Restrictions:
At most 2,000 sequences and 200,000 amino acids per submission; each sequence not more than 6,000 amino acids.
Confidentiality:
The sequences are kept confidential and will be deleted after processing.
CITATIONS
For publication of results, please cite:
SignalP 4.0: discriminating signal peptides from transmembrane regions
Thomas Nordahl Petersen, Søren Brunak, Gunnar von Heijne & Henrik Nielsen
Nature Methods, 8:785-786, 2011doi: 10.1038/nmeth.1701
PMID: 21959131
Supplementary materials: nmeth.1701-S1.pdf
Other relevant papers:
- Original paper (version 1.0):
Identification of prokaryotic and eukaryotic signal peptides and prediction of their cleavage sites.
Henrik Nielsen, Jacob Engelbrecht, Søren Brunak and Gunnar von Heijne.
Protein Engineering, 10:1-6, 1997. - SignalP-HMM (version 2.0):
Prediction of signal peptides and signal anchors by a hidden Markov model.
Henrik Nielsen and Anders Krogh.
Proceedings of the Sixth International Conference on Intelligent Systems for Molecular Biology (ISMB 6),
AAAI Press, Menlo Park, California, pp. 122-130, 1998. - Version 3.0:
Improved prediction of signal peptides: SignalP 3.0.
Jannick Dyrløv Bendtsen, Henrik Nielsen, Gunnar von Heijne and Søren Brunak.
J. Mol. Biol., 340:783-795, 2004.Download the full article in PDF.
- Paper about using SignalP and otherprotein subcellular localization prediction methods:
Locating proteins in the cell using TargetP, SignalP, and related tools
Olof Emanuelsson, Søren Brunak, Gunnar von Heijne, Henrik Nielsen
Nature Protocols 2:953-971 (2007).
Access the paper and supplementary materials.
Frequently Asked Questions
Changes from version 4.0 to 4.1
Changes from version 3 to 4
Biological background, signal peptides
Biological background, other sorting signals
Biological background, organism groups
History
Changes from version 4.0 to 4.1
— What's new?
Please see the Version history.
— Why do you present a choice between two cutoff settings? Can't you just decide on one?
The optimal cutoff really depends on what you want to use the methodfor. If it is important to find all signal peptides, use the sensitivecutoff. If you want an estimate of the number of signal peptides in agenome, use the default cutoff.
— Why have you imposed a minimum length?
Because we believe that predictions of signal peptides shorter than ten residues made by SignalP 4.1 are false. The shortestknown signal peptides are 11 residues long (with one exception, SP23_TENMO, which does not look like a signal peptide atall). Clickhere for an updated list of experimentally confirmed signalpeptides from UniProt of length 11 or shorter.
— What happened to the Background page?
It's here! The important material from the Background page has beenintegrated into this FAQ, we hope you like the new format.
Changes from version 3 to 4
— What's new?
Please see the Version history.
— What happened to the HMM part?
While making SignalP 4.0, we did retrain the Hidden Markov Model (HMM) part of SignalP. However, we found that it did not perform better thanthe neural networks in any of the performance parameters we tested.Therefore, we decided not to include it. If the HMM output is importantfor you, you can still use SignalP 3.0.
— Why is my favourite signal peptide no longer predicted correctly?SignalP 3.0 could do it!
As explained on the Performance page, SignalP 4 with the default cutoff has a lower sensitivity than SignalP3. Please try again with the new "Sensitive" setting.
— What happened to the Yes/No answers for max C score etc.?
SignalP 3.0 provided five Yes/No answers for the NN part. We found thatthis was confusing for users and obscured the fact that the D-score isthe best score for discriminating between signal peptides and non-signalpeptides.
Biological background, signal peptides
— What are signal peptides?
The term "signal peptide" is used with two meanings: In the broadsense (used in many textbooks), a signal peptide is any sorting signal embedded in the aminoacid sequence of a protein. In the narrow sense (used in most of the scientific literature), a signal peptideis an N-terminal signal that directs the protein across the ERmembrane in eukaryotes and across the plasma membrane in prokaryotes.Signal peptides in the narrow sense are also known as ER signalpeptides or secretory signal peptides. Read more in UniProt, inWikipedia,and in the Sequence feature ontology.
It is important to emphasize that SignalP predicts signal peptides inthe narrow sense only.
— Are signal peptides always N-terminal?
In the narrow sense: Yes, per definition. In the broad sense: No, there are several sorting signal that are C-terminal(e.g. the PTS1 signal for peroxisomal import)or internal (e.g. the nuclear localization signal).
— Are signal peptides (in the narrow sense) always cleaved?
No, there are rare cases of uncleaved signal peptides. For an updated list of such proteins annotated in UniProt, click here.These should not be confused with signal anchors, see below.In SignalP, these typically get high S-scores but low C- and Y-scores.
