GPCRVS in the detection of target-selective molecular scaffolds based on deep learning and gradient boosting

Motivation High drug selectivity minimizes unintended interactions that cause adverse effects, decreases drug potency, and are failure-prone factors in many clinical trials due to a poor balance between drug efficacy and toxicity. Sharing similar structural folds and modes of activation, G protein-coupled receptors may represent difficult drug targets in terms of optimization of pharmacodynamic profiles towards drug target selectivity. Here, we describe GPCRVS, a web service for GPCR virtual screening, designed to detect selective drug-like compounds. GPCRVS combines machine learning for ligand-based approaches using deep learning and gradient boosting, together with multi-target, structure-based virtual screening. Results GPCRVS performed well in binary classification (GPCR ligand vs non-GPCR ligand) with an F1 score of 0.875 and in multiclass classification for predicting drug target and drug target class with precision of 0.421 and 0.914, respectively, demonstrating its ability to detect active compounds with improved pharmacological characteristics. For receptors with sparse ChEMBL datasets such as active-state CCRL2, CXCR5, CCR6, CCR7, CCR8, CCR9, CCR10, and inactive-state CCR7, mixed training datasets based on receptor sequence homology were used for the ML model training with resulting accuracy of 85.5% obtained, e.g., for the CCR8-active ML model. In addition, GPCRVS detected target-selective, distinct molecular scaffolds with the LightGBM module such as fluorinated fragments in CCR6 inhibitors, or nitrogen-rich heterocyclic fragments in CXCR3 inhibitors. Availability and implementation GPCRVS is available on: https://gpcrvs.chem.uw.edu.pl.