PlantHelixSeek-Anno

PlantHelixSeek-Anno

Fine-tuned from PlantHelixSeek for nucleotide-resolution gene structure prediction. Each nucleotide is assigned one of 17 BILOU labels covering CDS, intron, 5' UTR, 3' UTR and intergenic regions.

Task

• Type: token classification (17-class BILOU)
• Labels: O; B/I/L/U-CDS; B/I/L/U-INTRON; B/I/L/U-UTR5; B/I/L/U-UTR3
• Input: 8,192 bp windows with a 4,096 bp step

Training data

Gene models from 13 plant species were converted from GFF3 annotations into BILOU labels. Training and validation chromosomes were separated at the chromosome level. Intergenic windows were down-sampled to approximately 15% of the training set. See the manuscript Methods for the gene-filtering criteria.

Training

Reported performance

From the manuscript (Figure 4, Supplemental Figure 12 and Supplemental Tables 5 and 9):

Usage

from transformers import AutoModelForTokenClassification, AutoTokenizer

model = AutoModelForTokenClassification.from_pretrained(
    "zhangtaolab/PlantHelixSeek-Anno", trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(
    "zhangtaolab/PlantHelixSeek-Anno", trust_remote_code=True
)

The genome-wide prediction and GFF3 decoding pipeline is in scripts/gene_annotation/.

Citation

Wu Y, He Y, Chen L, Liu Y, Wang T, Zhao H, Zhang Y, Zhang T. PlantHelixSeek learns transferable genome representations that capture conserved genome grammar for functional interpretation across plant species. Molecular Plant, 2026. Manuscript MOLECULAR-PLANT-D-26-00999 (DOI to be assigned).

License

CC-BY-NC 4.0 (research use; contact the authors for commercial licensing).