Genome annotation model that labels 14 genic and regulatory element classes at single-nucleotide resolution across DNA windows up to 50 kb.
No providers recorded yet. Browse all providers
Genome annotation — marking where genes, exons, splice sites, promoters, and enhancers sit along a chromosome — has long relied on a patchwork of specialist tools, each trained from scratch on one element class and a modest labeled dataset. That caps accuracy and transfers poorly to newly sequenced genomes, a growing problem as sequencing outpaces curation.
SegmentNT, built at InstaDeep with collaborators at BioNTech and the University of Copenhagen, replaces the patchwork with one model. It reframes annotation as instance segmentation, the task computer vision uses to label every pixel in an image, and applies it to DNA: every nucleotide in an input sequence receives a probability for each of 14 genic and regulatory element classes at once. Because the labels are not mutually exclusive, nested elements — an exon inside a protein-coding gene inside a promoter-proximal region — are handled natively.
The model attaches a one-dimensional U-Net decoder to a pretrained Nucleotide Transformer backbone, which supplies self-supervised representations learned from genomes across hundreds of species. The same framework was applied to two other DNA foundation models, Enformer and Borzoi, producing SegmentEnformer and SegmentBorzoi and showing the recipe is not specific to one encoder.
SegmentNT pairs the 500M-parameter nucleotide-transformer-v2-multi-species encoder with a 1D U-Net head of two downsampling and two upsampling convolutional blocks — roughly 53M additional parameters, 562M in total. Labels came from GENCODE V44 for gene elements and ENCODE's SCREEN registry for regulatory elements, with chromosomes 20 and 21 held out for testing and chromosome 22 for validation. Context was extended by curriculum: the 3 kb model saw 10.24 billion tokens, and successive checkpoints were retrained at 10 kb, 20 kb, and 30 kb.
Averaged over the 14 classes, SegmentNT-30kb reaches a Matthews correlation coefficient of 0.45, rising to 0.47 when inference runs on 50 kb sequences. Pretraining is what carries the result: at 3 kb context the model averages 0.37 against 0.16 for the identical architecture with a randomly initialized encoder, and 0.07 to 0.11 for one-hot U-Net baselines trained from scratch. On the SpliceAI test set it comes close to dedicated splice predictors, with auPRC of 0.93 for both donor and acceptor sites against 0.94 and 0.96 for SpliceAI and 0.94 for Pangolin on both. SegmentEnformer and SegmentBorzoi average 0.34 and 0.35 across the same elements. The multispecies variant averages 0.64 on animals close to human, 0.57 on distant animals, and 0.45 on plants.
The most direct use is annotating newly sequenced genomes: a lab working on a non-model organism can obtain gene structures and regulatory elements without assembling a species-specific training set, and the multispecies model beat the HMM gene finder AUGUSTUS on gene annotation in every species tested except Arabidopsis. In human genomics, the per-nucleotide outputs support asking what a candidate sequence does — screening synthetic constructs for cryptic splice sites, checking whether a designed regulatory element carries promoter or enhancer signatures, or annotating regions where curated coverage is thin. Running from sequence alone, with no alignments or expression data, it fits in-silico design loops that must score many variants quickly.
SegmentNT is among the clearest demonstrations that DNA foundation-model pretraining pays off on dense structured prediction rather than only on sequence-level classification: the gap between the pretrained and randomly initialized encoders is the paper's central result. The work also established segmentation as a reusable head for genomic backbones, with SegmentEnformer and SegmentBorzoi built by swapping the encoder alone. Two constraints bound practical use. The code and the two SegmentNT checkpoints — the human model and the multispecies model — are released under CC BY-NC-SA 4.0, which excludes commercial application, while the SegmentEnformer and SegmentBorzoi checkpoints carry no license tag at all on HuggingFace, leaving their reuse terms undefined. And the labels are human-derived, so element definitions carried into distant species inherit human annotation conventions — reflected in weaker scores on the diffuse classes, lncRNAs and CTCF-bound sites, even within human.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.