Advancing screening of aquatic persistent, mobile, and toxic substances via integration of graph convolutional...

The accurate identification of persistent, mobile, and toxic (PMT) substances, which are typical emerging contaminants, is critical for assessing risks to aquatic environments and safeguarding drinking water quality. Machine learning plays a significant role in predicting the properties of emerging contaminants, while the common use of single-task models may lead to overfitting. This study integrated molecular graphs with key molecular descriptors (MDs) to develop a graph convolutional network (GCN) model based on multitask learning (MTL), using a dataset of 6820 substances. By using multitask learning, this model emphasizes the interrelationships among three properties, enabling simultaneous multitask binary classification for persistence, mobility, and toxicity. The results demonstrate that the developed MTL-GCN-MD model outperforms both single-task GCN models and traditional machine learning models based on molecular descriptors. The multitask architecture of the MTL-GCN-MD model effectively reduces the risk of overfitting, enabling the model to achieve a balanced accuracy of over 0.81 across three distinct identification tasks on the test sets. Through the visualization of multi-level feature distributions and atomic attention weights, the model's prediction mechanism was systematically analyzed. Kernel density estimation, scaffold diversity, and Tanimoto coefficient were employed to define the model's application domain. An online server has been developed, providing accessible predictions for researchers and water authorities, facilitating practical application in water quality monitoring of PMT substances. This study offers early warning, screening, and regulatory management of hazardous chemicals in water systems.