— Which protease is responsible for signal peptidecleavage?
In bacteria, it is Signal Peptidase I (SPase I), also known as Leader Peptidase(Lep). In eukaryotes, it is the signal peptidase complex (SPC), whichconsists of four subunits in yeast and five in mammals. Read more in MEROPS.
— My protein has a signal peptide. Can I then safelyconclude that it is secreted?
No. You can only conclude that it enters the secretory pathway.
In eukaryotes, there are several opportunities for a protein with asignal peptide to escape secretion. It could:
- It could be retained in the endoplasmic reticulum (ER). Soluble ER-resident proteins have a C-terminal retention signal with the consensussequence KDEL, see PROSITE.
- be retained in the Golgi apparatus,
- be directed to the lysosome (vacuole in plants and fungi),
- have one or more transmembrane helices and therefore beretained in either the plasma membrane, or one of the membranes of thesecretory pathway (ER, Golgi, lysosome/vacuole), or
- have a signal for GPI-anchoring, a C-terminal cleavedpeptide which functions as a signal for attachment of aGlycophosphatidylinositol (GPI) group that anchors the protein to theouter face of the plasma membrane.
In Gram-positive bacteria, a protein with a signal peptide could:
- have one or more transmembrane helices, or
- be attached to the cell wall.
In Gram-negative bacteria, a protein with a signal peptide could:
- have one or more transmembrane helices,
- be retained in the periplasm, or
- be inserted into the outer membrane as a β-barrel transmembraneprotein.
— Does SignalP predict signal peptides of bacterial lipoproteins?
No. Bacterial lipoproteins have special signal peptides which arecleaved by Signal Peptidase II (SPase II), also known as Lipoproteinsignal peptidase (Lsp). A diacylglyceryl group is attached to a Cysteine residuein position +1 relative to the cleavage site, which bears no resemblanceto the SPase I cleavage site. See alsoMEROPSand PROSITE.
For prediction of prokaryotic lipoproteins we recommend using the LipoP server.
— Does SignalP predict TAT (Twin-arginine translocation) signal peptides?
Not very well. Bacterial TAT signal peptides, which direct their proteins throughan alternative translocon (TatABC), have a special motif containing twoArginines in the n-region. Additionally, they are in general longer and lesshydrophobic than normal (Sec) signal peptides. As a consequence, they are sometimes missed by SignalP, even though they are cleaved by SPaseI (Lep). See also PROSITE andInterPro.
For prediction of TAT signal peptides we recommend using the TatP server.
Biological background, other sorting signals
— What are signal anchors?
A signal anchor is a transmembrane helix located close to the N-terminusof a protein with an N-in orientation (i.e. the N-terminus is on thecytoplasmic side of the membrane). It functions much like a signalpeptide since it is recognized by the Signal Recognition Particle (SRP)and inserted into the translocon; but instead of being cleaved anddegraded it remains in the membrane and anchors the protein to it.Proteins anchored in this way are known as Type II transmembraneproteins.
![]() | Signal peptides (above) versus signal anchors (below) |
It is important to realize that the difference between signal peptidesand signal anchors is not a question of presence or absence of acleavage site. Instead, the most important difference seems to be thelength of the hydrophobic domain. It has been shown experimentally that it is possible to convert a cleavedsignal peptide to a signal anchor merely by lengthening theh-region, without altering the cleavage site(Chou & Kendall 1990; Nilsson, Whitley, & von Heijne 1994).
The introduction of the Hidden Markov Model (HMM) method in SignalPversion 2 made it possible to some extent to distinguish signal peptidesfrom signal anchors (in that version, only in eukaryotes). However,SignalP 4 (based entirely on the Neural Network (NN) method), does abetter job, since its negative set is not confined only to transmembranehelices annotated as signal anchors, but includes all types oftransmembrane segments close to the N-terminus.
— What should I use for predicting signal peptides in thebroad sense?
For mitochondrial and plastid import signals, also known as transitpeptides, we recommend TargetP. For othersorting signals, we refer to our paper "Locating proteins in the cell using TargetP, SignalP, and relatedtools".
— What should I use for predicting non-classical (leaderless) secreted proteins?
Not all secretory proteins carry signal peptides. Some proteins enter a non-classical secretory pathwaywithout any currently known sequence motif. In eukaryotes, these proteins are mostly growth factorsand extracellular matrix binding proteins. In Gram-negative bacteria, thetype I, III, IV and VI secretion systems function without signal peptides. For prediction of such proteins werecommend the SecretomePserver.
Biological background, organism groups
— Why is there no version for Archaea?
Because there are too few experimentally confirmed signalpeptides from this organism group in the UniProt database (click here for an updated list).
— Which version should I use for vira and bacteriophages?
You should use the version corresponding to the host organism. There aresome indications that viral signal peptides differ from those of thehost organism, but SignalP currently does not take that into account.
