BOTANIC-1

A plant GWAS interval usually hands a breeder thousands of candidate variants sitting inside one linkage block, statistically indistinguishable from one another, and the annotation available to rank them can only sort them into discrete consequence classes — synonymous, missense, intergenic. In plants that annotation is unusually thin: most species have a single reference genome, and regulatory mechanisms are far less well mapped than in mammals. BOTANIC-1 approaches the problem from the sequence instead. Trained only to fill in masked bases across hundreds of plant genomes, it yields a continuous measure of how constrained each position is, which separates candidates that consequence classes leave tied.

BOTANIC-1 is a family of four plant genomic language models from Living Models, a Paris company, posted as a bioRxiv technical report in September 2026. It succeeds BOTANIC-0, the company's first plant model family, and changes the recipe in three ways: a bidirectional Mamba-2 backbone in place of a transformer, single-nucleotide tokens in place of 6-mers, and a pretraining corpus built from 320 embryophyte species rather than 43. The authors frame the work around an agent-run "Model Factory" — coding agents that launch, monitor and retire the ablation runs behind the data pipeline — which is how they justify systematically ablating each corpus design decision at fixed parameter and token budget.

Because the backbone is a state-space model with no attention and no positional embeddings, context length is cheap to extend. The 318M model was carried from its 8,192 bp pretraining window out to 131,072 bp by a four-stage curriculum, putting a whole regulatory neighborhood inside a single forward pass.

#Key Features

• Bidirectional Mamba-2 encoder: Each block runs two independent Mamba-2 scans, forward and time-reversed, and averages them, giving bidirectional context without attention and without hard-wiring reverse-complement equivariance into the architecture.
• Four sizes, one recipe: Botanic1-S (318M), -M (688M), -L (2.1B) and -XL (3.2B) share an optimisation recipe and a 314.6B-token horizon, so the family doubles as a clean scaling study.
• Long-context extension: A four-stage curriculum doubles the training window at each stage, from 8,192 bp to 131,072 bp, validated by needle-in-a-haystack retrieval and by a chromatin accessibility task scored at 32,768 bp.
• Open pretraining corpus: The 8 kbp mixture selected by the ablations — 7,207,506 windows, 59 Gbp, with sequence case, augmentation margins, source manifests and assembly metadata preserved — is published as a Hugging Face dataset.
• Interpretability artifacts: A BatchTopK sparse autoencoder trained on layer 35 of Botanic1-M (8,192 features, k=64) is released alongside the weights, with features that track coding-region boundaries and splice-site motifs.

#Technical Details

Pretraining is masked language modelling with a 15% mask rate over single-nucleotide tokens, on 8,192-token sequences (a <cls> token plus 8,191 bases), with reverse-complement augmentation at probability 0.5 and a warmup-stable-decay schedule. The corpus was assembled from 5,074 candidate assemblies reduced to 326 annotated genomes, with function-biased window sampling, a MinHash redundancy filter at a 0.3 similarity threshold, and genome-length reweighting — each choice selected by a matched ablation. On the report's aggregate species-balanced score over 22 frozen-model tasks, the four checkpoints take the top four positions, from 0.758 for Botanic1-S to 0.769 for Botanic1-XL, ahead of PlantCAD2-L (0.756), GPN (0.731), Carbon-8B (0.727), PlantBiMoE (0.704) and Evo 2 7B (0.695). Botanic1-S reaches that band with half PlantCAD2-L's trainable parameters and roughly 13× fewer pretraining tokens. Adapted with LoRA, Botanic1-S surpasses ChromBPNet on base-resolution chromatin accessibility prediction, while the same architecture trained from scratch does not.

#Applications

The models suit plant and crop researchers doing causal-variant prioritisation after GWAS or bulk segregant analysis, regulatory element and splice-site annotation in under-annotated species, and chromatin accessibility prediction. The report introduces a causal-variant benchmark of 545 experimentally validated loci across 14 species and demonstrates the workflow on a melon sex-determination locus, where an LLM agent given Botanic1 as a callable scoring tool ranks the validated CmEIN3 substitution first among thousands of candidates.

#Impact

BOTANIC-1 argues that careful corpus construction and species breadth buy more on plant tasks than raw scale: a 318M model trained on 314.6B tokens outscores genomic models an order of magnitude larger. Publishing the pretraining mixture with its provenance, rather than only the weights, makes the data ablations that drive that result reproducible, which is rare in this field. The release is gated and non-commercial, and the XL checkpoint listed in the paper's availability section is not public, so practical adoption rests on the S, M and L weights. As a preprint from a commercial lab with a declared financial interest, its benchmark claims await independent replication.