TP-Transformer: An Interpretable Model for Predicting the Transformation Pathways of Organic Pollutants in Chemical Oxidation Processes
Zhenhua Dai, Jihong Xu, Jian Guan, Mingyang Feng, Yang Liu, Cuili Xing, Xuanying Cai, Shuchen Wang, Lushi Lian, Hongyu Dong, Zhiyong Jason Ren, Wei Shi, Alicia Kyoungjin An, Shifa Zhong, Xiaohong Guan
Abstract
Chemical oxidation is pivotal in remediating organic pollutants in aquatic systems; however, it frequently yields transformation products (TPs) with potential toxicological profiles surpassing those of the parent pollutants. Comprehensive identification of these TPs is imperative for environmental risk assessment and optimization of oxidation methodologies. Traditional experimental approaches for TP elucidation are often hindered by substantial financial and technical constraints, limiting their applicability in high-throughput scenarios. Here, we introduce TP-Transformer, an advanced deep learning framework designed to predict both the structures of TPs and their corresponding formation pathways. Trained on Chem_Oxi_2K, a meticulously curated data set comprising 2780 pollutant degradation reactions, TP-Transformer achieved a notable accuracy of 86.28% in TP prediction. The model adeptly reconstructs complete degradation pathways, addressing the intricate challenge of pathway elucidation. Attention analyses indicate that the TP-Transformer discerns reactive moieties within substrates and correlates them with specific reaction conditions, emulating expert-level chemical reasoning. Experimental validations corroborate the model's robustness, with accurate TP predictions ranging from 80.20 to 92.86% for five pollutants absent from the training data set. These findings underscore TP-Transformer's potential to transform environmental chemistry by offering a scalable, precise, and efficient alternative to traditional experimental methodologies, thereby enhancing water treatment strategies and safeguarding ecological and human health.