10 AMP MIC Predictors Tested: Which Can You Trust?

Summary. Researchers at Hunan Normal University in China reported in Small Methods, published online on 15 October, a retrospective benchmark of 10 publicly released tools that predict the minimum inhibitory concentration (MIC) of antimicrobial peptides (AMPs) directly from sequence. Such quantitative models, spanning classical machine learning and more recent deep learning, are increasingly used to prioritise candidate sequences before synthesis. The group ran each tool out of the box on an external dataset of newly reported AMPs with measured activity and on reconstructed versions of the tools' own native datasets, scoring absolute MIC regression, reproducibility, directional prediction patterns, ranking and virtual-screening enrichment.

Research context (RUO). The authors report that most released tools showed poor absolute MIC regression and limited transfer to the external benchmark, although selected tools kept useful relative-ranking and enrichment ability. Re-running the published implementations also revealed marked differences in how reproducibly they could be executed. Error-stratified analyses linked prediction errors to peptide residue composition and showed a general compression of the predicted activity range. The evaluated tools are listed with their public code repositories, and the new benchmark data are provided as supplementary files. The study is computational and makes no biological or clinical claims. The authors conclude that current quantitative AMP predictors are best used as early-stage screening filters rather than as substitutes for experimental MIC measurement. For peptide-design and microbiology laboratories, the paper is a practical reference for choosing and validating such tools.