A small Gemma 4 model, working with a plant DNA model called BOTANIC-1, ranked a known melon gene first among 2,494 candidate mutations, according to a write-up on Google DeepMind’s Gemma pages. The test was run by Living Models, a Paris-based AI biology lab, and it looked backward: the right answer had already been established by other scientists. The page’s headline pitch is that the approach could shrink years of crop-genetics work into hours of computing. The word “could” is doing real work, because the page reports a retrospective test, not a new plant.
BOTANIC-1 is what the page calls a foundation model for plants, meaning one large model trained on a broad body of data and then pointed at specific jobs. Here the data is DNA from 320 plant species. In plain terms, it learns statistical patterns in genetic code, including which positions stay nearly identical over millions of years of evolution. A swap at one of those positions probably breaks something, so the model scores each proposed mutation by how surprising it finds the swap given roughly 128,000 surrounding DNA letters.
The problem it targets is a genetic traffic jam. Breeders cross plants that differ in a trait, grow the offspring, sort them, and sequence the pools. That narrows the trait to a region of a chromosome but leaves thousands of suspects, because DNA travels in large blocks and every variant in a block looks equally guilty to statistics. In the melon case, standard mapping took 47,492 genome-wide variations down to 3,061 on Chromosome 2.
The trait itself comes from work led by Dr. Adnane Boualem, a Research Director at INRAE whom the page describes as a collaborator. A single-letter change in a melon gene called CmEIN3 shifts flowers from female to hermaphroditic, which affects pollination and fruit yield.
Living Models ran two setups. In the first, Gemma 4 acted as an automated bioinformatician with ordinary command-line tools. It recovered from tool errors in 9 of 16 cases, and in its best runs it landed on two coding mutations with identical scores, one of them the real driver and one unrelated. In the second, Gemma set aside the 567 candidates that were insertions, deletions or rearrangements and passed the remaining 2,494 single-letter changes to BOTANIC-1 for scoring. The known mutation came out on top with no ties.
Across 20 runs, the BOTANIC-1 setup placed the true mutation first 90 percent of the time (Recall@1 of 0.90), against 0.00 for an unguided Gemma and 0.15 when an expert guided it. It also produced no invented variants, while the classical-tool setups did in 3 and 4 of 20 runs. Gemma 4 E4B, the small model used, ran locally through Ollama on a single NVIDIA L4 GPU, which the page says keeps proprietary genomes in house.
A wider benchmark is more sobering than the melon demo. Across more than 500 published causal plant mutations, BOTANIC-1 put the real culprit in the top 0.1 percent of candidates in 15.9 percent of cases and in the top 1 percent in 48.8 percent. The best non-genomic baseline managed 4.6 percent, and classical pipelines 33.9 percent. Better than rivals, yes, but the true cause still falls outside the top 1 percent in just over half of cases.
Every number comes from Living Models’ own experiments as reported on Google’s page, with no independent replication mentioned. The page is also Google promoting its own model family.
The distance to a finished crop is the larger gap. A ranked list of suspect mutations is a prediction about DNA. A plant carrying a useful trait still has to be bred, confirmed in the greenhouse and tested in fields, a process the page itself describes in terms of seasons and years. The page cites climate pressure on food supply as the motivation. That is a reason for the work, not a result of it, and no bred crop is reported.
For anyone weighing AI tools in agricultural research, the figure to request is the hit rate on traits whose answers are not yet published, because the melon test measured a puzzle that already had a solution.
Google DeepMind, on its Gemma pages at deepmind.google; the scraped page carries no publication date and no byline.
