BeeEcoTox: A Large Language Model-Empowered Graph-Learning Framework for Predicting Ecotoxicity of Agrochemicals to...

Area of Science:

• Environmental Science
• Computational Chemistry
• Toxicology

Background:

• Agrochemicals pose significant ecological risks to bees.
• Existing ecotoxicity prediction models lack accuracy and generalizability due to single-type structural inputs.

Purpose of the Study:

• To develop a multimodal graph-learning framework, BeeEcoTox, for accurate prediction of agrochemical ecotoxicity to bees.
• To overcome limitations of existing models by integrating diverse molecular features.

Main Methods:

• Developed BeeEcoTox, a multimodal graph-learning framework fusing ChemFM semantic features with molecular graphs.
• Utilized a graph isomorphism network with internal batch normalization and structural features from 1,139 agrochemicals.
• Employed cost-sensitive learning to address class imbalance and prioritize high recall for ecotoxic compounds.

Main Results:

• BeeEcoTox achieved state-of-the-art performance with an area under the curve of 0.91 and recall of 0.90.
• Rigorous benchmarking confirmed superior performance against baseline models.
• GNNExplainer was used for model explainability, identifying key toxicophores.

Conclusions:

• BeeEcoTox offers a significant advancement in predicting agrochemical ecotoxicity to bees.
• The framework provides a publicly available web platform for ecological risk assessment.
• Supports the development of new approach methodologies for pesticide safety evaluation.