— Which version should I use for Tenericutes/Mollicutes(Mycoplasma and related genera)?
You shouldn't use SignalP at all for these organisms, since they seem tolack a type I signal peptidase completely!
— Which version should I use for metagenomic sequencesof unknown origin?
This is an unsolved question. Please use all three versions tosearch for signal peptides in such data.
— Is one version enough for all eukaryotic organisms, orare there differences within the eukaryotes?
It is known that some yeast signal peptides are not recognized bymammalian cells (Bird et al., 1987 and 1990). Therefore, it would be natural to assume that separate SignalP versionsfor yeast and Mammalia would provide better predictions than a commoneukaryotic version. While developing SignalP 4.0 we tried dividing theeukaryotic data into animals, fungi, and plants and training separatemethods for these three groups. However, this did not give anyimprovement, and performance for all three groups was better when usingthe method trained on all eukaryotic sequences together.
— Are two versions enough for all bacteria, orare there differences within the Gram-positive/Gram-negative bacterial groups?
The Gram-negative version of SignalP is almost certainly biased towardsE. coli and other γ-proteobacteria, since these constitute the bulkof the experimentally annotated bacterial proteins in UniProt.Unpublished results suggest that some bacteria have very divergentcleavage site motifs. Future versions of SignalP might therefore dividethe Gram-negative bacteria into several classes, if data are available.
Gram-positive bacteria probably constitute a more hom*ogenous group, butit is an open question whether there are differences in signal peptidesbetween Actinobacteria (high G+C Gram-positive bacteria) and Firmicutes (low G+C Gram-positive bacteria). More data onActinobacteria are needed before that can be answered.
History
— How are the various versions of SignalP related?
Please see the Version history.
— Was there ever a Nobel prize awarded for signal peptides?
Yes, for signal peptides in the broad sense. The importance of signal peptideswas emphasized in 1999 when Günter Blobel received the Nobel Prize inphysiology or medicine for his discovery "proteins have intrinsicsignal that govern their transport and localization in the cell". The press release can be read here.
— Was SignalP the first signal peptide predictor?
No, but it was, to our knowledge, the first to be implemented as aweb server (in 1996). Among the earlier methods were McGeoch (1985) and von Heijne (1986), both of which have been included in PSORT.
— How many times have the SignalP papers been cited?
This information is available on Henrik Nielsen's ResearcherID,Scopus,and GoogleScholar pages.
References
Main references:
- Original method (SignalP v. 1.1)
- Update to SignalP v. 2.0
- Update to SignalP v. 3.0
- Update to SignalP v. 4.0
- Update to SignalP v. 4.1 (current method)
Other publications
Henrik Nielsen's PhD thesis
Original method (SignalP v. 1.1)
Identification of prokaryotic and eukaryotic signal peptidesand prediction of their cleavage sites.
Henrik Nielsen, Jacob Engelbrecht, Søren Brunak and Gunnar vonHeijne.
Protein Engineering, 10:1-6 (1997).
We have developed a new method for the identification of signal peptides andtheir cleavage sites based on neural networks trained on separate sets ofprokaryotic and eukaryotic sequence. The method performs significantly betterthan previous prediction schemes and can easily be applied on genome-wide datasets. Discrimination between cleaved signal peptides and uncleaved N-terminalsignal-anchor sequences is also possible, though with lower precision.Predictions can be made on a publicly available WWW server.
PMID: 9051728(free full text pdfversion)
Update to SignalP v. 2.0
Prediction of signal peptides and signal anchors by a hidden Markovmodel.
Henrik Nielsen and Anders Krogh.
Proc Int Conf Intell Syst Mol Biol. (ISMB 6), 6:122-130 (1998).
A hidden Markov model of signal peptides has been developed. It containssubmodels for the N-terminal part, the hydrophobic region, and the regionaround the cleavage site. For known signal peptides, the model can be used toassign objective boundaries between these three regions. Applied to our data,the length distributions for the three regions are significantly different fromexpectations. For instance, the assigned hydrophobic region is between 8 and 12residues long in almost all eukaryotic signal peptides. This analysis alsomakes obvious the difference between eukaryotes, Gram-positive bacteria, andGram-negative bacteria. The model can be used to predict the location of thecleavage site, which it finds correctly in nearly 70% of signal peptides in across-validated test — almost the same accuracy as the best previous method. Oneof the problems for existing prediction methods is the poor discriminationbetween signal peptides and uncleaved signal anchors, but this is substantiallyimproved by the hidden Markov model when expanding it with a very simple signalanchor model.
PMID: 9783217
Update to SignalP v. 3.0
Improved prediction of signal peptides: SignalP 3.0.
Jannick Dyrløv Bendtsen, Henrik Nielsen, Gunnar von Heijne and Søren Brunak.
