Interpretable support vector classifier for reliable prediction of antibacterial activity of modified peptides against Escherichia coli

Abstract

Antimicrobial peptides (AMPs) are promising alternatives to traditional antibiotics, whose effectiveness is declining due to rising antimicrobial resistance (AMR). To accelerate AMP discovery, we developed ISCAPE (Interpretable Support Vector Classifier of Antibacterial Activity of Peptides against Escherichia coli), a machine learning (ML) model that addresses the limitations of current AMP predictors. ISCAPE requires only a Simplified Molecular-Input Line-Entry System (SMILES) string as input and can predict the activity of both natural and chemically modified peptides against E. coli ATCC 25922. Activity is defined by a minimum inhibitory concentration (MIC) threshold of ≤16 μg/mL. To ensure reliability, only MIC values obtained under comparable experimental conditions were included in our curated dataset. ISCAPE outperformed the state-of-the-art AntiMPmod, achieving an area under the receiver operating characteristic curve (AUROC) of 91.83% and a Matthew's correlation coefficient (MCC) of 71.86%. Features driving this performance include the fraction of carbon-carbon pairs and feature- and count-based extended connectivity fingerprints (ECFPs). Model interpretability is enhanced through SHapley Additive exPlanations (SHAP), which identifies the molecular features most critical for AMP activity. To our knowledge, ISCAPE is the first interpretable ML predictor capable of predicting antibacterial activity for both natural and modified peptides against a specific E. coli strain. It is a user-friendly tool that allows experimentalists to pinpoint key molecular features, reducing the need for extensive structure-activity relationship (SAR) studies and guiding the design of novel AMPs.