Improving Solubility-Based Environmental Risk Assessment with Probabilistic Correction Methods: Toward Reliable GeoAI Systems

Abstract

Models for predicting molecular solubility are indispensable for understanding chemical properties in location-specific environmental aware Artificial Intelligence. Aqueous solubility (LogS) is a key environmental hazard parameter that regulates transport of pollution, water quality, and ecosystems’ susceptibility to toxic chemical exposure. Here we present an interpretable machine learning pipeline for LogS prediction that was built using the ESOL (Estimated SOLubility) dataset and four chemical interpretable descriptors. We trained a Random Forest regression model that yielded favorable predictive performance (MAE = 0.59, RMSE = 0.86, R2 = 0.84). For transparency and explainability of the model’s output we further applied SHAP (SHapley Additive exPlanations) to quantify the contribution of each model descriptor and determined that hydrophobicity (LogP) and molecular weight contributed the most to predicting LogS.The contributions of this research are three-fold:(i) development of a lightweight and interpretable framework for LogS prediction and screening,(ii) addition of the SHAP method to support explainability, and (iii) demonstrating model applicability in GeoAI (geospatial AI) in environmental risk assessment. This study is crucial in supporting regulators, scientists, and policymakers in building trustworthiness and substantively assisting with a practical and trustworthy tool towards chemical risk assessment and sustainability in decision architecture for environmental decision-making.