henhaobio
Sign in

bioRxivBioinformaticsPreprint

Accurate ab initio gene prediction in eukaryotes with Tiberius in multiple clades

Lars GabrieliD, Tomáš BrůnaiD, Asees KauriD, Anish KrishnaniD, Felix Ortmann, Asaf A. SalamoviD, Samuel TalbotiD, Felix BeckeriD, Richard KriegiD, Christopher West WheatiD, Igor V. GrigorieviD, Mario StankeiD, Katharina J. HoffiD

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

Abstract

Background: Eukaryotic genome annotation is currently bottlenecked by limitations in the generality, scalability or accuracy of computational methods and fewer than 20% of the genomes available at NCBI Datasets have associated gene annotations. Evidence-based pipelines such as BRAKER3 achieve high accuracy but require substantial extrinsic evidence and compute. Deep learning approaches have recently achieved large improvements in ab initio gene prediction accuracy. The end-to-end deep learning gene predictor Tiberius approaches the accuracy of evidence-based annotation on mammalian genomes without using any extrinsic evidence, but its published models were trained on mammals only, limiting its applicability across eukaryotes. Results: We extend Tiberius beyond mammals by training lineage-specific models for Mesangiospermae, Fungi, Vertebrata, Insecta, Chlorophyta and Bacillariophyta, covering clades that account for 92% of currently available eukaryotic assemblies. Across a benchmark of 33 species, Tiberius achieved higher exon-, transcript- and gene-level accuracy than the other evaluated ab initio methods, Helixer and ANNEVO, improving gene-level F1 score by 12-37 percentage points over Helixer and by 10-22 percentage points over ANNEVO, while also having the fastest runtimes overall. Compared with BRAKER3, which incorporates RNA-Seq and protein evidence, Tiberius approaches state-of-the-art accuracy in Mesangiospermae, Fungi, Bacillariophyta and Chlorophyta, while being on average 80 times faster when using a GPU. A reimplementation of the Tiberius backend reduced runtime by 31% compared to the previous implementation. Tiberius and its models are also available through a web server, which allows annotation of submitted assemblies without local resources. Furthermore, the Vertebrata model has been applied to annotate 2,948 vertebrate assemblies totalling nearly 6 trillion base pairs. Conclusions: Tiberius is transferable well beyond Mammalia and reaches accuracy close to evidence-based annotation in several clades at a small fraction of the compute cost. This makes it a practical choice for highly accurate large-scale genome annotation, particularly when extrinsic evidence is unavailable or annotation throughput is limiting. Availability and implementation: Code: https://github.com/Gaius-Augustus/Tiberius, web server: https://bioinf.uni-greifswald.de/tiberius.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 3 (1 Oct 2026) · CC BY 4.0

Extended Tiberius models outscore Helixer and ANNEVO on all 33 genomes from the five clades all three tools cover, and approach BRAKER3 accuracy in four clades without extrinsic evidence.

Fewer than 20% of genomes at NCBI Datasets have gene annotations; evidence-based pipelines such as BRAKER3 need RNA-Seq, protein evidence and heavy compute; and the published Tiberius models covered only mammals. This work extends Tiberius to six more clades that hold 92% of available eukaryotic assemblies, including the first deep learning gene predictor for Bacillariophyta (protists).

Key results

  1. ShownIn the five clades where all three ab initio tools have models, Tiberius beat Helixer and ANNEVO at CDS-exon, transcript and gene level, and it did so on every individual test genome.

    Over 33 test species, gene-level F1 exceeded Helixer by 12-37 (per-clade averages) and ANNEVO by 10-22 percentage points. Exon-level F1 was about 5 points higher on average. Two-sided Wilcoxon signed-rank tests (n = 33 genomes) were significant at all three levels. Helixer had 17 of these species in its training or validation data and ANNEVO had two; Tiberius had none. · 3.1 Accuracy comparison with other ab initio…; Abstract

  2. ShownBRAKER3, which uses RNA-Seq and protein evidence, stays more accurate overall. Tiberius uses no extrinsic evidence but comes within three gene-level F1 points of it in Mesangiospermae, Fungi, Bacillariophyta and Chlorophyta.

    Overall, BRAKER3 led by about 12 gene-level and nine transcript-level F1 points. The gap was largest in non-mammalian Vertebrata (about 19 and 15). In the four clades named, gene-level F1 differed by under three points, and Tiberius led transcript-level F1 in Bacillariophyta by five points. Exon-level F1 averaged at most 1 point higher for Tiberius. · 3.2 Accuracy comparison with BRAKER3; Abstract

  3. ShownTiberius was the fastest tool overall, and with a GPU it was about 80 times faster than BRAKER3 on average. A rewritten backend cut its runtime by 31% without loss of accuracy.

    Mean runtime per genome over 33 species: 26 min Tiberius, 30 min ANNEVO, 178 min Helixer, 2170 min BRAKER3 (CPU only; the other three also used an NVIDIA A100 GPU). In non-mammalian Vertebrata, BRAKER3 averaged 3035 min per genome, while Tiberius and ANNEVO stayed under 60 min. The reimplementation cut mean runtime from 35 to 24 min over 37 species at matching accuracy. · Abstract; 3.4 Runtime comparison; 3.5 Effect of implementation changes; Tables

  4. ShownMost of Tiberius' advantage comes from long and multi-exon genes. Its Insecta model, however, heavily overpredicts single-exon genes.

