Machine Learning-based prediction of protein-ligand binding affinity

Machine Learning-based prediction of protein-ligand binding affinity F. Tanda1, R. Beccaria2,3,4, M. Civera1, G. Tiana2,5 1Università degli studi di Milano, Dipartimento di Chimica, via Golgi 19, 20133, Milano, Italia. 2Università degli studi di Milano, Dipartimento di Fisica, via Celoria 16, 20133, Milano, Italia. 3Heidelberg Institute for Theoretical Studies, Schloss-Wolfsbrunnenweg 35, 69118, Heidelberg, Germania 4Faculty of Physics, Heidelberg University, Im Neuenheimer Feld 227, 69120 Heidelberg, Germania 5INFN, via Celoria 16, 20133, Milano, Italia Accurately predicting ligand-protein binding affinity is crucial in computational drug discovery, as it enables the efficient identification and optimization of therapeutic compounds. Machine learning can capture complex molecular interaction patterns beyond traditional methods, but existing models are often limited by dataset quality and molecular representation.1,2 We propose a regression-based tool for ligand ranking that exploits an equivariant molecular representation.3 A curated dataset, which includes values of Kd, Ki, and IC50 for each ligand-protein complex, was used to train multiple models. The algorithm with the highest R² score during validation provides a reliable predictor of binding affinity and offers a valuable tool for structure-based drug design. Bibliography: [1] Thomas Gaillard, Evaluation of AutoDock and AutoDock Vina on the CASF-2013 Benchmark, Journal of Chemical Information and Modeling 2018 58 (8), 1697-1706 [2] Huiwen Wang, Prediction of protein–ligand binding affinity via deep learning models, Briefings in Bioinformatics, Volume 25, Issue 2, March 2024 [3] R. Beccaria, A. Lazzeri, and G. Tiana, Journal of Chemical Information and Modeling 2024 64 (17), 6758-6767