Structure-aware machine learning strategies for antimicrobial peptide discovery.

Area of Science:

• Biochemistry
• Computational Biology
• Peptide Science

Background:

• Machine learning models are increasingly used for bioactive peptide discovery and design.
• Current models often lack structural awareness, limiting comprehension of peptide mechanisms of action.
• Antimicrobial peptides (AMPs) exhibit diverse mechanisms, including membrane disruption, penetration, and protein binding.

Purpose of the Study:

• To investigate the mechanisms of action and structural landscape of antimicrobial peptides.
• To develop accurate machine learning models for classifying AMPs based on their function and structure.
• To address and overcome structural biases in existing predictive models.

Main Methods:

• Analysis of critical features like dipeptides and physicochemical descriptors.
• Development of initial machine learning models (1.0 and 2.0) for predicting peptide categories.
• Implementation of subset selection and data reduction strategies to mitigate structural bias.
• Creation of structure-specific and structure-agnostic predictive models.

Main Results:

• Initial models achieved high accuracy (86-88%) but showed bias towards specific structures (α-helical, coiled).
• Structure-specific models were developed for α-helical, coiled, and mixed structures.
• Structure-agnostic models were created by depleting over-represented structures.
• Identified sensitivity of key features to different peptide structure classes.

Conclusions:

• Machine learning models can effectively predict antimicrobial peptide mechanisms of action and structure.
• Addressing structural biases is crucial for developing robust and generalizable peptide prediction models.
• Feature importance varies across different peptide structure classes, offering insights into structure-activity relationships.