Executive Summary

Key numbers

Four cells. The first two are what this pairing managed in the melon case. The second two are how widely that result was measured beyond the single case, and the condition attached to that measurement.

Source: Google DeepMind and Living Models joint announcement, bioRxiv paper.

In melon, the sex of the flowers a single plant produces differs by cultivar. Working out which single base changed to split that trait is an old problem in plant genetics, and it is the problem in this case too. The classical method is bulked segregant analysis. Collect plants with differing traits, compare their genomes, and the interval holding the cause narrows. In this melon case, 47,492 variants across the whole genome came down to one interval on chromosome 2 and 3,061 candidates.

It goes no further than that. The paper describes the state as one where the causal variant lies within the chromosomal interval while thousands of candidates stay unresolved. What this method measures is correlation in inheritance, not whether a variant actually alters how the plant works. Variants inside the interval are inherited together, so they score alike. With conventional bioinformatics tools, the authors write, the analysis reaches the limit of what current standard methods can do and yields a shortlist of variants that still require further experimental work.

The joint announcement puts a number on that limit. Handed only standard command line tools, Gemma 4 in its best runs narrowed the pool to two identical-scoring coding mutations and stopped. One was the validated CmEIN3 driver; the other was an unrelated gene. With equal scores there is no basis for choosing between them, and the only way out is to check both at the bench. The qualifier about best runs has to be read alongside the rest: across twenty runs, the share that ranked the correct variant first was 0.15, so even reaching a two-way tie was not something that happened every time.

Attaching BOTANIC-1 as a tool changed the route. In the second scenario Gemma 4 was given a single specialized tool, the local BOTANIC-1 scoring function. The agent filtered 567 insertions, deletions and rearrangements out of the chromosome 2 candidates and sent the remaining 2,494 single-base variants to that function. The validated CmEIN3 variant came back first, with no tie.

The "four minutes" that coverage has carried around needs care at this point. AI Times, a Korean outlet, wrote that the work pinned the target variant at the top in under four minutes. That figure appears neither in the paper nor in the joint announcement. explainX, an overseas blog that used the same figure, states separately that the four minutes is wall-clock time for the agent run and not for a research program. The announcement reaches for a different scale altogether. It talks about compressing years of traditional crop genetics into hours of computation, and about untangling decades-old genetic puzzles in an afternoon. Four minutes and an afternoon are different stretches of the same job. The first covers the agent shortlisting and scoring candidates; the classical analysis that narrowed the interval sat outside it.

Gemma 4's seat in this arrangement is the direction of the analysis. The joint announcement calls Gemma 4 E4B the autonomous bioinformatic orchestrator, writing terminal pipelines and executing tool calls. It turns biological questions into structured queries, runs standard command line software, reads terminal output, and retries commands that fail. BOTANIC-1 is the scoring function that the conductor calls. It returns a number for how far a site with one base changed departs from the sequence patterns of plant genomes.

The hardware is modest as well. Gemma 4 E4B ran locally through Ollama on a single NVIDIA L4 graphics card. Swapping in a larger model brought no gain in accuracy and only raised the cost of running it, according to benchmarking the announcement released alongside the result.

Measure the two separately and the division of labor sharpens. By the share of twenty repeated runs that ranked the correct variant first, classical tools with no guidance come in at 0.00, expert-guided classical tools at 0.15, and the setup calling BOTANIC-1 at 0.90. Fabricated variants split the same way. The BOTANIC-1 side produced none across all twenty runs, while the two classical setups invented variants in 3 and in 4 of their twenty runs.

That role tends to shrink by a notch as it gets relayed. AI Times wrote that Gemma 4 organizes gene data and automates the analysis. The description is not wrong, but it is a different picture from an autonomous orchestrator that writes pipelines itself and retries failed commands. For anyone reading this case to learn about agent design, that gap is the whole point.

A single case can be luck, which is why a broader measurement sits alongside it. The joint announcement reports two figures from a set of more than 500 experimentally validated variants. The causal variant landed inside the top 0.1% in 15.9% of them, against 4.6% for the best baseline that uses no genomic language model. Widen the window to the top 1% and 48.8% fall inside it, while the classical pipeline reached 33.9%.

The paper counts an evaluation set of the same character as 545 sites across 14 species. The two numbers sit close together without matching digit for digit, so it is safer not to read them as one series. The percentages in the table above come from the announcement.

The margin is clearly real. At the same time, 48.8% reads in another direction too. Even handed a list cut to a hundredth of the original pile, more than half of the causal variants remain outside it. The figure looks inconvenient, yet it is the one closer to how the work actually gets used. Anyone walking into the lab with a top 1% list should put the odds that what they are after is on it at roughly half.

