Machine Learning Screens 50 Million Compounds for Bee Safe Pesticide Repellents

Researchers at the University of California, Riverside have used machine learning to screen more than 50 million chemical compounds for scents that could help keep honey bees away from pesticide treated areas. The work, reported in eLife and announced by the university on September 24, combines computational screening with laboratory and field behavior tests.

The team trained a model on chemical structures and previously recorded insect responses, then refined it with new behavioral data from honey bees and fruit flies. The screening process narrowed tens of millions of possible compounds to about 130 candidates predicted to have strong repellent potential.

Researchers then tested the strongest candidates experimentally. UC Riverside said all seven compounds taken into field experiments repelled freely foraging bees from honeycombs without harming them in the tests, an important distinction because the goal is to alter short term behavior rather than create another toxic exposure.

The potential agricultural application is straightforward. A repellent could eventually be incorporated into or used alongside certain pesticide programs so bees are less likely to contact treated crops during periods of risk, although any commercial formulation would still require product development, efficacy work and regulatory review.

The study also shows why machine learning can be useful in chemical discovery even when biological data are limited. Instead of physically testing an enormous chemical library, researchers can use high quality behavioral observations to rank candidates, test the most promising compounds and feed the results back into the model.

The research was funded in part by the California Research Alliance by BASF, and several authors are inventors on a patent application covering compounds discussed in the study. That commercial context does not establish a market ready product, but it does suggest a path from academic screening toward practical pollinator protection tools if later development is successful.