J. Mol. Biol., 340:783-795 (2004).
We describe improvements of the currently mostpopular method for prediction of classically secreted proteins,SignalP. SignalP consists of two different predictors based onneural network and hidden Markov model algorithms, and bothcomponents have been updated. Motivated by the idea that thecleavage site position and the amino acid composition of thesignal peptide are correlated, new features have been included asinput to the neural network. This addition, together with athorough error-correction of a new data set, have improved theperformance of the predictor significantly over SignalP version 2.In version 3, correctness of the cleavage site predictions haveincreased notably for all three organism groups, eukaryotes, Gramnegative and Gram positive bacteria. The accuracy of cleavage siteprediction has increased in the range from 6–17 % over theprevious version, whereas the signal peptide discriminationimprovement mainly is due to the elimination of false positivepredictions, as well as the introduction of a new discriminationscore for the neural network. The new method has also beenbenchmarked against other available methods.
PMID: 15223320 doi: 10.1016/j.jmb.2004.05.028
Update to SignalP v. 4.0
SignalP 4.0: discriminating signal peptides from transmembrane regions.
Thomas Nordahl Petersen, Søren Brunak, Gunnar von Heijne and Henrik Nielsen.
Nature Methods, 8:785-786 (2011).
This is a Correspondence, it has no abstract.
doi: 10.1038/nmeth.1701
Access to the paper: if you have a personal or institutional subscription to Nature Methods, use the doi: link above. Access to the supplementary materials: nmeth.1701-S1.pdf
Other publications
Locating proteins in the cell using TargetP,SignalP, and related tools
Olof Emanuelsson, Søren Brunak, Gunnar von Heijne, Henrik Nielsen
Nature Protocols, 2:953-971 (2007).
Determining the subcellular localization of a protein is an importantfirst step toward understanding its function. Here, we describe theproperties of three well-known N-terminal sequence motifs directingproteins to the secretory pathway, mitochondria and chloroplasts, andsketch a brief history of methods to predict subcellular localizationbased on these sorting signals and other sequence properties. We thenoutline how to use a number of internet-accessible tools to arrive at areliable subcellular localization prediction for eukaryotic andprokaryotic proteins. In particular, we provide detailed step-by-stepinstructions for the coupled use of the amino-acid sequence-basedpredictors TargetP, SignalP, ChloroP and TMHMM, which are all hosted atthe Center for Biological Sequence Analysis, Technical University ofDenmark. In addition, we describe and provide web references to otheruseful subcellular localization predictors. Finally, we discusspredictive performance measures in general and the performance ofTargetP and SignalP in particular.
PMID: 17446895
Please click here to access the paper and supplementary materials.
Machine learning approaches to the prediction of signal peptidesand other protein sorting signals.
Henrik Nielsen, Søren Brunak, and Gunnar von Heijne.
Protein Engineering, 12:3-9 (1999), Review.
Prediction of protein sorting signals from the sequence of amino acids hasgreat importance in the field of proteomics today. Recently, the growth ofprotein databases, combined with machine learning approaches, such as neuralnetworks and hidden Markov models, have made it possible to achieve a level ofreliability where practical use in, for example automatic database annotationis feasible. In this review, we concentrate on the present status and futureperspectives of SignalP, our neural network-based method for prediction of themost well-known sorting signal: the secretory signal peptide. We discuss theproblems associated with the use of SignalP on genomic sequences, showing thatsignal peptide prediction will improve further if integrated with predictionsof start codons and transmembrane helices. As a step towards this goal, ahidden Markov model version of SignalP has been developed, making it possibleto discriminate between cleaved signal peptides and uncleaved signal anchors.Furthermore, we show how SignalP can be used to characterize putative signalpeptides from an archaeon, Methanococcus jannaschii. Finally, we briefly reviewa few methods for predicting other protein sorting signals and discuss thefuture of protein sorting prediction in general.
PMID: 10065704
A neural network method for identification of prokaryotic and eukaryoticsignal peptides and prediction of their cleavage sites.
Henrik Nielsen, Jacob Engelbrecht, Søren Brunakand Gunnar von Heijne.
Int. J. Neural Sys., 8:581-599 (1997).
We have developed a new method for the identification of signal peptides andtheir cleavage sites based on neural networks trained on separate sets ofprokaryotic and eukaryotic sequences. The method performs significantly betterthan previous prediction schemes, and can easily be applied to genome-wide datasets. Discrimination between cleaved signal peptides and uncleaved N-terminalsignal-anchor sequences is also possible, though with lower precision.
PMID: 10065837
Defining a similarity threshold for a functional protein sequence pattern:the signal peptide cleavage site.
Henrik Nielsen, Jacob Engelbrecht, Gunnar von Heijneand Søren Brunak.
Proteins, 24:165-77 (1996).
