henhaobio
Sign in

Nature MethodsBioinformaticsPeer-reviewed journal article

Helixer--de novo Prediction of Primary Eukaryotic Gene Models Combining Deep Learning and a Hidden Markov Model

Felix Holst, Anthony Bolger, Felicitas KindeliD, Christopher Günther, Janina MaßiD, Sebastian ST TrieschiD, Niklas KieliD, Nima P. SaadatiD, Oliver EbenhöhiD, Björn UsadeliD, Rainer Schwacke, Andreas P.M. WeberiD, Marie BolgeriD, Alisandra K. DentoniD

PDFDOIPreprintCode
In-depth full-text digest available · Jump to digest ↓

Abstract

AO_SCPLOWBSTRACTC_SCPLOWGene structural annotation is a critical step in obtaining biological knowledge from genome sequences yet remains a major challenge in genomics projects. Current de novo Hidden Markov Models are limited in their capacity to model biological complexity; while current pipelines are resource-intensive and their results vary in quality with the available extrinsic data. Here, we build on our previous work in applying Deep Learning to gene calling to make a fully applicable, fast and user friendly tool for predicting primary gene models from DNA sequence alone. The quality is state-of-the-art, with predictions scoring closer by most measures to the references than to predictions from other de novo tools. Helixers predictions can be used as is or could be integrated in pipelines to boost quality further. Moreover, there is substantial potential for further improvements and advancements in gene calling with Deep Learning. Helixer is open source and available at https://github.com/weberlab-hhu/Helixer A web interface is available at https://www.plabipd.de/helixer_main.html

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (9 Feb 2023) · CC BY 4.0

Helixer, a hybrid convolutional/bLSTM network with a hidden Markov model post-processor, predicts eukaryotic primary gene models from DNA alone, outscoring GenemarkES and AUGUSTUS on most species.

Existing de novo hidden Markov model (HMM) gene callers cannot capture biological complexity alone, so annotation relies on resource-hungry, evidence-dependent pipelines with uneven quality, and only circa 24% of recent eukaryotic NCBI assemblies carry annotations. Earlier Helixer gave only base-wise genic-class calls; this version adds phase prediction and an HMM post-processor to emit finished gff3 gene models without extrinsic data or species-specific retraining.

Key results

  1. ShownHelixerPost gene models beat both de novo HMM baselines on base-level gene structure for nearly all test species, and HMM post-processing preserved, and sometimes slightly raised, the accuracy of the raw network calls.

    Subgenic F1 of HelixerPost exceeded GenemarkES in 42 of 45 species (average margin 0.38) and AUGUSTUS in 14 of 15 species (average margin 0.17); post-processed models scored slightly above raw Helixer predictions in 33 of 45 test species. · 4.1 The released Helixer models show…; F1 scores

  2. ShownPredicted proteomes were nearly as complete as the reference annotations and usually more complete than AUGUSTUS or GenemarkES output; HelixerPost trailed GenemarkES only on three invertebrates whose reference annotations had the lowest BUSCO counts.

    BUSCO: reference highest in 39 of 45 species; mean shortfall in BUSCOs vs reference 3.7% (HelixerPost), 12.8% (AUGUSTUS), 33.9% (GenemarkES); HelixerPost found more BUSCOs than GenemarkES in 41 of 45 and AUGUSTUS in 14 of 15 species, but trailed GenemarkES on three invertebrates. · 4.1 The released Helixer models show…; Homology comparisons

  3. ShownOn plant test genomes, homology-based quality metrics place Helixer's DNA-only proteomes closer to the evidence-supported references than to GenemarkES. Helixer has higher gene-family recall than the references but lower precision.

    Orthogroups covering all 12 plant species: 0.38% reference, 0.26% Helixer, 0.019% Genemark; covering two or more species: 36.5%, 31.1%, 21.2%. Mapman4 precision/recall/harmonic mean: reference 0.966/0.931/0.948, Helixer 0.878/0.958/0.914, Genemark 0.719/0.742/0.724. · 4.3 Helixer’s annotations approach reference…; Homology comparisons

  4. ShownIn Arabidopsis thaliana, Helixer matched most Araport11 loci exactly and recovered conserved genes absent from the curated references, for example an expressed Phosphatidylinositol N-acetylglucosaminyltransferase γ subunit missing from TAIR10 and merged with a neighbour in Araport11.

    Roughly 70% of Araport11 loci had an exact Helixer match; of the remaining roughly 30%, around 27% were highly similar but differed in length. Among unmatched loci, 102 of 711 Helixer vs 93 of 1401 Araport11 annotations were conserved in at least 20 streptophytes. The γ subunit is absent from TAIR10 and fused to a neighbour in Araport11 (4bp out-of-frame overlap). · 4.4 Applicability for filling gaps in even good…; 3.6 Arabidopsis thaliana-annotation comparison

  5. ShownRuntime scales approximately linearly within each phylogenetic group, with slightly different coefficients between groups, and is much shorter than the HMM baselines in single-threaded comparisons, although Helixer relies on a GPU.

