Abstract
Small molecules play a critical role in the biomedical, environmental, andagrochemical domains, each with distinct physicochemical requirements andsuccess criteria. Although biomedical research benefits from extensive datasetsand established benchmarks, agrochemical data remain scarce, particularly withrespect to species-specific toxicity. This work focuses on ApisTox, the mostcomprehensive dataset of experimentally validated chemical toxicity to thehoney bee (Apis mellifera), an ecologically vital pollinator. We evaluateApisTox using a diverse suite of machine learning approaches, includingmolecular fingerprints, graph kernels, and graph neural networks, as well aspretrained models. Comparative analysis with medicinal datasets from theMoleculeNet benchmark reveals that ApisTox represents a distinct chemicalspace. Performance degradation on non-medicinal datasets, such as ApisTox,demonstrates their limited generalizability of current state-of-the-artalgorithms trained solely on biomedical data. Our study highlights the need formore diverse datasets and for targeted model development geared toward theagrochemical domain.