Integrative QSAR modelling of multi-species pesticide baseline (narcosis) toxicity

Baseline toxicity modelling aims at estimating the lowest dose of pesticides leading to non-polar narcosis. Existing logK_ow linear regression models often exhibit less resistance to outliers and noise. In this work, from the viewpoint of quantitative structure-activity relationship (QSAR), we propose an integrative extreme gradient boosting model (XGBoost) model to combine multimodal features, comprising Mordred descriptors, PubChem fingerprints and species one-hot encodings, for in silico modelling of pesticide baseline (narcosis) toxicity. The integrative XGBoost model facilitates transferring knowledge across species at different trophic levels, including four fishes (Danio rerio, Bluegill, Fathead minnow, and Rainbow trout), invertebrate crustacean Daphnia magna, alga Pseudokirchneriella subcapitata, and bacterium Vibrio fischeri. K-fold stratified cross validation (k = 10) shows that the integrative XGBoost model achieves competitive performance, with overall 0.96 R², and per species R² ranging from 0.71 (Danio rerio) to 0.96 (Bluegill), significantly outperforming logK_ow linear regression, non-linear support vector regression (SVR), and individual XGBoost models. SHapley Additive exPlanations (SHAP) show that SLogP contributes most to baseline toxicity, confirming the significance of hydrophobicity SLogP/logKow in determining baseline toxicity. Lastly, the case studies on Daphnia magna and Pimephales promela show that the derived logTR values exhibit agreements with the known MoAs of control pesticides, indicating the potentials of the integrative XGBoost model to predict baseline and excess toxicities, from which to further infer polar-narcosis, reactive and specifically-acting MoAs.