Biochemical-knowledge-driven machine learning pipeline for generating potent antimicrobial peptides.

Area of Science:

• Biochemistry
• Computational Biology
• Drug Discovery

Background:

• Antimicrobial resistance (AMR) is a growing global health threat.
• Novel antimicrobial peptides (AMPs) are needed as alternative therapeutics.
• Existing computational AMP discovery methods lack biological interpretability and translational value.

Purpose of the Study:

• To develop a biochemical-knowledge-driven pipeline (CVAE-BIO) for discovering AMPs targeting drug-resistant bacteria, specifically *Escherichia coli*.
• To integrate a conditional variational autoencoder (CVAE) with biochemical constraints and a Random Forest classifier.
• To generate and validate novel AMP candidates with high antimicrobial activity and low toxicity.

Main Methods:

• Utilized a conditional variational autoencoder (CVAE) constrained by biochemical properties (MIC≤10 μg/mL, net charge > +2, length < 40, instability index < 40, Boman index < 0).
• Integrated the CVAE with a Random Forest classifier trained on 30 biochemical descriptors.
• Performed *in vitro* validation of generated peptides for antimicrobial activity and hemolytic toxicity.

Main Results:

• 18.5% of generated peptides showed strong activity (MIC≤10 μg/mL), and 38.9% reached MIC ≤50 μg/mL.
• Validated novel peptides demonstrated narrow-spectrum activity, primarily targeting *E. coli*.
• Identified 9 non-toxic and active AMP candidates, with insights into residue composition correlating with activity.

Conclusions:

• The CVAE-BIO framework enables efficient AMP discovery under specific biochemical constraints.
• The study yielded experimentally validated AMP candidates with significant translational potential.
• Findings suggest residue composition (low counts of tiny/small residues) can enhance AMP activity.