When preparing data sets of amino acid or nucleotide sequences it isnecessary to exclude redundant or hom*ologous sequences in order to avoidoverestimating the predictive performance of an algorithm. For some timemethods for doing this have been available in the area of protein structureprediction. We have developed a similar procedure based on pair-wisealignments for sequences with functional sites. We show how a correlationcoefficient between sequence similarity and functional hom*ology can be usedto compare the efficiency of different similarity measures and choose anonarbitrary threshold value for excluding redundant sequences. The impactof the choice of scoring matrix used in the alignments is examined. Wedemonstrate that the parameter determining the quality of the correlation isthe relative entropy of the matrix, rather than the assumed (PAM oridentity) substitution mode. Results are presented for the case ofprediction of cleavage sites in signal peptides. By inspection of the falsepositives, several errors in the database were found. The procedurepresented may be used as a general outline for finding a problem-specificsimilarity measure and threshold value for analysis of other functionalamino acid or nucleotide sequence patterns.
PMID: 8820484
From sequence to sorting: Prediction of signal peptides.
Henrik Nielsen.
Ph.D. thesis, defended at Department of Biochemistry,Stockholm University, Sweden, May 25, 1999.
In the present age of genome sequencing, a vast number of predictedgenes are initially known only by their putative nucleotidesequence. The newly established field of bioinformatics is concernedwith the computational prediction of structural and functionalproperties of genes and the proteins they encode, based on theirnucleotide and amino acid sequences.
Since one of the crucial properties of a protein is its subcellularlocation, prediction of protein sorting is an important question inbioinformatics. A fundamental distinction in protein sorting is thatbetween secretory and non-secretory proteins, determined by acleavable N-terminal sorting signal, the secretory signal peptide.
The main part of this thesis, including four of the six papers,concerns prediction of secretory signal peptides in both eukaryoticand bacterial data using two machine learning techniques: artificialneural networks and hidden Markov models. A central result is theSignalP prediction method, which has been made available as a WorldWide Web server and is very widely used.
Two additional prediction methods are also included, with one papereach. ChloroP predicts chloroplast transit peptides, anothercleavable N-terminal sorting signal; while NetStart predicts startcodons in eukaryotic genes. For prediction of all N-terminal signals,the assignment of correct start codon can be critical, which is whyprediction of translation initiation from the nucleotide sequence isalso important for protein sorting prediction.
This thesis comprises a detailed review of the molecular biology ofprotein secretion, a short introduction to the most important machinelearning algorithms in bioinformatics, and a critical review ofexisting methods for protein sorting prediction. In addition, it contains general treatment of the principles of data set constructionand performance evaluation for prediction methods in bioinformatics.
Access to the thesis (without the six included papers): PhDthesis.pdf; PhDthesis-cover.pdf
Instructions
1. Specify the input sequences
All the input sequences must be in one-letter amino acidcode. The allowed alphabet (not case sensitive) is as follows:
All the alphabetic symbols not in the allowed alphabetwill be converted to X before processing. All the non-alphabeticsymbols, including white space and digits, will be ignored.
The sequences can be input in the following two ways:
- Paste a single sequence (just the amino acids) or a number of sequences inFASTAformat into the upper window of the main server page.
- Select a FASTAfile on your local disk, either by typing the file name into the lower windowor by browsing the disk.
Both ways can be employed at the same time: all the specified sequences willbe processed. However, there may be not more than 2,000 sequences and200,000 amino acids in total in one submission. The sequencesmay not be longer than 6,000 amino acids.
2. Customize your run
- Organism group:
It is important for performance that you choose the correct organism group — Eukaryotes, Gram-negative bacteria or Gram-positive bacteria — since the signal peptides of these three groups are known to differ from each other.
Gram-positive bacteria correspond to Actinobacteria and Firmicutes in the NCBI Taxonomy.
Gram-negative bacteria are all other eubacteria, except Tenericutes (including Mycoplasma), which seem to lack a type I signal peptidase and therefore do not have standard signal peptides.
Unfortunately, we are unable to provide a SignalP version for Archaea, since there are too few experimentally confirmed signal peptides from this organism group in the UniProt database (click here to repeat the search). - D-cutoff values:
The default cutoff values for SignalP 4 are chosen to optimize the performance measured as Matthews Correlation Coefficient (MCC). This results in a lower sensitivity (true positive rate) than SignalP 3.0 had. In SignalP 4.1, we have introduced the option of setting the cutoff to a lower value which yields the same sensitivity as SignalP 3.0. This will make the false positive rate slightly higher, but still better than that of SignalP 3.0. Read more on the Performance page.
You can see which cutoff values are being used in the boxes marked "D-cutoff". They will change if you change the setting for "D-cutoff values" or "Organism group".
If you want to experiment with your own cutoff values, select "User defined" and the boxes will go blank, ready for you to fill in values between 0 and 1. - Graphics output:
In the default output, SignalP embeds one plot in PNG format per sequence, showing the C-, S-, and Y-scores for each position in the sequence. You can choose to avoid the plots (No graphics) or to add an Encapsulated PostScript (EPS) file for each sequence. The EPS files will be provided as links.
