Greening the fertilizer industry with genetically modified bacteria and machine learning.
December 18, 2023
Daniel Trettel, a postdoc at Los Alamos who specializes in bio-manufacturing, is building nanoreactors capable of producing hydrogen using next-to-no fossil fuels. Trettel started the project last year, in part, to solve the problem of greenhouse gas emissions tied to the production of ammonia, an inorganic fertilizer that accounts for 10 percent of the fertilizer used globally. Ammonia (NH3) is mostly hydrogen and right now, 95 percent of hydrogen is produced through steam methane reforming—a process that emits 13 metric tons of carbon dioxide for every ton of hydrogen it produces. Trettel is reconfiguring hydrogenase, an enzyme produced by bacteria that lets them eat biomass and burp hydrogen, turning them into perfect little hydrogen factories in every way but one. The enzyme only works in environments that lack oxygen. “I want to tweak hydrogenase so that it produces hydrogen in the open air,” says Trettel. Do that, and the process could be affordable enough to be industrialized.
Trettel’s idea is to wrap hydrogenases in an engineered protein shell that blocks oxygen while emitting hydrogen through pores, the parts of the proteins that allow molecules to pass. From trillions of different possible combinations of protein shapes and charges, he and his team must discover the few proteins with pores that could work. Done by hand, the search could take lifetimes. But Trettel is employing machine learning (ML), and he figures he needs six months.
It works like this. Trettel constructs dozens of different protein shells known to either block some oxygen, emit some hydrogen, or do a bit of both. He then tests each shell’s effect on the bacteria’s hydrogen-producing ability in oxygenated and deoxygenated environments. Along with the proteins’ amino acid sequences, Trettel uploads the results from these tests into an ML algorithm that analyzes the shells’ pores and fine tunes the pores’ shapes by digitally adjusting the amino acid sequences. There are 8000 possible sequences for the proteins, and the ML algorithm rapidly tests each one, learning as it computes which pore shapes are best at blocking oxygen and emitting hydrogen. The algorithm’s findings give Trettel the blueprints for the proteins with the best pore shapes.
With these designs in hand, Trettel reengineers the bacteria to express the ML-optimized proteins and retests the shells in the lab to see how well their pores emit hydrogen. He repeats the process dozens of times, with the algorithm suggesting subtler and subtler tweaks to the amino acid configuration. Eventually, Trettel and the machine will zero in on the shell and pore combination that emits the most hydrogen in open air. The result could be a step toward a cleaner industry.