Michał Sapa, Alicja Gawalska, Marcin Kołaczkowski, Adam Bucki
The use of machine learning-based sequential virtual screening in the search of new ligands of 5-HT6 receptor
2023-04-05
5-HT6 receptor takes part in learning and memory processes. For this reason, the use of ligands of this receptor in the treatment of neurodegenerative diseases such as Alzheimer’s disease, depression, or autism is being investigated. The development of machine learning (ML) and access to large compound databases allow for the increasing use of these methods in search of new drugs. The use of ML in pre-clinical tests allows for a reduction in time and costs of drug discovery.
In this study, we used a sequential virtual screening approach in search of new structures with potential affinity for the 5-HT6 receptor. Data from the ChEMBL database containing ligand binding affinities, measured as an inhibition constant (Ki), was used as the training dataset. Each step of the screening was based on machine learning models, the task of which was to classify compounds as potentially active and inactive. The first step included a ligand-based drug discovery (LBDD) approach, in which, using Klekota-Roth fingerprints and molecular descriptors of the ligands, a classification model was developed to select a preliminary group of candidates from the Otava chemical compound database. In the second step, a structure-based drug discovery (SBDD) approach was used. For this purpose, compounds were docked to the AlphaFold database-derived model of the 5-HT6 receptor, previously optimized by the Induced-Fit Docking tool and molecular dynamics. Docking poses were scored by a trained Extra Trees classifier. Interactions of a reference ligand with 14 binding site residues were used as features for the trained model.
The use of machine learning as a scoring function allowed for improvement in the virtual screening parameters compared to the Glide GScore scoring function. Based on the obtained model, it was also confirmed that the location of a ligand near Ser5.43 and Phe5.38 residues is important for binding. The procedure allowed to select 20 candidates characterized by novel chemical structure and a relatively low basic pKa compared to known ligands, and thus suspected to have a low affinity for hERG channels and good brain penetration.
Keywords: molecular docking, machine learning, structure-activity relationship, serotonin 6 receptor.
© Farm Pol, 2022, 78(11): 607–614