See the Output format for an example and explanation of the scores. - Output format:
You can choose between four output formats:- Standard
- Appropriate for most users. Shows one plot and one summary per sequence.
- Short
- Convenient if you submit lots of sequences. Shows only one line of output per sequence. Incompatible with graphics.
- Long
- Shows the C-, S-, and Y-scores for each position in the sequence in addition to the Standard output.
- All
- Shows the output scores of both neural network types (SignalP-TM and SignalP-noTM) for each position in the sequence. Incompatible with graphics.
- Method:
Signalp 4.1 contains two types of neural networks. SignalP-TM has been trained with sequences containing transmembrane segments in the data set, while SignalP-noTM has been trained without those sequences. Per default, SignalP 4.1 uses SignalP-TM as a preprocessor to determine whether to use SignalP-TM or SignalP-noTM in the final prediction (if 4 or more positions are predicted to be in a transmembrane state, SignalP-TM is used, otherwise SignalP-noTM).
An exception is Gram-positive bacteria, where SignalP-TM is used always.
If you are confident that there are no transmembrane segments in your data, you can get a slightly better performance by choosing "Input sequences do not include TM regions", which will tell SignalP 4.1 to use SignalP-noTM always. - Positional limits:
- Minimal predicted signal peptide length
- SignalP 4.0 could, in rare cases, erroneously predict signal peptides shorter than 10 residues. These errors have in SignalP 4.1 been eliminated by imposing a lower limit on the cleavage site position (signal peptide length). The minimum length is by default 10, but you can adjust it. Signal peptides shorter than 15 residues are very rare. If you want to disable this length restriction completely, enter 0 (zero).
- N-terminal truncation of input sequence
- By default, the predictor truncates each sequence to max. 70 residues before submitting it to the neural networks. If you want to predict extremely long signal peptides, you can try a higher value, or disable truncation completely by entering 0 (zero). Note: The neural networks are trained with sequences with a maximal length of 70, and they include the relative position in the sequence in their input. Therefore, general performance will deteriorate if you change this setting.
3. Submit the job
Click on the "Submit" button. The status of your job (either 'queued'or 'running') will be displayed and constantly updated until it terminates andthe server output appears in the browser window.
At any time during the wait you may enter your e-mail address and simply leavethe window. Your job will continue; you will be notified by e-mail when it hasterminated. The e-mail message will contain the URL under which the results arestored; they will remain on the server for 24 hours for you to collect them.
Output format
The neural networks in SignalP produce three output scores for eachposition in the input sequence: In the summary below the plot, the maximal values ofthe three scores are reported. In addition, the following two scores areshown: For non-secretory proteins all the scores represented in theSignalP output should ideally be very low (close to the negative targetvalue of 0.1).DESCRIPTION OF THE SCORES
The graphical output from SignalP (see below) shows the three differentscores, C, S and Y, for each position in thesequence.
Note the position numbering of the cleavage site: the C-score is trained to be high at the position immediately after the cleavage site (the first residue in the mature protein).EXAMPLE OUTPUT
By default the server produces the following output for each input sequence:Example: secretory protein — standard output format
The example below shows the output for thioredoxin domain containingprotein 4 precursor (endoplasmic reticulum protein ERp44), taken from theSwiss-ProtentryERP44_HUMAN.The signal peptide prediction is consistent with the database annotation.
# SignalP-4.1 euk predictions>sp_Q9BS26_ERP44_HUMAN Endoplasmic reticulum resident protein 44 OS_hom*o sapiens GN_ERP44 PE_1 SV_1# Measure Position Value Cutoff signal peptide? max. C 30 0.427 max. Y 30 0.586 max. S 9 0.950 mean S 1-29 0.821 D 1-29 0.713 0.450 YESName=sp_Q9BS26_ERP44_HUMANSP='YES' Cleavage site between pos. 29 and 30: VTT-EI D=0.713 D-cutoff=0.450 Networks=SignalP-noTM# data# gnuplot scriptSignal peptides: 1# processed fasta entries# gff file of processed entries
Below the summary for each sequence, two files are provided via links:"data" and "gnuplot script". If you have the free graphics program gnuplot on your computer, you canuse these two files to customize your plot.
Below the output for all the sequences, two other files are provided vialinks, if at least one signal peptide has been predicted. These are "processed fasta entries", a FASTA sequence file containing thesequences of protein that had predicted signal peptides, with the signalpeptide removed; and "gff file of processed entries", a file showing thesignal peptides feature of those proteins that had predicted signalpeptides in GFF format.
