Computational Screening of Plant-Derived Biopesticide Candidates Using Molecular Descriptors, Insect-Target Scoring, and Machine Learning

Abstract

Synthetic insecticides remain effective, but rising resistance and harm to non-target species have renewed interest in plant-derived alternatives. Compounds such as azadirachtin, the pyrethrins, rotenone, and nicotine have shaped the history of pest control, and many more plant metabolites are likely to have insecticidal potential. To explore this chemical space systematically, we assembled 97 phytochemicals, 51 with reported insecticidal activity and 46 common plant metabolites without such activity, and processed them through an open-source cheminformatics pipeline. For each molecule we computed 18 physicochemical descriptors, scored similarity to known actives across eight insect molecular targets (acetylcholinesterase, juvenile hormone esterase, ecdysone receptor, chitin synthase, voltage-gated sodium channel, GABA-gated chloride channel, nicotinic acetylcholine receptor, and midgut aminopeptidase N), and trained five classifiers. Random Forest performed best (test accuracy = 0.92, ROC-AUC = 0.987). The most discriminating features were lipophilicity (LogP), hydrogen-bond donor count, and molar refractivity, exactly the properties that govern insect cuticle penetration. A consensus score combining target similarity, similarity to actives, and insecticide-likeness ranked Tephrosin, Limonin, Quassin, Nimbin, Rotenone, Pyrethrin II, Gedunin, and Deguelin at the top, all compounds with established insecticidal activity. The pipeline recovers known biology from a small, heterogeneous dataset and offers a transparent triage tool for prioritising plant compounds in early biopesticide discovery.