Leveraging Different Distance Functions to Predict Antiviral Peptides with Geometric Deep Learning from...

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Machine learning accelerates peptide-based drug discovery.
• Graph learning frameworks utilize peptide structure graphs.
• Current methods rely on Euclidean distance thresholds without strong evidence.

Purpose of the Study:

• Investigate diverse distance functions for peptide structure graph generation.
• Train deep graph learning models for antiviral peptide prediction using these graphs.

Main Methods:

• Analyze amino acid closeness using various distance functions.
• Compare derived graphs based on different distance metrics and random graphs.
• Train and evaluate deep graph learning models with optimal graph representations.

Main Results:

• Different distance functions create dissimilar graphs encoding distinct chemical spaces.
• These alternative graphs improve the discriminative power of predictive models.
• Performance comparisons with state-of-the-art models were conducted.

Conclusions:

• Euclidean distance is insufficient for comprehensive peptide structure graph representation.
• Employing varied distance functions yields superior graph structures.
• Optimized graph representations lead to enhanced antiviral peptide prediction models.