DeepMind's BOTANIC-1 hunts plant DNA mutations with 90% accuracy

Crop genetics usually hits an infuriating wall called linkage disequilibrium: when a single nucleotide flip makes melon vines drought-resistant or changes flower sex to boost yields, that mutation travels with thousands of harmless passenger mutations. Sorting the causal driver from the passengers has traditionally meant years of cross-breeding and greenhouse screening.

Now, Google DeepMind and Paris-based AI biology lab Living Models have introduced BOTANIC-1, a specialized plant genomic foundation model paired with Gemma 4, built to untangle decades-old genetic bottlenecks in an afternoon.

The agent and the plant genome model

Instead of trying to brute-force a generalist LLM into reading raw DNA—which ruins token efficiency and degrades core reasoning—the team split the job in two.

BOTANIC-1 acts as the biological foundation. Pretrained on the DNA sequences of 320 plant species, it has essentially learned evolutionary constraints: what parts of plant genomes stay conserved across millions of years, and what changes actually disrupt biological function.

Sitting on top of it is Gemma 4 E4B, acting as an autonomous computational biologist. It takes natural language questions from researchers, plans hypotheses, writes terminal command-line pipelines, and fires queries to BOTANIC-1. Crucially, Living Models runs the E4B agent locally on a single NVIDIA L4 GPU via Ollama, keeping sensitive crop IP on-premise without requiring monster compute clusters.

Pinpointing the target with 90% accuracy

To prove it in the real world, the researchers tested a known melon trait identified by France's INRAE, where a single letter swap in the CmEIN3 gene shifts flower sex from female to hermaphroditic—a massive win for commercial fruit yields.

Classical bioinformatics narrowed the search to thousands of regional candidates, but hit a ceiling trying to break genetic ties among the remaining variants.

When handed nearly 2,500 candidate mutation positions, the Gemma and BOTANIC-1 pipeline analyzed the genomic landscape and successfully ranked the correct causal mutation in first place 90% of the time. What usually demands years of field trials and laborious CRISPR screening collapsed into hours of automated computation.

Why it matters

We think the real breakthrough here is the modular architecture. Agricultural research doesn't need a bloated multitrillion-parameter model pretending to understand biology; it needs domain-specialized models wrapped in smart, lightweight agent layers that can run on standard commercial hardware.

For crop scientists and breeders, this radically lowers the barrier to entry. Sequencing plant DNA has gotten cheap, but making sense of the unannotated spreadsheets full of SNPs was the real bottleneck. Compressing the timeline from seasons to hours changes how fast agriculture can react to climate shocks and disease outbreaks.

That said, ranking a known monogenic floral switch is a best-case proof of concept. The true test for BOTANIC-1 will be multi-gene traits like yield and drought tolerance, where dozens of subtle variants interact at once. But as an operational blueprint for lab biology, it looks remarkably practical.