BCDPi: An interpretable multitask deep neural network model for predicting chemical bioconcentration in fish.

Area of Science:

• Environmental chemistry
• Toxicology
• Computational chemistry

Background:

• Accurate prediction of chemical bioconcentration is vital for environmental risk and toxicological assessments.
• Existing methods may lack accuracy or interpretability for regulatory decision-making.

Purpose of the Study:

• To develop a robust multitask deep learning model for predicting chemical bioconcentration potential.
• To provide an interpretable tool for assessing environmental risks of chemical compounds.

Main Methods:

• A multitask deep learning model was developed to classify compounds into non-bioconcentrative (non-BC), weakly bioconcentrative (weak-BC), and strongly bioconcentrative (strong-BC) categories.
• SHapley Additive exPlanations (SHAP) technology was used for model interpretation.
• Binary and ternary classification models were trained and evaluated for predictive performance.

Main Results:

• Binary classification models achieved >90% accuracy and 0.95 AUC.
• The ternary classification model demonstrated an overall accuracy of 91.11%.
• Key physicochemical properties influencing bioconcentration include molecular weight, polar surface area, solubility, and hydrogen bonding; eight structural alerts were identified.

Conclusions:

• The developed deep learning model provides accurate and interpretable predictions of chemical bioconcentration potential.
• The BCdpi-predictor online tool is available for free use, supporting environmental risk assessment and regulatory frameworks like REACH.
• The findings offer valuable tools and insights for chemical bioconcentration prediction in environmental safety evaluations.