GraphKAN: Graph Kolmogorov Arnold Network for Small Molecule-Protein Interaction Predictions

Abstract

This study presents a proof of concept for utilizing Graph Kolmogorov Arnold Networks (GraphKAN/GKAN) in predicting the binding affinity of small molecules to protein targets. Working with three protein targets, we explored the potential of GraphKAN to infer binding affinities. We compared the performance of GraphKAN with MLP-based graph neural networks, 1D convolutional neural networks (1D CNN), and machine learning algorithms like random forests. Although the model did not achieve state-of-the-art performance, our results demonstrate its feasibility and highlight its promise as a novel approach in computational drug discovery. This work opens new research directions, suggesting that further refinement and exploration of GraphKAN could significantly impact the efficiency and accuracy of binding affinity predictions, ultimately aiding in the discovery of new therapeutic agents. Source code is available at - https://github.com/TashinAhmed/ferroin.