Written and maintained by CASRAI Editorial Board
Published · Last updated
On this page
SignalP is the DTU Health Tech tool that predicts signal peptides and their cleavage sites from protein sequences. This guide covers the current version, SignalP 6.0: how to run it on the web server or from the command line, what each output file contains, and how to read the prediction codes. It was checked against the DTU server page and the SignalP 6.0 installation instructions on 2026-11-07.
What SignalP is used for in research
SignalP predicts whether a protein carries an N-terminal signal peptide and where it is cleaved. Version 6.0 predicts all five signal peptide types: Sec/SPI, Sec/SPII, Tat/SPI, Tat/SPII and Sec/SPIII. It covers Archaea, Gram-positive bacteria, Gram-negative bacteria and Eukarya, and is built on a transformer protein language model combined with a conditional random field (CRF) for region and cleavage-site prediction.
A signal peptide prediction is evidence about export, not a statement of final location. The DTU documentation states that a signal peptide alone does not establish where the mature protein ends up. For eukaryotes, SignalP 6.0 predicts only the Sec/SPI type. If you also need mitochondrial, chloroplast or luminal-thylakoid targeting, see the alternatives table below.
Which SignalP version to run and cite
The server page lists 6.0 as current, with older versions 5.0b, 4.1g, 3.0 and 2.0 still available. The differences matter when you compare results with a published paper:
- 6.0 adds Tat/SPII and Sec/SPIII and region prediction.
- 5.0 used CNN/LSTM with a CRF and added Sec/SPII, Tat/SPI and archaea.
- 4.1 added a D-score cutoff option; 4.0 improved discrimination of signal peptides from transmembrane regions.
- 3.0 introduced the D-score; 2.0 was HMM-based; 1.1 was the original neural-network server from 1997.
Cite SignalP 6.0 as Teufel F., Almagro Armenteros J.J., et al. (2022), Nature Biotechnology, doi:10.1038/s41587-021-01156-3. Record the version string you ran (for example 6.0i) and the mode (fast or slow) in your methods; see FAIR4RS for why software versions belong in a citation.
How to install or access it
Web server
The SignalP 6.0 web server is hosted by DTU Health Tech at services.healthtech.dtu.dk/services/SignalP-6.0/. It accepts protein sequences of at least 10 amino acids. The form states a per-submission sequence limit and a maximum sequence length of 10,000 amino acids; the page’s own instructions and form wording differ on the sequence count, so read the limit on the live form before a large batch. The server page notes that long output for more than about 100 entries may time out, so use Short output for big sets.
Python package
The same functionality is available as a Python package from the DTU academic download page. Non-academic users must contact the DTU Health Tech Software Package Manager; the page does not state licence terms, so confirm them with DTU before commercial use. The installation instructions (github.com/fteufel/signalp-6.0, installation_instructions.md) give these requirements: Python up to 3.10, an operating system supported by PyTorch, and about 1 to 5 minutes for installation. A GPU is optional.
Step-by-step: predict signal peptides
On the web server
- Open the SignalP 6.0 page and paste FASTA sequences or upload a FASTA file.
- Choose the organism group: Eukarya or Other. Prokaryotes need no further organism information.
- Choose the output format: Long (a plot and summary per sequence) or Short (one line per sequence, no graphics).
- Choose the mode: Fast (default, a distilled model) or Slow (the full model, about six times slower, recommended when you need accurate region borders).
- Submit, then download the results. The server offers JSON, a prediction summary, processed entries (FASTA with the signal peptide removed) and GFF3.
From the command line (Python package)
- Install the package from the downloaded directory:
pip install signalp-6-package/. - For version 6.0h only, also run
pip install "numpy<2". - Find the package directory with
python -c "import signalp, os; print(os.path.dirname(signalp.__file__))", then copy the contents of the downloadedmodels/folder into<signalp directory>/model_weights/. - Convert the model weights for your hardware:
signalp6_convert_models gpuorsignalp6_convert_models cpu. - Run a prediction:
signalp6 --fastafile in.fasta --organism other --output_dir out --format txt --mode fast
| Option | Short form | Values and default |
|---|---|---|
--fastafile |
-ff |
Input FASTA (required) |
--output_dir |
-od |
Output directory (required); created if missing, and reruns overwrite its files |
--organism |
-org |
other (default) or eukarya; eukarya restricts prediction to Sec/SPI |
--format |
-fmt |
txt (default), png, eps, all or none |
--mode |
-m |
fast (default) or slow-sequential (about six times slower, same RAM) |
--model_dir |
-md |
Directory holding the model weights |
--bsize |
-bs |
Batch size, default 10 |
--torch_num_threads |
-tt |
PyTorch threads, default 8 |
--write_procs |
-wp |
Writer processes, default the CPU count, maximum 8 |
SignalP output and how to read it
The main result is prediction_results.txt, a tab-delimited file with one row per sequence: the sequence ID, the predicted type (Prediction), a probability for each type, and the cleavage site position (CS Position). The Prediction column uses six codes:
| Code | Meaning |
|---|---|
OTHER |
No signal peptide predicted |
SP |
Standard secretory signal peptide (Sec/SPI) |
LIPO |
Lipoprotein signal peptide (Sec/SPII) |
TAT |
Tat signal peptide (Tat/SPI) |
TATLIPO |
Tat lipoprotein signal peptide (Tat/SPII) |
PILIN |
Pilin-like signal peptide (Sec/SPIII) |
To read a row, look first at the Prediction, then at that type’s probability against the others, then at the CS Position, which marks where the signal peptide is cut. A mature-protein sequence is easier to get from the processed-entries FASTA than by trimming the original yourself. The other files written are:
processed_entries.fasta: input sequences with the predicted signal peptide removed.output.gff3andregion_output.gff3: predictions and region annotations for genome browsers and pipelines.output.json: the same predictions in machine-readable form.- Optional per-sequence plots (png, eps or text) when you request a graphics format.
