PlantGFM

Editing a plant gene with CRISPR or mutagenesis rearranges sequence a genome already carries; it cannot write a gene no genome has ever carried. PlantGFM was built to test whether a generative model of plant DNA can. It is a decoder-only genomic foundation model that reads and writes nucleotide sequence at single-base resolution across a 64 kb window — long enough to hold an entire plant locus, introns and flanking regions included.

Single-nucleotide resolution and long context normally pull against each other: at one token per base a 64 kb window is roughly 64,000 tokens, and the cost of self-attention grows with the square of that. PlantGFM keeps the resolution and buys the context by replacing the multi-head attention of a LLaMA-style decoder with the Hyena operator, which captures local and global dependencies through gated long and short convolutions rather than explicit query–key products. Complexity drops from O(n²) to O(n log n), and memory falls far enough to make 64 kb pretraining practical. Because the objective is plain next-token prediction, the same checkpoint that scores a genomic window can also emit one.

PlantGFM was developed at Hubei Hongshan Laboratory with Huazhong Agricultural University, Yazhouwan National Laboratory and Huazhong University of Science and Technology, and published in Advanced Science in 2026. It is a sibling of the same group's PlantGeneAnn, which targets annotation alone. Early releases used the name PlantGLM, and the group's repository under that name now redirects to this one.

#Key Features

• Whole-locus context at base resolution: A 64 kb window tokenized one nucleotide at a time covers a complete plant gene with its introns and regulatory flanks, without the k-mer merging that blurs single-base signals.
• Prediction and generation in one checkpoint: The autoregressive objective covers both scoring genomic windows and sampling new ones, so annotation and sequence design draw on the same pretrained weights.
• Cross-species annotation: Paired with a U-Net segmentation head, it labels gene, exon and CDS regions in species it never saw annotated during fine-tuning.
• Experimentally validated de novo genes: Prompt-based retraining on natural plant genes produced candidates that plant cells transcribe, splice and in some cases translate.
• Function-guided family design: Retraining on canonical NLR immune-receptor genes with an "NLR" prompt yields candidates carrying the Rx_N, NB-ARC and LRR domain architecture.

#Technical Details

The architecture is a token embedding layer, a positional encoding module, 16 Hyena blocks (each combining an MLP, RMSNorm and the Hyena operator) and an output projection head, totaling roughly 220 million parameters. Pretraining ran autoregressively over the complete genomes of 12 model plant species — 10.84 billion nucleotides — using a three-stage length warm-up that began at 1 kb fragments, moved to 8 kb, and finished at 64 kb, converging in 16 epochs.

For gene prediction, PlantGFM was fine-tuned on ten annotated genomes with Arabidopsis thaliana and Brassica rapa held out for checkpoint selection. Across four unseen species it reached a mean Matthews correlation coefficient of 0.771, ahead of Helixer at 0.765 and far ahead of AUGUSTUS (0.548) and SegmentNT (0.402), while trailing the specialist ANNEVO at 0.779. A matched-token ablation at 60 kb isolated the two design choices: PlantGFM scored 0.764 against 0.721 for GPN and lower still for k-mer-tokenized baselines. Regulatory tasks transfer by fine-tuning — mean Pearson correlation 0.739 for gene expression across four species, and mean AUROC 0.988 with AUPRC 0.574 for chromatin accessibility over 19 maize tissues.

#Applications

The annotation head suits crop genomics groups working on newly sequenced or non-model species, where RNA-seq and homology evidence are thin and conventional pipelines degrade. The generative half is aimed at plant synthetic biology: sampling candidate genes outside the natural sequence space, then filtering them computationally before synthesis. The NLR results point at disease-resistance breeding, where the supply of natural resistance genes limits what can be deployed. Fine-tuned heads cover the routine regulatory-genomics workload — expression, chromatin accessibility, transcription-factor binding sites, cis-regulatory element strength — on the same backbone.

#Impact

The paper's central result is wet-lab, not computational. The generation model produced 3,000 candidates in three batches of 1,000; an AI–human screening pipeline narrowed one such batch to 454, then 92, then 30, and seven of those were synthesized and transiently expressed in Nicotiana benthamiana. All seven were transcribed — 2- to 7-fold above the empty-vector control by RNA-seq, 11- to 220-fold by RT-qPCR, with one locus producing multiple spliced isoforms — and two accumulated stable protein detectable by GFP fluorescence and immunoblot. None shared more than 10% nucleotide identity with a natural gene. That chain from generated DNA through RNA to protein had not been shown in plants before. The limits sit in the same numbers: two of seven reached protein, annotation still ranks behind dedicated tools, and every downstream task needs fine-tuning rather than zero-shot use. Code and all checkpoints are public, but the repository carries an MIT badge with no license file and the model repositories declare no license, leaving reuse terms unsettled.