Example: secretory protein — short output format
# SignalP-4.0 euk predictions# name Cmax pos Ymax pos Smax pos Smean D ? Dmaxcut Networks-usedERP44_HUMAN 0.427 30 0.586 30 0.950 9 0.821 0.713 Y 0.450 SignalP-noTM
Performance of SignalP 4.1
Correlation
In the SignalP 4.0 article, we show that SignalP 4.0 is superior in performance to SignalP 3.0 and ten competingmethods (five dedicated signal peptide predictors and five transmembranetopology predictors with built-in signal peptide models), when theperformance is measured by Matthews Correlation Coefficient (MCC).
Matthews Correlation Coefficient is a very widely used measure forperformance in bioinformatics. It is defined thus:
where
- tp is the number of true positives (signal peptides predicted as such)
- tn is the number of true negatives (non-signal peptides predicted as such)
- fp is the number of false positives (erroneous signal peptide predictions)
- fn is the number of false negatives (missed signal peptides)
and it takes the value of 1 for a perfect prediction, 0 for a random (non-informative) prediction,and -1 for a consistently wrong prediction.
In Table E (pp. 10-11) of the supplementary materials you can see the MCC valuesfor SignalP and the competing methods.
Sensitivity, false positive rate and cutoff choice
However, SignalP 4.0 is not superior to SignalP 3.0 according to allperformance measures. Notably, the sensitivity is lower when youuse the default cutoff. Sensitivity is the proportion of the true signal peptides that are correctly predicted:
All prediction methods that make a classification from a numericaloutput have a choice to make: where to place the cutoff (also known asthreshold) for the output? If you use a high cutoff, you will getfew false positives, but also a low sensitivity; if you lower thecutoff, you will get a better sensitivity at the price of more falsepositives. The false positive rate is defined as:
There is no single correct answer to the problem of choosing thecutoff, it depends on the contet in which the prediction method is used.For SignalP, we have used a cutoff on the D-score (see the Output format for a definition) thatmaximizes the MCC.
ROC curves
The trade-off between sensitivity and false positive rate is oftenillustrated graphically as a so-called ROC curvewhich has false positive rateon the x-axis and sensitivity on the y-axis for varyingvalues of the cutoff. The better a predition method is, the closer to the upper left corner the ROC curve will be, while a random (non-informative) prediction will followthe diagonal. This is anexcellent way to compare different predictors, since it is not dependenton cutoff choice.
Below, you can see ROC curves for SignalP 3 and 4 for the threedifferent organism groups. Note: in contrast to the values in Table E,these are not evaluation performances; they are made by applying the finishedmethods to the Total data set before hom*ology reduction.
![SignalP 4.1 - DTU Health Tech (3) SignalP 4.1 - DTU Health Tech (3)](https://i0.wp.com/services.healthtech.dtu.dk/euk.roc.png)
![SignalP 4.1 - DTU Health Tech (4) SignalP 4.1 - DTU Health Tech (4)](https://i0.wp.com/services.healthtech.dtu.dk/gram+.roc.png)
![SignalP 4.1 - DTU Health Tech (5) SignalP 4.1 - DTU Health Tech (5)](https://i0.wp.com/services.healthtech.dtu.dk/gram-.roc.png)
These ROC curves show that:
- When there are TM segments in the data ("all data"), SignalP 4.0 is clearly better than SignalP 3.0 (compare the pink and green curves)
- When TM segments are excluded from the data ("no TM"), SignalP 4.0 performance is practically equal to that of SignalP 3.0 — except in the Gram-positives, where it is better(compare the blue and red curves)
- SignalP 4.0 and 3.0 default cutoffs are placed at very different points on the ROC curves, leading to lower sensitivity (and much lower FP rates)in SignalP 4.0.
The cutoff choice in SignalP 4.1
SignalP 4.1 offers the users an option of using cutoff values which reproduce the sensitivity of SignalP 3.0. The price is, of course, aslightly higher false positive rate.
In the table below, the performace values are shown for SignalP 3.0, SignalP 4.1 with default cutoff, and SignalP 4.1 with "sensitive" (SignalP-3.0 compliant) cutoff. Note, again, that these are notevaluation performances and should not be used to compare SignalP tocompeting methods, they are merely for the purpose of comparing SignalPversions.
