Kotli, M.; Piir, G.; Maran, U. Multi-task prediction of pesticide toxicity to pollinators with censored data and interpretable attributions on graph neural networks. Journal of Chemical Information and Modeling

Abstract

Bee populations are declining globally due to multiple stressors, with pesticide exposure being a critical factor. Predicting the toxicity of pesticides to bees is essential for risk assessment and regulatory decision-making, as experimental testing remains costly and time-consuming. In this study, we present a multi-endpoint graph neural network model for predicting the acute toxicity of pesticides to honeybees (Apis mellifera) and bumblebees (Bombus terrestris). The proposed model is based on a curated dataset of 780 compounds with four toxicity endpoints (acute contact and oral LD50 for both species) and a novel architecture that incorporates endpoint-specific embeddings, cross-endpoint attention mechanisms, and a custom loss function to handle the prevalent interval-censored measurements in ecotoxicological data. The model can predict four endpoints simultaneously and achieves strong predictive performance. The external validation R² values reach 0.79 and 0.78 for honeybee contact and oral toxicity, respectively. For high-confidence bumblebee classification endpoints, Matthews Correlation Coefficients reach 1.00 and 0.77 for contact and oral toxicity, respectively. An interpretable saliency-based attribution framework is introduced that identifies chemically relevant structural fragments that drive toxicity predictions, including known toxic moieties such as the 2,2-dimethylcyclopropyl group characteristic of pyrethrins. The model provides estimates of both aleatoric and epistemic uncertainty, calibrated predictive intervals, and a two-axis Applicability Domain. Proposed approach advances computational toxicology for pollinators by simultaneously addressing handling of censored data, multi-species prediction, and model interpretability within a unified framework.