Antimicrobial activity predictors benchmarking analysis using shuffled and designed synthetic peptides

The antimicrobial activity prediction tools aim to help the novel antimicrobial peptides (AMP) sequences discovery, utilizing machine learning methods. Such approaches have gained increasing importance in the generation of novel synthetic peptides by means of rational design techniques. This study focused on predictive ability of such approaches to determine the antimicrobial sequence activities, which were previously characterized at the protein level by in vitro studies. Using four web servers and one standalone software, we evaluated 78 sequences generated by the so-called linguistic model, being 40 designed and 38 shuffled sequences, with ∼60 and ∼25% of identity to AMPs, respectively. The ab initio molecular modelling of such sequences indicated that the structure does not affect the predictions, as both sets present similar structures. Overall, the systems failed on predicting shuffled versions of designed peptides, as they are identical in AMPs composition, which implies in accuracies below 30%. The prediction accuracy is negatively affected by the low specificity of all systems here evaluated, as they, on the other hand, reached 100% of sensitivity. Our results suggest that complementary approaches with high specificity, not necessarily high accuracy, should be developed to be used together with the current systems, overcoming their limitations.