温暖化に強い作物を育種するAI、DeepMindが研究公開 Engineering more resilient crops for a warming climate

Google DeepMind has outlined a research effort applying machine learning to help breed crops that can sustain yields under a warming climate. Framed around food security, the work taps into rising concern about how heat waves, drought and shifting growing seasons threaten staple agriculture worldwide.

According to the post, DeepMind researchers are modeling the genes, enzymes and metabolic pathways that govern photosynthesis efficiency and tolerance to heat and water stress. The aim is to give plant scientists better computational tools for designing resilient varieties, building on protein-structure expertise developed through AlphaFold and extending it into plant biology. Wet-lab validation is expected to take place in collaboration with external plant-science institutions rather than inside DeepMind itself.

The scientific motivation is well established. Photosynthesis is famously inefficient: the enzyme RuBisCO frequently fixes oxygen instead of carbon dioxide, especially at higher temperatures, triggering photorespiration that wastes energy and reduces growth. International consortia such as the RIPE project have spent more than a decade trying to re-engineer photosynthetic pathways in rice, soy and cowpea. By predicting which enzyme variants or regulatory elements are most likely to improve carbon fixation or thermal stability, machine learning could meaningfully shorten the slow design-build-test cycles that constrain plant breeding today.

DeepMind's move sits within a broader industry trend. NVIDIA has been pushing BioNeMo-style foundation models into genomics, Microsoft has agricultural research partnerships, and startups such as Inari and Pivot Bio combine gene editing or microbial engineering with computational design. Inside Alphabet, X's Mineral graduated as an independent precision-agriculture company focused on phenotyping. DeepMind's contribution appears more upstream and basic-science oriented, complementing rather than competing with those commercial plays.

There are caveats. Translating in silico predictions into field-ready cultivars is notoriously difficult: regulatory pathways for genetically engineered crops differ sharply between regions, and complex traits like drought tolerance involve many genes interacting with environment. Even strong computational candidates may take years of trials before reaching farmers. The blog post is, by its own framing, a research direction rather than a product announcement, and concrete benchmarks or peer-reviewed results are likely to follow rather than accompany it.

Still, the direction is significant. If protein-structure prediction transformed structural biology in just a few years, similar gains in plant systems biology could become an important lever for climate adaptation. Expect more detailed publications, partnerships with CGIAR-affiliated centers or university breeding programs, and possibly open model releases tailored to crop genomes as the work matures.