Loading…
Phylogeny-aware genome annotation model predicting exons, introns, UTRs and repeats directly from eukaryotic DNA, with no RNA or protein evidence.
No providers recorded yet. Browse all providers
A freshly assembled genome is a string of nucleotides with nothing marked on it. Turning it into a usable resource has traditionally meant training a species-specific hidden Markov model in the AUGUSTUS mould, or running an evidence pipeline such as BRAKER3 that needs RNA-seq reads and homologous proteins from a close relative. Both assume resources newly sequenced organisms rarely have, which is why roughly half of NCBI's reference genomes still carry no annotation.
OrionGeno, developed by BGI Research with the School of Artificial Intelligence at the University of Chinese Academy of Sciences, closes that gap by making species identity an input rather than a training condition. The complete NCBI taxonomic hierarchy — Kingdom through Species — for 22,444 eukaryotes is written out as structured text and embedded with the Qwen-7B language model, giving a 22,444 × 4,096 matrix that principal component analysis compresses to one 384-dimensional vector per species. That vector is fused with convolutional features of the input DNA, so annotating an organism the model has never seen means supplying its lineage string rather than fitting anything new. The model learns the grammar shared across eukaryotes while keeping lineage-specific variation in intron length and repeat composition distinguishable.
From a genome FASTA alone, it emits GTF gene models — exons, introns, splice sites, start and stop codons, both UTRs — alongside a base-resolution repeat track, removing the separate repeat-masking step annotation pipelines normally run first. The first preprint appeared in April 2026; the August 2026 revision scales the work past 5,300 genomes.
OrionGeno is a roughly 121-million-parameter hybrid: a three-stage down-sampling convolutional encoder, a stack of bidirectional Mamba (BiMamba) blocks holding about 74% of the weights, and a mirrored convolutional decoder with skip connections that restores single-base resolution. Training runs in two stages over the same backbone: BERT-style masked language modelling on genomic sequence, then supervised multi-task fine-tuning under a combined cross-entropy and F1-oriented loss built for the sparsity of exons in a largely non-coding background. Training used assemblies from 1,282 species — 393 vertebrates, 185 invertebrates, 99 Viridiplantae, 556 fungi and 49 protists — drawn mainly from RefSeq. On a benchmark of 36 well-curated species strictly excluded from training, OrionGeno improves exon-level F1 by 12.8% over ANNEVO and gene-level F1 by 32.0% over the best competing method, and raises BUSCO proteome completeness by 6.6%. Its protein-level advantage widens with alignment stringency: 4.3% at 60% coverage, 37.2% when predicted proteins must match references exactly. Folding predicted and reference proteomes with ESMFold gave mean TM-scores of 0.85, 0.93 and 0.89 for human, Arabidopsis thaliana and zebrafish.
The authors ran the model over 5,310 unannotated chromosome-level NCBI assemblies — a week on 40 GPUs — reaching a median BUSCO completeness of 97.3%, with 85.9% of genomes above 90%, and supplying gene models for 718 taxonomic families that had no publicly annotated representative — a 47.4% lift in family-level coverage. Biodiversity consortia, crop and livestock groups, and anyone working on an organism without transcriptome support can obtain gene models from an assembly alone, locally or through the hosted CNGBdb service. Predictions also pay off on well-curated genomes: EggNOG analysis surfaced seed orthologs missing from reference protein-coding annotation in human, zebrafish and zebra finch, of which 86, 64 and 52 loci showed transcriptional support above 5 TPM.
Learned annotation has arrived from two directions — genomic language models such as SegmentNT and GENATATOR that classify sequence at base resolution, and supervised clade-specific gene predictors. OrionGeno shows that taxonomic conditioning can cover the eukaryotic tree while still producing structurally coherent gene models. From sequence alone it rivalled the evidence-integrating BRAKER3 pipeline, a 4.4% relative BUSCO gain across six species with deep public RNA-seq. The limitations are real: accuracy falls in isolated lineages such as Hydrozoa and Trebouxiophyceae where few close references exist, UTRs stay harder to recover than coding regions, and only the longest coding transcript per locus is reported, so alternative splicing is not represented. Code and weights carry a custom academic, non-commercial licence, and the work remains a preprint awaiting peer review.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.