A 320-genome corpus spanning the land plants

Earlier studies all trained on fewer than 100 species (42 plant genomes for PlantBiMoE, 48 crop genomes for AgroNT and Botanic0, 65 for PlantCAD2), whereas the new Model Factory employs a 320-genome corpus spanning the land plants, thus yielding, to our knowledge, the broadest taxonomic coverage of any plant gLM to date.

The filters of the data pipeline leave 322 training and 4 held-out genomes (one per species); 320 of the training species contribute at least one window, for 7,207,506 windows and 59.0 Gbp of training sequence at 8,192 bp. Although the corpus spans 48 orders and 102 families, its coverage is strongly concentrated within angiosperms: 310 of the 320 species belong to the angiosperm crown group, with 241 eudicots and 61 monocots.

Windows per species

Every rectangle is one species, sized by its number of 8,192 bp training windows and grouped by order. Colour is genome size, which the length weighting below deliberately keeps from dictating a species' share of the corpus.

Correcting for genome length

The smallest genome in the corpus, Arabidopsis thaliana (0.12 Gbp), receives a weight of 1; the weight decreases with genome length until it reaches the floor of 0.5 at 0.48 Gbp, after which all larger genomes receive the same weight. Thus, for example, a 0.5 Gbp genome and the 11.9 Gbp genome of Vicia faba receive the same correction despite having widely different lengths. As a result, the correction reduces the tendency for larger genomes to contribute more windows but without erasing it completely. Empirically, doubling genome length increases a species' share of the 8 kbp windows by about 39% on average, compared with 48% without correction. Species with genomes larger than 2 Gbp therefore still occupy a larger median fraction of the corpus (0.50%) than species below 0.48 Gbp (0.19%), whereas perfectly uniform sampling across these 306 species would assign 0.33% to each.

Near-duplicate windows

Windows containing more than 20% N bases are removed. Near-duplicate windows are then filtered within each selected assembly, removing near-copies of the same window within species, but not orthologous sequence between species. Two windows are compared using the fraction of k-mers they have in common (Jaccard similarity) estimated with a per-window MinHash sketch (128 smallest hash values, each k-mer is canonicalised with its reverse complement so that strand orientation alone does not make two windows different). A window is discarded when its estimated Jaccard similarity to any window in the current set reaches 0.3. This removes roughly 15 to 25% of the sampled windows depending on the species pool and recipe; on the full 330-species ablation pool it removes 20.6% (65.6 to 52.1 Gbp).

Orders

The five context-extension datasets

To expose the model to longer contexts, we create five datasets with window sizes 8,192, 16,384, 32,768, 65,536, and 131,072 bp. Those are produced by re-running our data pipeline at the target window length and increasing the contig-N50 floor accordingly. This limit rises faster than assembly contiguity improves, causing fewer species to qualify at each step. The 128 kbp dataset requires at least one long-read assembly per species, which is why it has the fewest species. Total base pairs per dataset range between 59 and 116 Gbp: fewer but longer windows offset the shrinking species pool.

The 8 kbp corpus is released as the Botanic1-pretraining dataset, with the original sequence case, augmentation margins, train/test separation, source manifests and assembly metadata.