BOTANIC-1 and Gemma: separating reasoning from genomic evidence
Living Models pairs a plant-genome specialist with a general model that coordinates the analysis.
Gemma and BOTANIC-1 analysis workflow. Credit: Living Models · Google DeepMind. Original image ↗
Living Models pairs Gemma 4 with BOTANIC-1 to prioritize candidate mutations in plant genomes. The Google DeepMind case study describes a modular workflow: Gemma coordinates tools, while BOTANIC-1 supplies a domain-specific genomic score.
Keep the reasoning assistant and the biological model distinct
Living Models pairs a conversational system with a specialist model. A question about a mutation must first become a well-specified biological query: the relevant organism, sequence region, candidate change, and predicted effect all matter.
Gemma helps organize that investigation; BOTANIC-1 contributes plant-genome predictions. Neither role makes a generated explanation an experimental observation. The DeepMind case study is useful because it describes the tools and the evidence behind a specific biological result.
Why use two models?
A general language model can write analysis pipelines and recover from tool errors. That does not automatically give it a reliable representation of DNA. A genomic model supplies different evidence, learned from sequence data.
In the case study, Gemma 4 E4B runs locally and queries BOTANIC-1. The specialist was pretrained across 320 plant species.
The reported experiment
The team studies a previously validated melon mutation. Conventional filtering leaves linked candidates difficult to distinguish. Adding BOTANIC-1 ranks the known causal mutation first among 2,494 single-nucleotide candidates in the described analysis.
That result concerns the reported experiment. It should not be expanded into a claim that the system establishes causation for arbitrary crops or replaces laboratory validation.
Work through the likelihood ratio
The case study describes a log-likelihood ratio comparing an alternative nucleotide with the reference, given surrounding sequence context. A useful teaching example starts with invented probabilities so the calculation can be checked independently of a model:
If the alternative receives probability 0.01 and the reference 0.50, their ratio is 0.02 and its natural logarithm is approximately −3.912. If the alternative is also 0.50, the score is zero. The graph changes only the alternative probability while keeping the reference fixed. These probabilities are illustrative; they are not BOTANIC-1 outputs from the melon experiment.
A negative score says that the model considers the alternative less compatible with the supplied sequence context than the reference. It is not the probability that a mutation causes a phenotype. Nor does a more negative score tell you whether the effect is desirable for breeding. The model's training distribution, species coverage, genomic coordinates, and surrounding context all affect interpretation. A ranking is a way to prioritize investigation, not to skip it.
Keep the causal result distinct from the ranking result
The reported mutation was already validated before this ranking experiment. That matters: the team could test whether the workflow recovered a known causal candidate. Recovering that candidate from a filtered set is a different question from discovering and validating an entirely unknown mechanism. Both can be valuable, but they need different experimental designs.
For a new dataset, record how the candidate set was formed before interpreting rank one. A model could rank the best of a poor set without containing the actual causal variant. Also preserve candidates that the filter excluded, and state whether insertions, deletions, and rearrangements were scored by the same tool. A benchmark on single-nucleotide substitutions cannot silently establish performance on every variant class.
A practical research note should retain the reference genome version, candidate coordinates, variant definitions, context preparation, scores, and validation status. The assistant can organize that record and help inspect tool errors. Its fluent account should not substitute for the record itself. The useful result of the two-model architecture is a prioritized, auditable hypothesis that a scientist can evaluate with evidence appropriate to the biological question.
Interpret the sign of a log-likelihood ratio
Focus the graph and use arrow keys, or tap a point, to inspect values. The complete data table is available below.
View data table
| Alternative probability | Natural-log ratio |
|---|---|
| 0.01 | -3.912023005428146 |
| 0.05 | -2.3025850929940455 |
| 0.1 | -1.6094379124341005 |
| 0.25 | -0.6931471805599453 |
| 0.5 | 0 |
Read the melon result at its actual scale
The reported ranking concerns a particular tested mutation among 2,494 candidates. It supports a claim about that experiment. It does not mean the model has solved crop improvement generally, or that the same ranking accuracy will transfer unchanged to another species, trait, or growth condition.
When reading a new result, identify the candidate set and outcome that were evaluated. Were the alternatives experimentally tested, computationally scored, or selected after other filtering? How was the useful candidate confirmed? These details determine what a headline ranking can establish.
Build a reviewable research question
A practical starting point is to write down one hypothesis and the evidence needed to test it. Specify the organism and genomic context, list the alternative variants, and separate the predicted effect from the measurement that would validate it. Ask the assistant to retain links and tool outputs for each inference.
If the question depends on a dataset or tool that the public release does not provide, record that dependency instead of inventing a setup command. Read the model and access material linked from the case study before planning a local workflow. The story illustrates a way to connect language models with specialist scientific tools; its credibility comes from the experimental chain, not from the fluency of the explanation.
