Topological deep learning for enhancing peptide-protein complex prediction.

Area of Science:

• Computational biology
• Structural biology
• Drug discovery

Background:

• Peptide-protein interactions are crucial for biological processes and drug development.
• Selecting accurate models of peptide-protein complexes is difficult due to high false positive rates.
• Existing methods lack sufficient precision in evaluating interface quality.

Purpose of the Study:

• To develop a novel deep learning model, TopoDockQ, for accurate prediction of peptide-protein interface quality.
• To enhance the precision of model selection and mitigate false positives in peptide-protein complex prediction.
• To introduce ResidueX, a workflow for incorporating non-canonical amino acids into peptide designs.

Main Methods:

• Developed TopoDockQ, a topological deep learning model utilizing persistent combinatorial Laplacian (PCL) features.
• Predicted DockQ scores (p-DockQ) to evaluate peptide-protein interface quality.
• Integrated ResidueX workflow for non-canonical amino acid (ncAA) incorporation into peptide scaffolds.

Main Results:

• TopoDockQ reduced false positives by at least 42% compared to AlphaFold2's confidence score.
• Achieved a 6.7% increase in precision across five evaluation datasets.
• Maintained high recall and F1 scores while significantly improving accuracy.

Conclusions:

• TopoDockQ offers a more precise method for evaluating peptide-protein complex models, reducing false positives.
• ResidueX enables flexible peptide design with non-canonical amino acids.
• These advancements accelerate the development of next-generation peptide therapeutics through improved modeling and design capabilities.