Living Models Says Its AI Pair Ranked a Known Melon Mutation First Among 2,494

A known mutation affecting melon flowers rose to the top of 2,494 candidates when Living Models paired a general-purpose AI with a plant-genome specialist. In an October 6 account published by Google DeepMind, the lab described how Gemma 4 E4B and BOTANIC-1 separated genetic suspects that conventional analysis had left tied—helping scientists choose what to validate first.

Why inherited DNA leaves a tie

The experiment revisited a validated change in the melon gene CmEIN3 that shifts flowers from female to hermaphroditic. Standard genetic mapping had narrowed 47,492 genome-wide variations to 3,061 candidates on chromosome 2. That located the neighborhood, but did not identify the culprit.

Plants inherit DNA in blocks, so harmless changes can travel alongside the mutation responsible for a trait. Those changes can show the same inheritance pattern. Separating them may require growing more plants to find rare genetic reshuffling events, or testing individual changes in the laboratory.

With standard bioinformatics tools, Gemma wrote analysis pipelines, filtered sequencing data and annotated mutations. Its best runs ended with two equally scored coding mutations: the validated CmEIN3 change and one in an unrelated gene. Deterministic scripts using the same data reached the same deadlock, the team said.

A specialist supplies a different signal

The second setup gave Gemma access to BOTANIC-1’s scoring function. Rather than ask which mutations travel with a trait, the specialist estimates how compatible each DNA change is with its surrounding sequence. It compares the probabilities of the changed and original DNA letters; strongly negative scores suggest disruption at positions constrained by evolution.

BOTANIC-1 was trained on a corpus of 320 plant genomes and supports sequence context up to about 128,000 DNA letters. Gemma removed 567 insertions, deletions and rearrangements from the regional candidates, then sent the remaining single-letter changes for scoring. The known causal mutation finished first without a tie.

Living Models reported a 90% first-place success rate across 20 runs with BOTANIC-1, versus 15% for expert-guided classical analysis and zero for unguided analysis. The specialist-equipped runs produced no fabricated variants. Classical-tool runs did so in three unguided runs and four expert-guided runs.

48.8%BOTANIC-1: causal mutation in top 1%

Living Models reported this result across a retrospective evaluation of more than 500 published, experimentally validated causal plant mutations.

33.9%Classical pipelines: causal mutation in top 1%

Classical pipelines reached this rate in the same retrospective comparison.

A shortlist, not a new biological discovery

The broader evaluation also placed causal mutations in the top 0.1% of candidates in 15.9% of cases, against 4.6% for the best baseline without a genomic language model. The September 23 technical report describes a causal-variant benchmark containing 545 experimentally validated genetic sites across 14 species.

These are retrospective rankings of known causes, not newly validated crop improvements. Living Models describes the output as a prioritized hypothesis for laboratory testing. The division of labor is deliberate: Gemma organizes the investigation and calls tools; BOTANIC-1 supplies the DNA-specific assessment.

The lab also said larger orchestration models added overhead without improving accuracy. It runs Gemma E4B locally through Ollama on one NVIDIA L4 GPU, keeping the analysis on premises rather than requiring proprietary genomes to be sent away.