Holo-Mol: An Explainable Hybrid Deep Learning Framework for Predicting Reactivity of Hydroxyl Radical to Water Contaminants Based on Holographic Fused Molecular Representations

Abstract

The reaction rate constants of hydroxyl radical (OH•) in the aqueous phase are crucial physicochemical parameters for assessing the persistence of organic compounds in water environments. However, the continuous influx of newly emerging organic compounds into water creates both experimental and theoretical challenges in understanding the reactivity of their degradation processes. Recently, deep learning (DL) methods have proven their ability to predict chemical properties. This study presented Holo-Mol, a DL model for predicting the OH• rate constants in the aqueous phase. Holo-Mol utilized a fused learning approach based on holographic molecular representations, which provides comprehensive molecular information from three dimensions. Also, the 2D/3D graph attention mechanism and fusion gate were utilized to enhance molecular feature extraction. The model was evaluated on a dataset of 1374 organic compounds and resulted in a high goodness-of-fit that outperforms previous methods. This research further used interpretability methods to analyze the effects of combining representations from different dimensions based on accurate predictions of the model. It demonstrated that different dimensions of representations can provide different patterns of "knowledge" to the model, assisting in its learning process in molecular property prediction. Finally, based on interpretability techniques, this study identified key substructures susceptible to attack by OH•. In summary, the results demonstrated the potential of Holo-Mol in expanding the application of advanced oxidative technologies for treating contaminants in water. Additionally, it offers novel mechanistic insights into the reaction mechanism of OH•.

Keywords: Deep learning, molecular representation learning, model interpretability, Advanced oxidation processes, hydroxyl radical (OH·)