Vipsania Web Server

Vipsania is an unsupervised deep learning ab initio gene finder for eukaryotic genomes. Unlike other deep learning gene finders, it is never shown a reference annotation: it is trained on raw genomic sequence alone with a masked language modelling objective, and gene structure emerges from a hidden Markov model inside the network. No extrinsic evidence is needed to annotate a genome.

Upload a genome FASTA file, select the clade that best matches your organism, and receive a gene annotation in GFF3 format. No registration required.

No finetuning on this server. Vipsania can be finetuned on the genome it is about to annotate (vipsania annotate --finetune), and its authors recommend doing so. This web server does not offer finetuning, because we do not have the GPU capacity for it: every job is annotated with the pretrained clade model as it is. Finetuning may make the annotation somewhat more accurate. If you need that, please run Vipsania locally with --finetune; a GPU is strongly recommended for finetuning.

Note: Accepted file types are .fa.gz and .fasta.gz (gzip-compressed FASTA, max 500 MB, genomes up to 300 Mbp). Results are stored for 45 days; email addresses deleted after the result notification is sent.

Genome FASTA * Help
Clade model * Help	Choose the most specific clade that contains your organism. The pretrained model of that clade is used as it is (no finetuning).
Email Help	Optional. If provided, results and error notifications are sent to this address.
Help	I confirm that the uploaded genome does not contain personalized human genomic data. *
Help	I consent to the processing of my personal data (uploaded genome sequence and, if provided, email address) in accordance with the Data Privacy Protection Declaration. *

* required field

Read the tutorial

Output files

GFF3, coding sequences, and protein sequences (all gzip-compressed).

Runtime estimates

Runtime depends on genome size and resource availability. Jobs are automatically routed to GPU nodes when available (substantially faster) or CPU nodes otherwise. In our measurements, genomes of 27–73 Mbp took 2–7 min on GPU and 1–7 h on CPU. Because any job may run on CPU, this server accepts genomes of up to 300 Mbp; please run Vipsania locally for larger genomes.

Citation

If you use results from this server, please cite:

Krieg R., Becker F., Saenko S., Diehl J., Stanke M. (2026). Vipsania: Unsupervised Deep Gene Finding. bioRxiv. https://doi.org/10.64898/2026.08.26.747235

Source code

Vipsania is open-source software (license: MIT).
→ github.com/gaius-augustus/vipsania

Funding

This web server was developed with support from the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation), Projektnummer 552910312 (project AI-GUSTUS).