and (2) the ability to represent complex spatial interactions with fewer parameters. We extend a state-of -the -art spatial GNN for protein-ligand complexes with two innovations. First, we introduce a 2-timestep GCN+GRU model that learns structure changes before and after short molecular dynamics simulations. Second, we compress the model head using variational quantum circuits, preserving expressivity while removing 66% of parameters. On the full PDBbind-koff-2020 benchmark, temporal integration improves predictive accuracy, and quantum compression matches the full classical models accuracy at a fraction of its size. Our results show that time and quantum structure are underexplored, high-leverage axes for advancing kinetic ML models used in drug design.