    Helixer and ANNEVO declined most for genes with CDS over 2 kb and with more CDS exons; Tiberius and BRAKER3 held up. For Insecta single-exon genes, Tiberius precision was about 18% vs about 74% for ANNEVO. Fusions hit under 3% of reference loci except in P. striiformis (7.8%). Tiberius split 3.7% of reference genes vs 3.1% for ANNEVO. · 3.3 Characterization of prediction errors

  5. ShownThe Vertebrata model has been run on 2,948 public vertebrate assemblies. The annotations are released for download, and 2,100 can be browsed in the UCSC Genome Browser.

    The assemblies total 5,987,909,637,336 bp: 362 primates, 1120 other mammals, 497 birds, 650 fishes and 319 other vertebrates. · 3.6 Large-scale vertebrate annotation; 8 Data availability

Methods, data and reproducibility

Lineage-specific Tiberius models (an end-to-end deep learning gene finder with a differentiable HMM layer) were trained on unmasked genomes for six clades. They were benchmarked on held-out species against Helixer, ANNEVO and evidence-based BRAKER3. Accuracy was scored with gffcompare as exact CDS-exon, transcript and gene matches to reference annotations, allowing 3 bp tolerance at the outer CDS ends.

  • Held-out test genomes with reference annotations from NCBI RefSeq, Ensembl, GenBank, Phytozome or genome projects (Supplementary Tables S1-S9) · Mesangiospermae, Fungi, Vertebrata, Insecta, Chlorophyta, Bacillariophyta · n = 37 test species (33 in the four-tool comparison)
  • Training and validation genomes per clade, unmasked; Insecta and Bacillariophyta supplemented with BRAKER annotations from the authors' group · Mesangiospermae, Fungi, Vertebrata, Insecta, Chlorophyta, Bacillariophyta
  • Mesangiospermae assemblies and annotations not publicly released at the time of the study · Mesangiospermae · 10.6084/m9.figshare.32086728
  • BRAKER3 evidence: RNA-Seq input generated using VARUS; protein evidence from the combined proteomes of the Tiberius training species in each clade · multiple eukaryotic clades
  • Public vertebrate assemblies annotated with the Tiberius Vertebrata model · Vertebrata (primates, other mammals, birds, fishes, other vertebrates) · n = 2,948 assemblies (5,987,909,637,336 bp)
  • Eukaryotic assemblies at NCBI Datasets (retrieved 2026-03-25), used to estimate clade coverage · eukaryotes · n = 60,750 assemblies, 55,786 (91.8%) in covered clades
Methods & tools
Tiberius v2.0.0, bricks2marble, hidten, Helixer, ANNEVO, BRAKER3, VARUS, gffcompare v0.12.10, DIAMOND blastp, UniProt SwissProt, TimeTree, TensorFlow
Software & environment
Tiberius v2.0.0 (MIT licence) was used for all analyses. The reimplemented backend is split into the bricks2marble and hidten packages and runs on TensorFlow versions above 2.13. The seven clade-specific pretrained model archives are split across the code and data lists.
Assay conditions & compute
Benchmark: 72 threads of an AMD EPYC 7773X 64-Core Processor, plus an NVIDIA A100-SXM4-80GB GPU for the deep learning tools (BRAKER3 ran on CPU only). The web server uses an A100 GPU node or a CPU node with 24 cores and 63 GB RAM. Training hardware is in Supplementary Tables S29-S30.
Reproducibility gaps
Some Mesangiospermae training data are unreleased; the Figshare copy falls under the Fort Lauderdale Accord (contact providers before independent use). Web server: gzipped assemblies up to 500 Mb, results kept 45 days. Accessions only in Supplementary Tables S1-S9.
Limitations (6)
  • Tiberius and the other deep learning ab initio tools predict one isoform per locus. This limits transcript-level accuracy and gives BRAKER3 a systematic edge at gene level. Reconstructing complete gene structures in complex genomes remains the main weakness, not exon-boundary detection. · stated by the authors
  • Coverage of 92% of assemblies reflects sequencing effort, not species diversity, and does not mean uniform accuracy. Accuracy varies with phylogenetic distance, repeat content and gene architecture. Trans-splicing, non-canonical splice sites and very short introns or exons were not specifically tested. · stated by the authors
  • Accuracy is limited by how many reference annotations exist and how good they are. Chlorophyta training data are sparse and heterogeneous; Fungi annotations are heterogeneous across a diverse clade; Insecta training data were dominated by BRAKER annotations, which may explain the single-exon overprediction. · stated by the authors
  • The comparisons are confounded. Helixer had 17 and ANNEVO two of the 33 test species in their training or validation data, which favours them. The runtime comparison with BRAKER3 is also uneven, because only the deep learning tools used a GPU. · stated by the authors
  • Chlorophyta results rest on two test species, with Helixer winning on one and Tiberius on the other. Bacillariophyta was compared only with BRAKER3, so the evidence for these clades is thin. · noted in this summary
  • No accuracy check is reported for the 2,948 released vertebrate annotations, although the benchmark shows weak results on distant, repeat-rich vertebrates such as the 5.9 Gb Pristiophorus japonicus genome. · noted in this summary

Read this if you annotate protein-coding genes across many eukaryotic genomes without RNA-Seq or protein evidence and want to know how close ab initio deep learning now gets to BRAKER3, clade by clade.

Citations
0

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 '24Gabriel '24Brůna '21Zimin '26Brůna '24Krieg '26
74% semantic matchBioinformatics
Lars Gabriel, Felix Becker, Katharina J. Hoff+1 more·Bioinformatics·Nov 13, 2024· 45 citations