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

Agrochemical exposure can threaten bees with substantial
ecological
risks. The rapid and accurate prediction of agrochemical ecotoxicity
to bees is thus urgently needed; however, existing models are constrained
by single-type structural inputs, resulting in limited accuracy and
generalizability. This study presents BeeEcoTox, a multimodal graph-learning
framework for predicting agrochemical ecotoxicity to bees. The model
fuses ChemFM-derived semantic features with molecular graphs through
a graph isomorphism network with internal batch normalization, combined
with structural features from 1,139 agrochemicals. Inherent class
imbalance in the curated data set is addressed through a cost-sensitive
learning approach to ensure that the model prioritizes high recall
in detecting ecotoxic agrochemicals without compromising overall performance.
BeeEcoTox achieves state-of-the-art performance, with area under the
curve and recall values of 0.91 and 0.90, respectively, following
rigorous benchmarking against a suite of baseline models. Model explainability
is enhanced through the use of GNNExplainer, a model-agnostic approach
for identifying key toxicophores. BeeEcoTox is deployed as a publicly
available web-based platform (https://www.ai4environ.cn/BeeEcoTox) to support advanced ecological risk assessments of agrochemicals
and the development of new approach methodologies.