The buzz about machine learning and pesticide toxicity

Bees are vital to agriculture as pollinators, but the diversity of wild bee species is declining. Pesticides are likely playing a role in this decline, and new chemicals must constantly be developed as pests become resistant. Experiments can determine the toxicity of a new pesticide to bees, but they are expensive and time-consuming.

Yang et al. developed a machine learning approach to predict toxicity to bees from the structure of the pesticide molecule. The team used experimental toxicity data to train their model.

“Imagine representing each pesticide molecule as a point in this 3D room,” said author Cory Simon. “The support vector machine draws a dividing plane between the toxic and nontoxic examples.”

A graph represents the molecular structure, with the vertices corresponding to atoms and the edges denoting bonds. Random walks then traverse different vertices along edges, constructing a sequence of atom and bond types along their journey. Accounting for all the walks implicitly places the pesticide molecule in a vector space.

“Suppose you randomly walk in a city. Based on the sequence of the types of places you visit, you could figure out the region of the city (e.g., Pacific Northwest),” said Simon. “The same principle applies for random walks on molecules to distinguish between toxic and nontoxic pesticides.”

The random walk representation performed similarly compared to a classical molecular representation, which looks for specific chemical fragments curated for drug discovery. Remarkably, the random walk representation enabled performant toxicity classification without much prior knowledge, but it is difficult to determine which specific walks result in toxicity.

Other molecular machine learning tasks, such as drug discovery, could employ the same random walk representations.

Source: “Classifying the toxicity of pesticides to honey bees via support vector machines with random walk graph kernels,” by Ping Yang, E. Adrian Henle, Xiaoli Z. Fern, and Cory M. Simon, Journal of Chemical Physics (2022). The article can be accessed at https://doi.org/10.1063/5.0090573 .

This paper is part of the Chemical Design by Artificial Intelligence Collection, learn more here .