Vipsania

Every deep learning gene finder before Vipsania was taught what a gene looks like by being shown one. Helixer, Tiberius and ANNEVO all train on curated reference annotations, so their reach ends where those annotations do — and fewer than 20% of eukaryotic assemblies in NCBI Datasets carry one. The clades that most need an ab initio predictor, the eukaryotic algae above all, are exactly those with too few annotated genomes to train on.

Vipsania, from Mario Stanke's group at the University of Greifswald — the lab behind AUGUSTUS, BRAKER and Tiberius — removes the requirement. It is trained only to fill in masked nucleotides in raw sequence, and never sees an annotation. Gene structure emerges anyway: a differentiable hidden Markov layer sits mid-network, and the model can explain the sequence well only by routing it through that layer's grammar — reading frames maintained across introns, canonical splice sites, no in-frame stop codons inside exons. At inference a Viterbi path through those states decodes into a GFF3 annotation.

Because training needs nothing but sequence, the genome a user wants annotated is itself valid training data. An optional --finetune pass adapts the checkpoint to the target genome first and saves the adapted weights for reuse — an enhancement, not a precondition: the released checkpoints annotate held-out species on their own.

#Key Features

• No annotation anywhere in the loop: Training is masked-nucleotide prediction on unannotated genomes, so the model cannot inherit the errors of the evidence-based pipelines behind most reference annotations.
• Differentiable HMM inside the stack: The hidden Markov layer sits mid-network rather than bolted on as a decoder, so its structural constraint shapes the representations being learned rather than filtering them afterwards.
• Pan-eukaryotic checkpoint family: Seventeen clade models cover 99.4% of the 1,716,067 eukaryotic taxa in the NCBI Taxonomy, with an eighteenth general model for the rest. Clade names resolve to model IDs, fetched on first use.
• Unsupervised test-time adaptation: Finetuning needs only the target FASTA, takes roughly 160 minutes on an A100, and writes a reusable checkpoint next to the output.
• Non-standard genetic codes: Adapting to a species using NCBI translation table 6 amounts to removing TAG and TAA from the HMM's stop-codon emissions — a set hard-coded in ANNEVO and Helixer.

#Technical Details

Vipsania is a 25M-parameter model: 16 layers on an uninterrupted residual stream of width 320 at single-base-pair resolution, with attention replaced by Linear Recurrent Units and SwiGLU blocks, and no convolutional layers. The differentiable HMM sits after the eighth layer, with 18 states per phase and an emission alphabet of 160, run as two parameter-sharing copies for both strands. Training uses a BERT-style cross-entropy loss over up to 5% masked positions in a 20,000 bp context, extended by a "spliced loss" that upweights nucleotides near the positions the HMM currently believes are exon boundaries. The released checkpoints were pretrained on 910 Gbp from 1,559 genomes, typically for 1,000 epochs and around ten days on an A100.

On test species disjoint from each model's training species, the pretrained checkpoints alone reach average locus F1 of 0.696 on Discoba, 0.645 on Fungi and 0.579 on Streptophyta; finetuning adds a few points in almost every clade. Mean locus F1 exceeds the supervised deep gene finders in all clades but Vertebrata, and Vipsania matches or beats Tiberius on 38 of 60 test species, ANNEVO on 82 of 92, Helixer on 91 of 92, and the unsupervised GeneMark on 123 of 126. Average BUSCO completeness is 86.22%, against 80.35% for Tiberius and 65.78% for GeneMark.

#Applications

The intended user has a newly sequenced eukaryotic genome, no RNA-seq, and no well-annotated relative — increasingly the norm as sequencing outpaces annotation across the Earth BioGenome Project's target space. Vipsania produces a standard-format structural annotation from the assembly alone, which makes it usable for the basal clades, protists, algae and invertebrate lineages that supervised tools have no checkpoint for. In evidence-rich settings it is not expected to displace RNA-seq and homology pipelines, but complements them on genes those pipelines miss.

#Impact

The result that travels furthest concerns phylogenetic leakage. Supervised gene finders lose three to five percentage points of locus F1 for every ten-point drop in proteome identity between target and training species, while Vipsania's trend is significantly flatter, because it has no gene set to overfit to. That also qualifies a recent negative claim about DNA language models: GENATATOR's authors concluded pretrained embeddings do not capture the features needed for precise gene segmentation, whereas here unsupervised pretraining produces them, given the right inductive bias. Vertebrata remains the honest weak spot — long introns and intergenic regions are hard to pin down without labels, and it is the one clade where supervised methods win. The work is a bioRxiv preprint awaiting peer review.