The accelerating emergence of antibiotic-resistant pathogens such as Staphylococcus aureus intensifies the need for faster antimicrobial discovery strategies. Artificial intelligence (AI) enables large-scale inference of candidate antimicrobial peptides (AMPs), but computational predictions alone cannot establish biological function. This study tests a genome-guided approach that couples biologically grounded candidate generation with deep-learning prioritization and experimental validation to assess whether AI-prioritized, genome-derived peptide fragments can yield validated antimicrobial candidates.
Rather than exploring unconstrained or randomly generated sequences, the authors used genome-guided mining to derive candidate fragments from two genomes of the skin-associated yeast Malassezia furfur. The rationale for this source is ecological and evolutionary: organisms that coexist with bacterial colonizers have sequence space that has been shaped by biological interactions and selective pressures, potentially enriching for bioactive peptide motifs.
Peptide fragments obtained from the two M. furfur genomes were first filtered by physicochemical properties to remove sequences unlikely to be compatible with peptide synthesis or biological testing. After physicochemical filtering, the fragments were prioritized using deep-learning AMP predictors. The prioritized list generated by AI served as the basis for experimental synthesis and subsequent functional characterization.
A subset of AI-prioritized, physicochemically filtered fragments was synthesized for empirical testing. The study closes the loop from computational prediction to bench validation by moving selected sequences through synthesis and into a series of biological and biophysical assays designed to evaluate antimicrobial activity, mechanism-relevant interactions, and host cytotoxicity.
Selected synthesized peptides underwent cross-kingdom antimicrobial screening against Staphylococcus aureus. Experimental characterization combined kinetic growth assays, which measure bacterial growth dynamics in the presence of peptide candidates, with ultrastructural assays to observe morphological effects on bacterial cells. These assays provided direct functional readouts of peptide antimicrobial potential.
To complement experimental testing, the authors performed in silico structural prediction for selected peptides and analyzed lipid–membrane interaction properties. These computational characterizations aimed to interpret potential mechanisms of antimicrobial action, such as membrane disruption or interaction with lipid bilayers, and to contextualize experimental observations from growth and ultrastructural assays.
Because the peptide source is skin-associated and potential applications could involve contact with human tissues, the study included cytotoxicity screening on human keratinocytes. This step assessed whether AI-prioritized, genome-derived peptides with antimicrobial activity also exhibited unacceptable toxicity toward a relevant human cell type.
The integrated pipeline demonstrated that AI-guided genomic mining enriched biologically motivated sequence space for peptides with measurable antimicrobial activity against S. aureus. At the same time, the experimental results revealed biases and limits to the generalizability of current deep-learning AMP predictors. The authors report that AI prioritization improved the likelihood of identifying functional candidates relative to unguided sequence exploration, yet model-driven inference did not eliminate the need for empirical validation and exposed areas where predictors can misrank or miss biologically relevant sequences.
This study provides an experimental assessment of a model-guided route for AMP discovery that links genome-derived candidate generation, AI-based inference, peptide synthesis, and functional characterization. By applying this pipeline to fragments from Malassezia furfur and testing activity against Staphylococcus aureus, the authors illustrate both the promise and the current limitations of AI-assisted antimicrobial discovery. The work offers a reproducible workflow and highlights the importance of integrating in silico prediction with diverse experimental assays to validate candidate AMPs and to identify areas where AI models require refinement.
Note: The article is a preprint posted on August 26, 2026. The authors declared no competing interests. Specific numeric results, sequence identities, model names, and experimental metrics were not reported in the summary source provided and therefore are not detailed here.