Single-task regression naturally adapts to multi-species (eco)toxicological modelling: a case study on animals

Area of Science:

• Computational toxicology and environmental science.
• Development of predictive models for chemical safety assessment.

Background:

• In silico (eco)toxicological modeling is crucial for assessing chemical risks to ecosystems, animals, and humans.
• Current local and multi-task models have limitations in multi-species applicability and require common data points across species.

Purpose of the Study:

• To propose a single-task regression strategy for adaptable multi-species (eco)toxicological modeling.
• To overcome the limitations of existing models in handling diverse species datasets without common pesticides.

Main Methods:

• Aggregated 37,305 ecotoxicological measurements for 29,140 pesticides across 10 animal groups.
• Trained four machine learning models: extreme gradient boosting (XGBoost), deep neural networks (DNN), random forest (RF), and support vector regression (SVR).
• Employed five-fold stratified cross-validation to evaluate model performance.

Main Results:

• XGBoost demonstrated superior performance with R² of 0.67, RMSE of 0.44, and MAE of 0.29.
• The single-task regression model achieved a significant R² increase of 0.08–0.49 compared to local models.
• Morgan bit 389 (a five-atom aromatic ring fragment) was identified as a key predictor by XGBoost.

Conclusions:

• The proposed single-task regression strategy effectively adapts ecotoxicological modeling to multiple species.
• This approach facilitates data augmentation and inter-species knowledge transfer, enhancing predictive accuracy.
• Case studies confirmed the model's credibility by analyzing toxicity and structural similarities of pesticides.