A Machine Learning framework for predicting peptide inhibitors of insect voltage-gated sodium channels

Peptides represent environmentally safer and target-specific alternatives to conventional insecticides, offering advantages such as biodegradability and reduced off-target effects. However, their practical application remains limited by the complexity and cost of chemical synthesis and large-scale screening. In this study, we developed a machine learning (ML) framework to predict peptide inhibitors of insect voltage-gated sodium channels (VGSCs), which are key targets in neuronal signaling and pest control. Six well-established ML algorithms were systematically evaluated, with the Support Vector Classifier (SVC) achieving the best predictive performance. Model interpretability analysis using SHAP revealed that physicochemical descriptors, particularly those describing structural relationships and amino acid interactions, were the most influential for model predictions, consistent with the structural determinants of VGSC-toxin interactions. The top-ranked plant-derived peptides predicted by the model (UniProt: P0DKH7, P56552 and P81930) were further validated through molecular docking and molecular dynamics simulations with the Drosophila suzukii VGSC, a major pest of soft-skinned fruits, confirming stable interactions at the pore and voltage-sensing domains. These peptides, classified as cysteine-rich defensins, exhibited structural patterns compatible with known ion channel modulators. Although the limited availability of experimentally validated insect VGSC inhibitors constrains model generalization, the proposed approach demonstrates the potential of ML-driven sequence analysis to accelerate peptide discovery. By integrating predictive modeling with molecular simulations, this work provides a computationally efficient and biologically meaningful strategy for identifying novel bioactive peptides for sustainable pest management. • The model learned key physicochemical and structural activity determinants • Predicted peptides showed stable interactions and potential inhibition activity • In silico analysis supports a target-specific machine learning approach