AI-Based Approach To Screen Opioid Receptor Ligands Via Grind Fingerprints and Convolutional Neural Network

Abstract

Opioid receptors belong to G-Protein Coupled Receptor (GPCR) family, and are divided into 3 subtypes of Mu, Kappa and Delta. The receptors are targeted by agonists and antagonists to manage variety of medical issues. Although main endogenous ligands for opioid receptors are peptides, a diverse set of non-peptide molecules can also bind orthosteric and allosteric sites of the receptor. Virtual screening for active agonists and antagonists using structure-based molecular docking is challenged by receptor flexibility, multimerization and orthosteric/allosteric sites coordination. Meanwhile, ligand-based virtual screening is more reliable. In this research grid-based alignment-independent molecular fingerprint (GRIND) and convolutional neural network (CNN) models are used to classify database of potential opioid ligands of experimentally measured activities. Six GRIND-CNN models were trained to recognize features within the fingerprint that are relevant to agonistic and antagonistic activities against each of Mu, Kappa and Delta receptors. The models performance measured by Area Under the Curve of Receiver Operating Characteristic (AUC-ROC), Matthews Correlation Coefficient (MCC) and Balanced Accuracy (BA). The GRIND-CNN models show mean AUC-ROC values (from 0.70 to 0.83) which are comparable to PubChem and Extended-connectivity fingerprints used for deep neural network (DNN) models (PubChem-DNN and ECFP-DNN, respectively. However, both of GRIND-CNN and PubChem-DNN models outperform ECFP-DNN model in recognition of activity-relevant features within relevant fingerprint (mean MCC values for Y-randomized training across all ligands groups are 0.05, 0.03 and 0.16, respectively). By using combined predictions of both GRIND-CNN and PubChem-DNN models with 0.5 coefficient for each, the mean AUC-ROC is improved (0.77 to 0.91) and MCC for Y-randomized training is reduced to 0.02, which indicates improved feature recognition. Saliency maps of GRIND fingerprints for models trained to differentiate agonists vs antagonists, and differentiate ligands vs non-ligands for different opioid receptor subtypes showed statically significant differences at correlogram bins (pixels) level. Such differences are related to message-address theory of opioid ligands. The research provides tools for virtual screening of opioid ligands that follow activity-relevant features recognition irrespective to molecular scaffold or size, in a manner similar to how a hand senses through distinct fingers the geometry of an object. The models are freely available on https://github.com/mohammednooraldeen/GRIND-CNN.