Generative models for antimicrobial peptide design: auto-encoders and beyond.

Area of Science:

• Computational biology
• Drug discovery
• Bioinformatics

Background:

• Rising threat of multi-resistant pathogens necessitates novel antimicrobial drug development.
• Antimicrobial peptides (AMPs) offer a promising alternative due to lower resistance induction.
• Generative deep learning accelerates in silico mining for potential AMP candidates.

Purpose of the Study:

• To comparatively analyze generative deep learning models for novel antimicrobial peptide synthesis.
• To evaluate Variational Auto-Encoders, Wasserstein Auto-Encoder, Recurrent Neural Network, and Language Model performance.
• To identify optimal model-sampling strategy combinations for diverse antimicrobial peptide design objectives.

Main Methods:

• Comparative analysis of generative deep learning models: Variational Auto-Encoders, Wasserstein Auto-Encoder, Recurrent Neural Network, Language Model.
• Systematic evaluation of model generative performance and sampling strategies.
• Assessment of generated peptide physicochemical properties and diversity.

Main Results:

• All models produced peptides with physicochemical profiles akin to natural AMPs.
• Auto-encoders, particularly the Wasserstein variant, excelled in generating diverse and balanced peptides.
• Embedding-space analysis confirmed auto-encoders avoided overfitting; model-specific preferences were noted in AMP predictor evaluations.

Conclusions:

• Generative models exhibit distinct strengths and weaknesses for antimicrobial peptide generation.
• Tailoring model selection and evaluation metrics to specific design goals is crucial.
• Practical guidance is provided on combining model types and sampling strategies for targeted antimicrobial peptide discovery.