Check a surprising result against a curated record in UniProt or against domain and family annotations in InterPro before relying on it. To compute properties of the mature sequence, ProtParam works on the processed FASTA, and structure predictions can be browsed in the AlphaFold Protein Structure Database.
Troubleshooting
- The web job times out or returns nothing. Long output for more than about 100 entries may time out. Resubmit in smaller batches with Short output.
- A sequence is rejected. Input sequences must be at least 10 amino acids, and no longer than 10,000.
- A eukaryotic protein shows no lipoprotein, Tat or pilin type. With
--organism eukaryathe model predicts only Sec/SPI. Run withotheronly if the organism is not eukaryotic. - NumPy errors on 6.0h. The installation instructions require
pip install "numpy<2"for that release. - The model cannot be found. Confirm the model files were copied into
model_weights/inside the package directory and that you ransignalp6_convert_modelsfor gpu or cpu. - Python version problems. The instructions state Python up to 3.10. A clean environment helps; see sharing a reproducible conda environment.
- Results vanished after a rerun. Reruns into the same output directory overwrite earlier files; use a new directory for each run.
- Region borders look rough. Rerun with slow mode, which uses the full model.
Free alternatives and complements
These tools answer overlapping questions. Each row is taken from the tool’s own page or publication, read on 2026-11-07.
| Tool | What it predicts | Access | Reference |
|---|---|---|---|
| TargetP 2.0 | Signal peptides (SP), mitochondrial (mTP), chloroplast (cTP) and luminal thylakoid (lTP) transit peptides, with cleavage sites; Plant or Non-plant setting | Free DTU web server, sequences of at least 10 amino acids, up to 5,000 per submission; code on GitHub under CC BY-NC-SA | Life Science Alliance 2(5): e201900429 |
| Phobius | Transmembrane topology and signal peptides together, in one hidden Markov model | Free web interface | Käll, Krogh and Sonnhammer (2007), Nucleic Acids Research, PMID 17483518 |
TargetP 2.0 is the natural second tool for eukaryotic sets, because SignalP 6.0 stops at Sec/SPI for Eukarya. Phobius is useful when signal peptides and transmembrane helices may be confused. For broader context, see what bioinformatics covers.
FAQ
How do I cite SignalP 6.0?
Cite Teufel F., Almagro Armenteros J.J., et al. (2022), Nature Biotechnology, doi:10.1038/s41587-021-01156-3, and state the version (such as 6.0i) and mode you used.
How do I install SignalP?
Download the package from the DTU academic download page, install it with pip install signalp-6-package/, copy the model files into model_weights/, then run signalp6_convert_models gpu or cpu. Python up to 3.10 is required.
What does SignalP 6.0 predict?
Signal peptides of five types (Sec/SPI, Sec/SPII, Tat/SPI, Tat/SPII, Sec/SPIII) and their cleavage sites, across Archaea, Gram-positive bacteria, Gram-negative bacteria and Eukarya. For Eukarya only Sec/SPI is predicted.
What is the difference between fast and slow mode?
Fast mode is the default and uses a distilled model. Slow mode uses the full model, takes about six times longer with the same RAM, and is intended for accurate region borders.
Does a predicted signal peptide mean the protein is secreted?
Not on its own. The DTU documentation states that a signal peptide alone does not establish final localization.
Sources
- DTU Health Tech, SignalP 6.0 server page: services.healthtech.dtu.dk/services/SignalP-6.0/ (read 2026-11-07)
- SignalP 6.0 installation instructions: github.com/fteufel/signalp-6.0 (read 2026-11-07)
- Teufel F. et al. (2022), Nature Biotechnology, doi:10.1038/s41587-021-01156-3
- DTU Health Tech, TargetP 2.0 server page: services.healthtech.dtu.dk/services/TargetP-2.0/ (read 2026-11-07)
- Käll L., Krogh A., Sonnhammer E.L.L. (2007), Nucleic Acids Research: PubMed 17483518







