Optimization of protein-ligand kinetics can influence the efficacy of a drug <i>in vivo </i>[1] and decrease its toxicity [2]. Therefore, exploring structure-kinetic relationships both experimentally and computationally is an important task. However, there is a lack of high-throughput methods for kinetic parameters estimation as describing the transition state geometry is very challenging and computationally expensive.<br/><br/>In this work, we apply graph neural networks to the task of binding kinetics parameters prediction. Namely, we build graph neural networks based on graph convolutional (GCN), graph attention (GAT) and graph isomorphism (GIN) architectures to train structure-kinetic relationships in protein-ligand complexes. We employ KIND [3] database of experimental kinetic measurements and our literature-curated dataset to obtain a docking-based structural database [4] of over 3800 data samples. In addition, we discuss the applicability of binding affinity transfer learning and attention mechanism to derive and visualize the structural features that affect model prediction power.

Graph neural networks in protein-ligand binding kinetics prediction

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.