Predicting pesticide vapour pressures: the power of functional groups

Abstract

Explainable machine learning can aid in deriving chemical rules, which in combination with inverse molecular design methods, can support humans to optimise classes of molecules such as pesticides. This study demonstrates that pesticide vapour pressures can be predicted (77.1% and 83.3% within one order of magnitude) using kernel ridge regression (krr) and XGBoost on quantum chemical molecular properties but these models lack easy interpretability. However, insights (via Shapley additive explanations (SHAP)) can be gained when a framework of functional groups is employed instead. Functional group-based models (krr: 66.7% within one order of magnitude) reveal that aromatic compounds, sulfonic acid derivatives, and carboxylic acid derivatives influence the vapour pressure the most. SHAP value trends indicate a linear relationship between reduced vapour pressure and the frequency of functional groups. A provided list of functional group contributions enables molecular modifications to optimise pesticide vapour pressures.