Predicting adsorption capacities of pharmaceutical pollutants using chemoinformatics and machine learning techniques.

Area of Science:

• Environmental Chemistry
• Computational Chemistry
• Water Treatment

Background:

• Pharmaceutical pollutants are persistent emerging contaminants in aquatic environments, posing ecological and human health risks.
• Their removal via adsorption is promising but highly variable, necessitating predictive modeling.
• Concerns include bioactivity and the spread of antimicrobial resistance.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting pharmaceutical pollutant adsorption capacity (Qe).
• To identify key molecular descriptors and experimental conditions influencing adsorption.
• To create a user-friendly application for predicting Qe.

Main Methods:

• Machine learning models (SVR, XGB, ANN) were trained using chemoinformatics descriptors from SMILES strings and experimental data.
• Feature selection (LassoCV) and multicollinearity analysis refined the descriptor set.
• Model optimization (Optuna) and cross-validation assessed predictive performance.

Main Results:

• Extreme Gradient Boosting (XGB) achieved the highest predictive accuracy (R² = 0.997, RMSE = 2.62 mg/g).
• SHAP analysis identified surface area and nitro groups as influential features.
• The best model was deployed in a Streamlit application with applicability domain checks.

Conclusions:

• Machine learning, particularly XGBoost, effectively predicts pharmaceutical adsorption capacity.
• Chemically interpretable descriptors enhance model understanding and applicability.
• The developed tool facilitates efficient prediction of pharmaceutical removal in water treatment.