AquaTox-Predictor: An MMoE-Enhanced Multimodal Deep Learning Framework for Aquatic Ecotoxicity Prediction

As the concentration of compounds in aquatic environments rises, the degradation of aquatic habitats underscores the increasing importance of studying the impact of chemicals on different aquatic populations. In this study, we developed AquaTox-Predictor, a multitask, multimodal classification model that integrates molecular fingerprints, molecular graphs, and a Multi-gate Mixture-of-Experts (MMoE) framework for prediction of seven aquatic ecotoxicities of chemicals. The results of the validation sets demonstrated that our model outperformed other methods, achieving the highest average AUC with an improvement of up to 5.75% compared to existing approaches. Our method can correctly identify the substructures that define molecular properties as well as the differences and relationships between end points, thus enhancing interpretability. These results demonstrate that AquaTox-Predictor can achieve high accuracy in predicting various aquatic toxicities and conduct interpretability analysis on compounds, which is conducive to its application in research on other more complex toxicity end points. All the code and datasets are freely available online at https://github.com/jwZhang-lab/AquaTox-Predictor.