Graph neural network-based prediction and interpretation of Daphnia toxicity using distinct scale molecular representations

Abstract

The lethal concentration (LC50) of Daphnia is a significant toxicity index for water quality assessment and chemical management; however, obtaining LC50 values for all substances poses difficulties in terms of cost and time. This study developed a deep learning-based framework to overcome these challenges that integrated molecular- and macro-level features to predict LC50 values and interpret toxicity mechanisms. In molecular structure-based graph machine learning, the presentation of molecular structure data plays a crucial role. In this study, to improve interpretability and stability, we applied three graph representations, including original molecules, Simplified Molecular Input Line Entry System (SMILES) notations, and SMILES Arbitrary Target Specification (SMARTS) notations for machine learning modeling. Consequently, the framework using SMARTSbased fragment graph representations consistently achieved a coefficient of determination (R2) above 0.825 regardless of the deep learning architectures, establishing it as the most appropriate representation method. These findings demonstrated the robustness and expressivity of simpler fragment-level graph representations for training toxicity-relevant features. Moreover, for this representation, standard graph neural network models outperformed more complex variants (R2 = 0.905 +/- 0.023). Model interpretation further identified the key functional groups and macro-level factors responsible for toxicity, with distinct patterns observed across the five MOA classes (e.g., narcosis, polar narcosis, unspecific reactivity, specific mechanism, unclassified). This advanced multimodal molecular graph-based neural network framework facilitates the management of extensive chemical inventories by enabling streamlined prioritization, rapid hazard identification, and mechanistic hypothesis generation for Daphnia toxicity assessments.