Improving Docking Power for Short Peptides Using Random Forest.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Bioinformatics

Background:

• Therapeutic peptides are increasingly important in medicine, with many approved drugs and more in clinical trials.
• Current peptide docking methods struggle due to scoring functions optimized for small molecules.
• This limitation hinders rational drug design for peptide-based therapeutics.

Purpose of the Study:

• To develop improved computational methods for accurate peptide docking.
• To enhance the success rate of peptide docking in drug design.
• To create a high-quality dataset for training and benchmarking peptide docking models.

Main Methods:

• Compiled the ProptPep37_2021 dataset of 322 crystallographic protein-peptide complexes.
• Developed and trained random forest classifiers to discriminate correctly docked peptide poses.
• Utilized a benchmark testing set of 47 protein-peptide complexes to evaluate classifier performance.

Main Results:

• Random forest classifiers significantly improved peptide docking success rates from ~25% to an average of ~70%.
• The developed classifiers demonstrated high accuracy on a diverse testing set.
• The ProptPep37_2021 dataset provides valuable resources for future peptide docking research.

Conclusions:

• The new random forest classifiers offer a substantial improvement for peptide docking accuracy.
• These methods facilitate peptide-based drug design, bringing it closer to small molecule docking success rates.
• The freely available dataset and classifiers will advance the field of computational peptide drug discovery.