PepScorer::RMSD: An Improved Machine Learning Scoring Function for Protein-Peptide Docking

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Pharmaceutical peptides offer advantages over small molecules but require specialized computational tools.
• Existing molecular docking methods struggle with peptide flexibility and scoring, limiting their effectiveness in drug discovery.
• Current computational tools are primarily optimized for small molecules, necessitating adaptation for peptide-based drug candidates.

Purpose of the Study:

• To develop a novel machine learning-based scoring function, PepScorer::RMSD, for accurate peptide pose prediction in molecular docking.
• To enhance the docking power (DP) and pose selection capabilities for virtual screening of peptide libraries.
• To address the limitations of current scoring functions in handling the conformational flexibility of peptides.

Main Methods:

• Developed PepScorer::RMSD, a machine learning model predicting root-mean-squared deviation (RMSD) of peptide poses.
• Utilized a curated dataset of protein-peptide complexes (3-10 amino acids) for model training and evaluation.
• Benchmarked the PLANTS-based workflow, incorporating PepScorer::RMSD, against AlphaFold-Multimer predictions.

Main Results:

• PepScorer::RMSD achieved a Pearson correlation of 0.70 and a mean absolute error of 1.77 Å.
• Demonstrated high top-1 docking power (DP) of 92% on an evaluation set and 81% on an external test set.
• Outperformed conventional, ML-based, and existing peptide-specific scoring functions in accuracy and efficiency.

Conclusions:

• PepScorer::RMSD significantly improves the accuracy of peptide pose prediction and virtual screening.
• The developed tool and dataset provide a robust solution for computational peptide drug discovery.
• Freely available resources (PepScorer::RMSD and dataset) facilitate further research in peptide-based therapeutics.