Deep learning in GPCR drug discovery: benchmarking the path to accurate peptide binding

Area of Science:

• Computational Biology
• Structural Biology
• Pharmacology

Background:

• Deep learning (DL) has revolutionized structure-based drug discovery by predicting protein structures from sequences.
• Recent DL models show increased accuracy in predicting multi-chain protein complexes.

Purpose of the Study:

• To evaluate DL tools for predicting interactions between G protein-coupled receptors (GPCRs) and their peptide ligands.
• To benchmark AlphaFold 2.3 (AF2), AlphaFold 3 (AF3), and other DL models on GPCR-peptide binding prediction.

Main Methods:

• Benchmarking of DL tools including AF2, AF3, Chai-1, NeuralPLexer, RoseTTAFold-AllAtom, Peptriever, ESMFold, and D-SCRIPT.
• Evaluation of peptide binding classification, ligand discovery, and binding pose accuracy.
• Analysis of confidence scores for predicting structural accuracy.

Main Results:

• Structure-aware DL models outperformed language models in peptide binding classification (AUC 0.86).
• AF2 demonstrated superior binding pose prediction (94% accuracy) compared to AF3 and Chai-1.
• Confidence scores correlate with binding mode accuracy, guiding interface prediction interpretation.

Conclusions:

• DL models reliably rediscover peptide binders and can aid peptide drug discovery for GPCRs.
• The study provides a guide for selecting optimal DL tools for GPCR-targeted therapies.
• An independent benchmarking set is provided for future model evaluations.