2021/11/30 by Carter Knutson, Knutson, Carter, Mridula Bontha +5
Biochemistry, Genetics and Molecular Biology · Computer Science · Materials Science · #Biomolecules (q-bio.BM) #Computational Drug Discovery Methods #FOS: Biological sciences #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Machine Learning in Materials Science #Protein Structure and Dynamics #Quantitative Methods (q-bio.QM)
paper · pdf · doi:10.48550/arxiv.2111.15144
openalex publication_date 2021/11/30 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Protein-ligand interactions (PLIs) are fundamental to biochemical research\nand their identification is crucial for estimating biophysical and biochemical\nproperties for rational therapeutic design. Currently, experimental\ncharacterization of these properties is the most accurate method, however, this\nis very time-consuming and labor-intensive. A number of computational methods\nhave been developed in this context but most of the existing PLI prediction\nheavily depends on 2D protein sequence data. Here, we present a novel parallel\ngraph neural network (GNN) to integrate knowledge representation and reasoning\nfor PLI prediction to perform deep learning guided by expert knowledge and\ninformed by 3D structural data. We develop two distinct GNN architectures, GNNF\nis the base implementation that employs distinct featurization to enhance\ndomain-awareness, while GNNP is a novel implementation that can predict with no\nprior knowledge of the intermolecular interactions. The comprehensive\nevaluation demonstrated that GNN can successfully capture the binary\ninteractions between ligand and proteins 3D structure with 0.979 test accuracy\nfor GNNF and 0.958 for GNNP for predicting activity of a protein-ligand\ncomplex. These models are further adapted for regression tasks to predict\nexperimental binding affinities and pIC50 is crucial for drugs potency and\nefficacy. We achieve a Pearson correlation coefficient of 0.66 and 0.65 on\nexperimental affinity and 0.50 and 0.51 on pIC50 with GNNF and GNNP,\nrespectively, outperforming similar 2D sequence-based models. Our method can\nserve as an interpretable and explainable artificial intelligence (AI) tool for\npredicted activity, potency, and biophysical properties of lead candidates. To\nthis end, we show the utility of GNNP on SARS-Cov-2 protein targets by\nscreening a large compound library and comparing our prediction with the\nexperimentally measured data.\n