Whether the scorecard can be picked up and used is a separate question. The model the paper puts out front is Botanic1-XL at 3.2 billion parameters, and the recall AUC of 0.752 that leads causal-variant discovery is that model's figure. What the paper writes it will release, though, is S, M and L, and the Hugging Face collection carries those same three at 318 million, 688 million and 2.1 billion parameters. XL is not on the list. The model that posted the score and the models available for download are different, so whether the released ones produce the same number cannot be answered from this material.

The paper does not hide what the melon case is. It describes the work in its own text as a retrospective analysis of a published melon sex-determination locus, and cites the earlier study in its reference list. The variant that came out on top is a value that experiment confirmed and a paper published.

The researchers did not leave the conditions loose, either. The agent was given the candidates and no access to the publication, which amounts to clearing the sheet with the answer written on it off the desk. One thing could not be cleared away. BOTANIC-1, which did the scoring, is a model trained on public plant genomes, and what went into that corpus decides what the score means. explainX puts its finger on this point: in a retrospective test the answer exists in the literature, and a model trained on public plant genomes could in principle have seen related sequence data.

This condition is not confined to the melon. The set of more than 500 in section 3 and the paper's 545 are all variants that experiment has confirmed, and confirmed means published somewhere. So the whole body of evidence the authors put forward rests on tests built the same way. In explainX's phrasing, this is a benchmark of recovery, not a new discovery.

explainX also keeps a separate list of what this work shows and what it does not. On the second list are the success rate across variants in general, any grounds for skipping experimental validation, performance on regulatory variants outside coding regions, traits where several genes act together, and prediction made before the answer is known. That last item is exactly what this article is asking about.

The melon funnel has a boundary of the same kind drawn around it. The agent cleared 567 insertions, deletions and rearrangements out of the chromosome 2 candidates and scored only single-base variants, and the two tied candidates in the announcement were both coding mutations. Had the cause sat among the 567 that were cleared, or outside the coding regions, this route would not have reached it.

The paper claims no general supremacy either. It states that no single model dominates across all tasks, and adds that although plant genomic language models can solve problems where standard bioinformatics pipelines stumble, these more classical approaches can still be applied successfully to other problems. Nor does it say that linkage disequilibrium, the underlying difficulty in pinning a cause down precisely, has been dissolved.

That is not left in the abstract. The paper reports that on 22 frozen-model evaluation tasks its four Botanic1 models occupy the top four positions in aggregate, ahead of every other model, and in the same text it notes that the autoregressive Evo 2 leads variant-effect prediction, where its 7B model reaches an LLR AUROC of 0.724 against 0.721 for Botanic1-XL. Leading in aggregate and leading in every cell are two different statements. The conclusion the authors reach is not a choice between the two camps. The most efficient approach, they write, appears to be combining genomic language models and bioinformatics tools, using each where it performs best.

For anyone designing agents, this case makes one thing plain. The conductor stayed where it was, the tools changed, and 0.15 became 0.90. Swapping in a heavier generalist model was already checked in the announcement and found to do nothing. What raised performance was not depth of reasoning but a layer trained separately on the data of the field. When the question is which layer to invest in, this is a case worth having on hand.

On the evaluation side, something else comes into view. Retrospective sets are easy to build. Gather variants already confirmed, hide them, and ask for them back. The trouble is that a set built that way measures recovery, while what the person using the model wants to know is the score where no answer exists yet. The distance between the two does not shrink by running the model more. It shrinks when the account of how the evaluation set was built arrives with the model.

So a record has to come packaged with the model. When the answers used in evaluation were settled, and whether that moment falls before or after the training data was collected. Whether there is even one case the model named first and experiment confirmed afterward. Without those two lines, any score, however high, draws the same question again. Preparing data for AI to use covers more than cleaning up training data; it covers keeping the provenance of the data used for scoring.

Thank you for reading this far. The paper is on bioRxiv, and the step-by-step figures for the melon case are in the Google DeepMind joint announcement. When your team looks at a scorecard that came with an outside model, does it also ask what kind of problem produced that score? We would be glad to hear it.

• 1.Living Models. (2026). "BOTANIC-1: a series of long-context plant genomic foundation models in the agentic era." bioRxiv, DOI 10.64898/2026.09.04.749355.
• 2.Google DeepMind. (2026). "Living Models — Gemmaverse." Source of the step-by-step melon figures and the Recall@1 benchmark.
• 3.Park, Chan. (2026). "AI finds in four minutes the key gene that raises melon yield." AI Times, 13 October 2026 (in Korean).
• 4.explainX. (2026). "Gemma 4 + BOTANIC-1: Plant DNA Variants Ranked in 4 Min." Analysis blog that flags the limits of retrospective validation.
• 5.Living Models. (2026). "Botanic1 collection." Hugging Face. Source for which model sizes (S, M, L) were released.