DeepPANB: integrating protein language model with PaiNN equivariant graph neural networks for prediction of protein-nucleic acid binding sites.

Protein-nucleic acid interactions are fundamental to various biological processes, and accurately identifying nucleic acid binding sites on proteins is essential for understanding gene regulation mechanisms and advancing drug design. Here, we present DeepPANB, an effective deep learning model for the issue. DeepPANB is built upon the Polarizable Atom Interaction Neural Network (PaiNN) architecture, an E(3)-equivariant graph neural network, which explicitly couples scalar and vector channels through lightweight message passing, thereby enabling effective modeling of the direction- and distance-dependent residue interactions. DeepPANB integrates multiple feature types including sequence embeddings from the Ankh pretrained model, residue-nucleotide pairwise propensity extracted by us from the large-scale datasets, as well as residue intrinsic disorder, physicochemical properties, and structural descriptors. To our best knowledge, PaiNN and Ankh are first introduced here for protein-nucleic acid binding site prediction. On the protein-DNA test set, DeepPANB achieves state-of-the-art performance, outperforming the existing methods. On protein-RNA test set, DeepPANB, fine-tuned with the feature types unchanged, shows competitive performance. Overall, DeepPANB demonstrates robust and generalizable performance, providing an effective tool for protein-nucleic acid binding site prediction. Freely available at https://github.com/ChunhuaLab/DeepPANB.