Deep Learning–Driven De Novo Design of Antimicrobial Peptides Active in Human Serum: A Systematic Review of Architectures, Serum Conditions, and Activity Retention

Abstract

Antimicrobial resistance (AMR) is a major global health problem, contributing to an estimated 5 million deaths annually, while current antibiotic development remains insufficient. Antimicrobial peptides (AMPs) are a potential alternative due to their membrane-disrupting mechanisms and lower likelihood of inducing resistance. However, their clinical application is limited by reduced activity in human serum, where proteins, salts, and proteases can bind to or degrade peptides. Recent advances in deep learning (DL) have enabled de novo design of AMPs using models such as generative adversarial networks (GANs), variational autoencoders (VAEs), diffusion models, transformers, and large language models (LLMs), producing large numbers of novel peptide sequences. However, there is no standard approach for testing AMP activity in human serum. This systematic review aims to identify the DL architectures used in AMP design, describe the serum concentrations used in experimental studies, and assess reported activity retention in human serum. The findings will highlight gaps between computational design and experimental validation and may help inform more consistent serum testing approaches in future studies.