De Novo Antimicrobial Peptide Design with Feedback Generative Adversarial Networks

Area of Science:

• Computational chemistry and drug discovery
• Bioinformatics and machine learning applications in life sciences

Background:

• Antimicrobial peptides (AMPs) show potential as novel antibiotics due to broad-spectrum activity and low resistance development.
• Deep learning, particularly generative models, can accelerate AMP discovery and optimization.
• Feedback Generative Adversarial Network (FBGAN) is a deep generative model incorporating a classifier for training.

Purpose of the Study:

• To investigate the impact of enhanced classifiers on the generative performance of the FBGAN framework.
• To develop improved FBGAN models for de novo design of antimicrobial peptides.

Main Methods:

• Introduced two novel classifiers for the FBGAN framework: one based on k-mers and another using transfer learning from the ESM2 protein language model.
• Integrated these enhanced classifiers into the FBGAN architecture.
• Evaluated the generative performance of the modified FBGAN models against the original FBGAN and other established methods (AMPGAN, HydrAMP).

Main Results:

• Both introduced classifiers demonstrated higher accuracy than the original FBGAN classifier.
• The enhanced FBGAN models showed significant performance improvements over the original FBGAN.
• The proposed models achieved performance comparable or superior to existing AMP design tools like AMPGAN and HydrAMP.

Conclusions:

• Leveraging advanced classifiers within the FBGAN framework significantly enhances its computational robustness for antimicrobial peptide de novo design.
• The improved FBGAN models offer a competitive and effective approach for discovering novel antimicrobial peptides, comparable to current state-of-the-art methods.