Graph neural networks-enhanced relation prediction for ecotoxicology (GRAPE).

Area of Science:

• Environmental toxicology
• Computational chemistry
• Machine learning

Background:

• Chemicals pose risks to aquatic ecosystems.
• Traditional ecotoxicity testing is time-consuming and resource-intensive.
• Novel computational methods are needed to predict chemical impacts.

Purpose of the Study:

• To develop and evaluate a Graph Neural Network (GNN) model, named GRAPE, for predicting aquatic toxicity.
• To provide an alternative or complementary approach to traditional in vivo ecotoxicity testing.
• To integrate diverse aquatic toxicity data within a unified graph framework.

Main Methods:

• Formulated ecotoxicology as a relation prediction task using GNNs.
• Developed the GRAPE model to simultaneously model 444 aquatic species and 2826 chemicals.
• Augmented species and chemical features to enhance prediction accuracy.
• Compared GRAPE against Logistic Regression (LR) and Multi-Layer Perceptron (MLP) models.

Main Results:

• GRAPE significantly outperformed LR and MLP, with recall improvements up to 30%.
• GRAPE demonstrated superior performance in predicting toxicity for novel chemicals (≥100% recall improvement) and new species (up to 13% recall improvement).
• GRAPE accurately predicted effects of novel chemicals (104/126) and on new species (7/8), and showed high accuracy for metallic and organic chemicals.

Conclusions:

• GRAPE represents a pioneering application of GNNs in ecotoxicology.
• The model's ability to integrate species and chemical data enhances prediction of aquatic toxicity.
• GRAPE offers a robust and scalable computational tool for assessing chemical risks to aquatic environments.