A Retrospective Benchmark of Quantitative Antimicrobial Peptide Activity Prediction Tools

Antimicrobial peptides (AMPs) are critical candidates for addressing the global antibiotic resistance crisis. AI-driven quantitative models capable of directly predicting peptide antimicrobial activity, specifically minimum inhibitory concentration (MIC), are gaining scientific prominence by providing precise guidance for the rational design of AMPs. To assess their real-world, out-of-the-box performance, we conducted a retrospective benchmark of 10 publicly released quantitative AMP prediction tools spanning classical machine learning and contemporary deep learning approaches. Using an external dataset and reconstructed native datasets, we evaluated absolute MIC regression, retrospective reproducibility, directional prediction patterns, ranking capability, and virtual-screening enrichment. Most released tools exhibited poor absolute MIC regression and limited transferability to the external benchmark, while selected tools retained useful relative-ranking and enrichment capabilities. Retrospective evaluation further revealed marked differences in implementation-level reproducibility among the released tools. Error-stratified analyses identified systematic associations between prediction errors and peptide residue composition, together with a general compression of the predicted activity range. Collectively, these findings characterize the practical performance boundaries of currently accessible quantitative AMP prediction tools and support their use primarily as early-stage screening filters rather than precise substitutes for experimental MIC measurements.