Evaluating machine learning models for predicting pesticide toxicity to honey bees

Area of Science:

• Agrochemical research
• Environmental toxicology
• Computational chemistry

Background:

• Small molecules are crucial in various scientific domains, including agrochemicals.
• Agrochemical research, particularly for species-specific toxicity, lacks extensive datasets compared to biomedical research.
• Honey bees (Apis mellifera) are vital pollinators, making their toxicity data critical.

Purpose of the Study:

• To assess the effectiveness of diverse machine learning (ML) methods for modeling chemical toxicity in honey bees.
• To evaluate the generalizability of ML models trained on biomedical data when applied to agrochemical datasets.
• To underscore the need for specialized datasets and ML models for the agrochemical sector.

Main Methods:

• Development and utilization of ApisTox, a novel dataset of experimentally validated chemical toxicity to honey bees.
• Application of various ML techniques, including molecular fingerprints, graph kernels, and graph neural networks.
• Comparative analysis of model performance on ApisTox versus established medicinal datasets (MoleculeNet).

Main Results:

• ApisTox represents a unique chemical space distinct from typical medicinal datasets.
• State-of-the-art ML algorithms trained on biomedical data exhibit performance degradation when applied to agrochemical toxicity data.
• Current ML models show limited generalizability beyond their training domain, particularly for non-medicinal applications.

Conclusions:

• Existing ML models trained on biomedical data are not directly transferable to agrochemical toxicity prediction.
• There is a critical need for the development of more diverse datasets tailored to the agrochemical domain.
• Targeted ML model development is essential for accurate prediction of agrochemical toxicity and environmental safety.