A machine learning method could help chemists formulate pesticides that target harmful insects but leave bees alone.
What’s new: Researchers at Oregon State University developed models that classify whether or not a chemical is fatally toxic to bees. The authors believe their approach could be used to screen pesticide formulations for potential harm to these crucial pollinators.
How it works: The authors trained two support vector machines to classify molecules as lethal or nonlethal. The dataset was 382 graphs of pesticide molecules, in which each atom is a node and each bond between atoms is an edge, labeled for toxicity. The researchers used a different method to train each model.
- In one method, the authors translated each graph into a vector that represented structural keys, arrangements of atoms that biochemists use to compare molecules. For instance, one feature indicated that a molecule includes phosphorus atoms. The model took these vectors as input.
- In the other method, the model’s input was a vector that counted the number of occurrences of all possible chains of four connected atoms. Similarly toxic molecules may share similar numbers of such groups.
Results: The two models performed similarly. They accurately classified 81 to 82 percent of molecules as lethal or nonlethal to bees. Of the molecules classified as lethal, 67 to 68 percent were truly lethal.
Behind the news: Bees play a crucial role in pollinating many agricultural products. Without them, yields of important crops like cotton, avocados, and most fruit would drop precipitously. Numerous studies have shown that pesticides are harmful to bees. Pesticides have contributed to increased mortality among domesticated honey bees as well as a decline in the number of wild bee species.
Why it matters: Pesticides, herbicides, and fungicides have their dangers, but they help enable farms to produce sufficient food to feed a growing global population. Machine learning may help chemists engineer pesticides that are benign to all creatures except their intended targets.
We’re thinking: It’s good to see machine learning take some of the sting out of using pesticides.