Insight into machine learning models to predict toxicity of organophosphorus insecticides to Photobacterium phosphoreum based on a small dataset

Organophosphorus insecticides (OPIs) pose escalating environmental and health threats due to toxicity. To address limitations of traditional methods relying on lab trials, in this study, quantitative structure-activity relationship (QSAR) models were developed to predict toxicity of OPIs to Photobacterium phosphoreum by machine learning in seven dimensions based on a small dataset. Filtering QSAR variables by feature descriptors could simplify and enhance their ability to capture the correlation between OPIs structure and toxicity. Ensemble prediction and Leave-One-Out Cross-Validation ensure model robustness and prevent overfitting. Results indicate that the best model was determined to be the ensemble model (R 2 = 0.961, RMSE = 0.184, MAE = 0.156). Interpretable analyses of the model revealed that charge balance and electrophilic potential points are key determinants of the toxicity of OPIs. Combined with structural fragmentation analysis, replacing chlorophenyl with fluorophenyl, sulfur with oxygen, and long alkyl chains with short alkyl chains can mitigate OPIs toxicity. This work contributes to providing reference information for the design of green OPIs, which is conducive to the application of computational toxicology in practical engineering.