ML-PLA: Enhancing Protein–Ligand Binding Affinity Prediction with Microenvironment and Long-Range Interaction-Aware Graph Neural Networks

Accurately predicting protein-ligand binding affinity (PLA) is essential in drug discovery for identifying lead compounds. The sequence and structural contexts of an amino acid residue (i.e., microenvironment) describe the surrounding chemical properties and geometric features. While recent graph-based models have shown considerable promise, they often construct microenvironment representations using a shallow fusion of sequence and structural features, potentially failing to capture their full synergistic effects. Furthermore, the common reliance on a fixed distance threshold to define interaction space, while computationally efficient, inherently limits the ability to model key nonlocal biological phenomena. To address these issues, we propose a novel method named ML-PLA. Specifically, ML-PLA employs a heterogeneous graph neural network to model protein microenvironments by aggregating both sequence and structure information from neighboring nodes. Furthermore, we incorporate a vector quantized-variational autoencoder to capture the diversity and complexity of microenvironments, producing chemically meaningful, fine-grained representations. To effectively exploit long-range interaction information, ML-PLA projects protein-ligand complex atoms into multiple virtual atoms using a multihead attention mechanism, rather than simply increasing the number of graph neural network layers. This approach effectively embeds the interaction information into the complex atom features while simultaneously avoiding oversmoothing. Extensive experiments on the CASF-2016 and CASF-2013 benchmark data sets demonstrate the significant effectiveness and robust generalization capabilities of ML-PLA compared with state-of-the-art methods.