Scoring pesticide-likeness of organic compounds via explainable one-class machine learning

Pesticides, especially eco-friendly biopesticides, are indispensable for crops protection and food security. In this work, we propose a scoring framework to screen or repurpose existing organic compounds, natural products and pharmaceutical drugs as sustainable agrochemicals. This scoring strategy adopts one-class support vector machine (OCSVM) as base learner to capture the patterns of pharmacophore-relevant fragments of organic compounds exhibiting pesticidal effects, in which only pesticides are required as training data, and the uncertainty and error-proneness of conventional sampling of negative training data are eliminated. T-SNE demonstrates that Morgan fingerprint is effective to transform molecule-level heterogeneity into fragment-level homogeneity to comply the requirement of OCSVM cohesive space. Furthermoe, Morgan bits are explained via SHapley Additive exPlanations (SHAP) in terms of significance, and are visualized via reverse engineering to bridge the gap between abstract representation of local chemical environments and intuitive structural fragments. Hierarchical clustering analyses show that two molecules exhibiting close substructural/fragmental similarities tend to be physically proximal in the inlier subspace, justifying the metric validity and explainability of OCSVM hyperplane distances as pesticide-likeness scores. K-fold cross validation (k = 5) shows that the OCSVM model achieves 4% novelty or outlier rate with the empirically and heuristically determined hyperparameters ([Formula: see text]). External test on 38 independent biopesticides shows that the OCSVM model recognises 92% of external biopesticides, equivalent to or better than two binary supervised classifiers (82.05% and 81.57%). Further model sanity check demonstrates that the OCSVM model has the competitive potentials of screening natural products molecules or fragments as eco-friendly biopesticides.