    Single-threaded walltime was 6.2-20.1 fold lower than GenemarkES and Augustus on the smallest and largest fungal and smallest plant test genomes; the 263Mbp Oryza brachyantha took 27min and the 3.3Gbp human genome took just under 8.5 hours; GenemarkES did not finish 10.7Gbp Triticum dicoccoides in 1 week on 24 threads. · 4.5 Benchmarking; 3.3 Annotation quality comparison; 3.8 Benchmarking

Methods, data and reproducibility

Convolutional layers followed by bidirectional LSTM layers map one-hot DNA to per-base genic class (intergenic, UTR, CDS, intron) and codon phase, trained with transition-weighted loss; HelixerPost then decodes genic regions by Viterbi over a 73-state HMM enforcing codon, splice-site and minimal-intron constraints, emitting gff3 primary gene models.

  • Annotated reference genomes from RefSeq and Phytozome13, split per lineage (vertebrates, invertebrates, land plants, fungi) into training, validation and test species · multiple eukaryotes (vertebrates, invertebrates, land plants, fungi)
  • Held-out test genomes for genome-wide Subgenic F1 and BUSCO evaluation against GenemarkES (all but Triticum dicoccoides) and AUGUSTUS (where trained models existed) · multiple eukaryotes · n = 45 test species (15 with AUGUSTUS models)
  • Plant test genomes for OrthoFinder orthogroup and Mapman4 gene-family evaluation, including Triticum dicoccoides and Papaver somniferum · land plants · n = 13 species (12 in orthogroup clustering)
  • Arabidopsis thaliana genome compared with the TAIR10 and Araport11 reference annotations, plus RNA-seq for locus-level expression · Arabidopsis thaliana
  • Proteomes from streptophyte genome projects used to verify orthologs of unmatched A. thaliana loci via reciprocal best hits · streptophytes (multiple plant species) · n = 420 genome projects
  • Runtime benchmark genomes including Oryza brachyantha and the human genome · Oryza brachyantha; human · n = 263Mbp and 3.3Gbp genomes
Methods & tools
Helixer (v0.3.0; hybrid convolutional/bLSTM network), HelixerPost (HMM post-processor with Viterbi decoding), SpeciesSelector (automated training-species selection), AUGUSTUS 3.3.2, GenemarkES 4.71_lic, BUSCO 5.2.2, Mercator4 5.0 / Mapman4, OrthoFinder 2.5.4, BLAST, Docker / Singularity
Software & environment
Docker or Singularity images bundle all three components plus a Helixer.py wrapper (fasta in, gff3 out); tag v0.3.0 used; baselines AUGUSTUS 3.3.2, GenemarkES 4.71_lic; BUSCO 5.2.2, Mercator4 5.0, OrthoFinder 2.5.4.
Assay conditions & compute
Single-threaded benchmarks on workstations with GeForce GTX 1080 Ti or GTX 1080 GPUs (AMD Ryzen 9 5950X or Intel Xeon CPUs, 32GiB or 64GiB RAM, SSD storage); human genome in just under 8.5 hours on workstation A.
Reproducibility gaps
No data URLs or accessions are given: genomes are cited only as RefSeq and Phytozome13, and species splits, parameters and per-species scores are in supplementary tables.
Limitations (6)
  • Speed comparisons were single-threaded; the authors note GenemarkES and AUGUSTUS support more complete multi-threading and need no GPU, and Helixer's runtime depends on disk and GPU speed. · stated by the authors
  • Ablation conclusions are tentative: each variant was trained once, without replicates or a fixed random seed, so effects of random initialization were not separated from the ablated change. · stated by the authors
  • Metrics treat reference annotations, themselves pipeline outputs, as ground truth, limiting assessment near reference quality; overall quality still trails evidence-supported pipelines, with lower precision on fragmented assemblies (Papaver somniferum, Triticum dicoccoides). · stated by the authors
  • Invertebrate models are the newest and least optimized; on three invertebrates with the weakest reference BUSCO counts, HelixerPost scored below GenemarkES on both Subgenic F1 and BUSCO completeness. · stated by the authors
  • AUGUSTUS was run only with pre-existing trained models (15 of 45 species, no retraining), and evidence-driven pipelines such as BRAKER2 or MAKER2 were not benchmarked, so comparisons cover only de novo HMM baselines. · noted in this summary
  • Helixer outputs one primary model per locus with no alternative isoforms, and in the highlighted Arabidopsis locus HelixerPost truncated UTR and intron features relative to RNA-seq, so isoform and UTR accuracy remain open. · noted in this summary

Read this if you need structural gene models for a newly assembled eukaryotic genome without RNA-seq or protein evidence, or want to weigh deep-learning gene calling against GenemarkES and AUGUSTUS.

Citations
160
FWCI
38.86
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISGabriel '24Krieg '26Zhang '20Shmelev '26Chan '17Zhang '20
67% semantic matchBioinformatics
Lars Gabriel, Felix Becker, Katharina J. Hoff+1 more·Bioinformatics·Nov 13, 2024· 45 citations