A novel deep learning framework for predicting protein-ligand interaction fingerprints from sequence data: integrating graph inductive bias transformer with Kolmogorov-Arnold networks

Abstract

Accurately modeling protein–ligand interactions is a central challenge in computational protein design and drug discovery. Traditional interaction fingerprint (IFP) approaches, while valuable, struggle to capture subtle binding features and adapt to diverse structural contexts. To address these limitations, we propose GITK, a deep learning framework that integrates a modified graph inductive bias transformer (GRIT) with Kolmogorov–Arnold networks (KANs) for interpretable interaction fingerprint prediction. GRIT introduces inductive bias to effectively represent both local and global graph structures of proteins and ligands, while KAN provides a principled functional decomposition that enhances nonlinear feature learning and interpretability. Benchmarking across multiple datasets demonstrates that GITK outperforms state-of-the-art models in binding affinity prediction, functional effect classification, and virtual screening. Moreover, GITK enables reliable selectivity analysis, successfully highlighting conformational differences and key residues in adenosine receptor subtypes, consistent with experimental findings such as the selectivity of the A1 antagonist DPCPX.