Living Models pairs Gemma 4 with BOTANIC-1 to help decode plant DNA
Short summary
Living Models paired Gemma 4 with BOTANIC-1, a plant genome foundation model, to identify causal mutations in crop breeding. The system solved a decades-old genetics problem—distinguishing which mutation actually drives a trait from dozens of hitchhiking variants—in hours instead of months of field trials.
- Classical bioinformatics tools hit a wall on the melon test case, narrowing 47,492 variants to 3,061 candidates but couldn't break linkage disequilibrium to identify the true causal CmEIN3 mutation; BOTANIC-1 ranked it #1 out of 2,494 SNPs using evolutionary constraint scoring.
- BOTANIC-1 placed true causal mutations in the top 0.1% of candidates in 15.9% of cases across 500+ validated mutations, versus 4.6% for the best non-GLM baseline, and in the top 1% in 48.8% of cases versus 33.9% for classical pipelines.
- The modular architecture runs Gemma 4 E4B locally on a single NVIDIA L4 GPU via Ollama, keeping proprietary genomes on-premises while slashing costs and making domain-specialist AI accessible on standard hardware.
Full summary
Living Models paired Gemma 4, Google's open language model, with BOTANIC-1, a specialized AI trained on plant genomes, to solve a real bottleneck in crop breeding. The problem: after field trials and genetic crosses narrow down thousands of mutations to a smaller set, scientists still can't tell which single letter change actually causes a desired trait versus mutations just tagging along for the ride. This genetic hitchhiking—called linkage disequilibrium—has historically forced breeders into months of expensive lab validation, testing mutations one by one with CRISPR. The modular approach lets Gemma handle the conversational layer and tool orchestration while BOTANIC-1 does the heavy genomic reasoning, all running locally on a single GPU to keep proprietary genetic data private.
The team tested this on a real case: a melon mutation in the CmEIN3 gene that shifts flowers from female to hermaphroditic, boosting pollination and yield. Standard mapping left researchers with 3,061 candidate mutations on one chromosome. Using classical bioinformatics tools, Gemma 4 could narrow the field to two candidates tied for top score—one was the real driver, one was noise. A coin toss essentially. When equipped with BOTANIC-1's scoring function instead, Gemma ranked the validated mutation #1 out of 2,494 candidates cleanly, zero ties. BOTANIC-1 works by measuring evolutionary constraint: if a DNA position stayed unchanged across millions of years of plant evolution, swapping in a different letter gets a harsh penalty score. This orthogonal signal—biological consequence rather than population frequency—cuts through the correlation ceiling that classical tools hit.
Across a broader benchmark of over 500 published, experimentally confirmed causal mutations, BOTANIC-1 placed the true culprit in the top 0.1% of candidates 15.9% of the time, versus 4.6% for the best non-genomic models. The practical win: compress years of seasonal breeding cycles and greenhouse screening into hours of computation. Scientists get high-confidence targets to validate rather than guessing between look-alike mutations.