AlloGNN: Data and Pre-Trained Models for Allosteric Binding Site Prediction in Human Protein Kinases

This repository provides the data, pre-computed representations, and pre-trained model resources required to reproduce the experiments reported in the manuscript "ALLOGNN: A Multi-Modal Graph Neural Network with Residue-Level Gated Fusion for Allosteric Binding Site Prediction in Human Protein Kinases". AlloGNN is a multi-modal graph neural network designed for residue-level prediction of active and allosteric binding sites in human protein kinases. The framework integrates three complementary modalities: protein sequence representations extracted using the ESM2-650M protein language model, structural information represented as Cα residue contact graphs, and local surface/biophysical features including relative solvent-accessible surface area (SASA), crystallographic B-factor, and secondary structure. The release contains the pre-computed PyTorch Geometric graph dataset comprising 9,217 protein chains, pre-extracted ESM2 embeddings, model checkpoints, and the associated data required to reproduce the reported evaluation results. The dataset includes an unseen gene-cold test set containing 812 protein chains. The release is intended to facilitate reproducibility and independent evaluation of the AlloGNN framework for allosteric binding site prediction in human protein kinases. Code and training/evaluation scripts are available in the associated GitHub repository:https://github.com/srajalcodes/AlloGNN