GPCRLigNet: rapid screening for GPCR active ligands using machine learning.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Pharmacology

Background:

• G protein-coupled receptors (GPCRs) are crucial drug targets, but identifying bioactive small molecules is challenging.
• Vast chemical libraries require efficient methods for virtual screening to find GPCR-active compounds.

Purpose of the Study:

• To compare various machine learning models for binary classification of GPCR-active molecules.
• To develop and validate an effective computational tool for high-throughput virtual screening of GPCR ligands.

Main Methods:

• Trained and validated machine learning models using over 600,000 active, inactive, and decoy compounds.
• Compared dilated graph convolutional networks with other models, including feedforward dense neural networks.
• Integrated the best model, GPCRLigNet, into a virtual screening workflow with molecular docking against the pituitary adenylate cyclase-activating polypeptide receptor type 1.

Main Results:

• A simple feedforward dense neural network (GPCRLigNet) mapping Morgan fingerprints to activity outperformed complex models.
• GPCRLigNet successfully enriched the selection of potentially potent molecules when used in conjunction with molecular docking.
• Demonstrated an enrichment of potent molecules compared to traditional virtual screening methods.

Conclusions:

• GPCRLigNet provides a proof of principle for effectively narrowing the chemical search space for GPCR-active ligands.
• The study highlights the utility of machine learning in accelerating the discovery of small molecules targeting GPCRs.
• GPCRLigNet can significantly enhance the efficiency of drug discovery pipelines for GPCR-related therapeutics.