Machine Learning Framework for Multi-Label Antimicrobial Peptide Classification with Interpretable Feature Insights

Antimicrobial peptides are promising alternatives to combat antibiotic resistance, but their multifunctionality poses significant classification challenges. This study introduces a multilabel classification framework using adaptation techniques in Extreme Gradient Boosting (XGB_multi), Random Forest (RF_multi), and Convolutional Neural Networks (CNN_multi) to predict AMP functionalities, including antibacterial, antifungal, antiviral, anticancer, and mammalian cell targeting. A dataset containing \((6,845)\) peptide sequences was curated from the dbAMPv2 database. Sequences were filtered for canonical amino acids and clustered with a \((90%)\) sequence identity threshold using CD-HIT (Cluster Database at High Identity with Tolerance), ensuring diversity and reducing redundancy. Features were extracted using the iFeatureOmega toolkit, combining sequence-based descriptors like Amino Acid Composition (AAC) and Pseudo-Amino Acid Composition (PAAC) with physicochemical properties like hydrophobicity, charge, and secondary structure. XGB_multi achieved the highest accuracy of \((0.919)\), outperforming RF_multi and CNN_multi in other functional predictions. CNN_multi excelled in antibacterial, antifungal, and antiviral tasks with an AUC of \((0.892)\), while RF_multi demonstrated high precision (\((0.861)\)) and subset accuracy (\((0.689)\)). These models outperformed existing tools like AMPfun, MultiPep, iAMPpred, and AMP_scanner v2, achieving up to \((7.9%)\) improvement in AUC for certain functionalities. Feature importance analysis identified charge, hydrophobicity, and structural attributes as critical contributors, with sequence-derived features like PAAC and hydrophobicity5 emerging as highly discriminative. These findings provide a robust foundation for designing innovative antimicrobial therapeutics with multifunctional capabilities to combat drug-resistant pathogens.