Accurate RNA-Ligand Binding Site Prediction Based on a Multi-Channel Graph Neural Network

RNA-ligand binding-site prediction is a challenging task in RNA molecular analysis. Binding regions are often sparse, structurally heterogeneous, and difficult to delineate accurately at the nucleotide level. Existing sequence-based methods lack explicit structural modeling, while conventional graph neural networks tend to mix signals around binding/non-binding transition regions. In this paper, BC-GNN, a multi-channel graph neural network for nucleotide-level RNA-ligand binding-site prediction, is proposed. BC-GNN integrates sequence-informed auxiliary transition estimation, boundary-aware propagation (BAP), microenvironment-aware channel recalibration (MACR), and hierarchical multi-scale integration (HMSI) to improve structural representation learning. When evaluated on a benchmark derived from RNAmigos2 using the official leakage-controlled 0.75 split, BC-GNN achieves an AUC of 0.8280, an F1-score of 0.6086, and an MCC of 0.4511, outperforming multiple re-evaluated baselines under the same rigorous protocol. These results demonstrate that BC-GNN is effective for RNA-ligand binding-site prediction.