Method | Cutoff, SignalP-noTM | Cutoff, SignalP-TM | Sensitivity | FP rate, no TM | FP rate, all data | MCC, no TM | MCC, all data |
---|---|---|---|---|---|---|---|
Eukaryotic data | |||||||
SignalP 3.0 | 0.43 | 0.988 | 0.008 | 0.117 | 0.978 | 0.781 | |
SignalP 4.1 default | 0.45 | 0.50 | 0.967 | 0.003 | 0.011 | 0.972 | 0.955 |
SignalP 4.1 sensitive | 0.34 | 0.34 | 0.988 | 0.009 | 0.043 | 0.976 | 0.903 |
Gram-positive data | |||||||
SignalP 3.0 | 0.45 | 0.961 | 0.008 | 0.033 | 0.937 | 0.814 | |
SignalP 4.1 default | 0.57 | 0.45 | 0.950 | 0.000 | 0.001 | 0.973 | 0.967 |
SignalP 4.1 sensitive | 0.42 | 0.42 | 0.961 | 0.000 | 0.003 | 0.978 | 0.958 |
Gram-negative data | |||||||
SignalP 3.0 | 0.44 | 0.955 | 0.004 | 0.061 | 0.949 | 0.691 | |
SignalP 4.1 default | 0.57 | 0.51 | 0.924 | 0.000 | 0.001 | 0.957 | 0.949 |
SignalP 4.1 sensitive | 0.42 | 0.42 | 0.955 | 0.002 | 0.006 | 0.963 | 0.937 |
Examples of proteome predictions for three organism types
Eukaryota - Human proteom GRCh37.62
short
gff
mature
Gram positive bacteria - B.subtilis EB2
short
gff
mature
Gram negative bacteria - E.coli K12
short
gff
mature
Training and testing data sets
These are the annotated sequence data described in Table A of theSupplementary Materials. The entiredatasets correspond to the "Total" columns in the table (before hom*ology reduction). Sequences labeled "Train" correspond to the "Train" columnsin the table, while sequences labeled "Evaluation" correspond to the"Comp." columns in the table (used for comparing the performance to SignalP 3.0 and other methods). Sequences used to train SignalP 3.0 (orhom*ologous to those used to train SignalP 3.0) have been removed fromthe "Comp." sets.
Note that the "Comp." sets are subsets of the "Train" sets. Theevaluation of SignalP 4.0 was done using a nested cross-validationapproach, where different partitions were used for training, optimization and evaluation, see Supplementary Materials for details.
166 AJL2_ANGJA EvaluationMVSFKLPAFLCVAVLSSMALVSHGAVLGLCEGACPEGWVEHKNRCYLHVAEKKTWLDAELNCLHHGGNLASEHSEDEHQFLKDLHKGSDDPFWIGLSAVHEGRSWLWSDGTSASAEGDFSMWNPGEPNDAGGKEDCVHDNYGGQKHWNDIKCDLLFPSICVLRMVESSSSSSSSSSSSSSSSSSSSSSSS.............................................................................................................................................. 503 A1BG_BOVIN Evaluation TrainMSAWAALLLLWGLSLSPVTEQATFFDPRPSLWAEAGSPLAPWADVTLTCQSPLPTQEFQLLKDGVGQEPVHLESPAHEHRFPLGPVTSTTRGLYRCSYKGNNDWISPSNLVEVTGAEPLPAPSISTSPVSWITPGLNTTLLCLSGLRGVTFLLRLEGEDQFLEVAEAPEATQATFPVHRAGNYSCSYRTHAAGTPSEPSATVTIEELDPPPAPTLTVDRESAKVLRPGSSASLTCVAPLSGVDFQLRRGAEEQLVPRASTSPDRVFFRLSALAAGDGSGYTCRYRLRSELAAWSRDSAPAELVLSDGTLPAPELSAEPAILSPTPGALVQLRCRAPRAGVRFALVRKDAGGRQVQRVLSPAGPEAQFELRGVSAVDSGNYSCVYVDTSPPfa*gSKPSATLELRVDGPLPRPQLRALWTGALTPGRDAVLRCEAEVPDVSFLLLRAGEEEPLAVAWSTHGPADLVLTSVGPQHAGTYSCRYRTGGPRSLLSELSDPVELRVAGSSSSSSSSSSSSSSSSSSSSSS..................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................
The format is:
- First a header line with number of amino acids, sequence name (UniProt ID) and possibly a description field ('Evaluation'/'Train').
- The protein sequence.
- The annotations, one for each amino acid.
Annotations:
S— Amino acid is part of a Signal peptide (experimentally verified)
T— Amino acid is part of a Transmembrane region (experimentally verified)
t— Amino acid is part of a Transmembrane region (not experimentally verified)
.— An annotation different from those shown above
Eukaryota sequence data
Gram positive sequence data
Gram negative sequence data
Version history
Please click on the version number to activate the corresponding server where available.
4.1 | The current server. New in this version:
Main publication:
|
4.0 | New in this version:
Main publication:
|
3.0 | New in this version:
Main publication:
|
2.0 | New in this version:
Main publication:
|
1.1 | The original server: the method based on artificial neural networks. Main publication:
|
Software Downloads
- Version 6.0h
- slow_sequential
- fast
- Version 5.0b
- Linux
- Darwin
- Version 4.1g
- Linux
- IRIX64
- Darwin
- CYGWIN
- Version 3.0
- SunOS
- OSF1
- Linux
- IRIX
- AIX
- Version 2.0
- SunOS
- OSF1
- Linux
- IRIX
- AIX
GETTING HELP
If you need help regarding technical issues (e.g. errors or missing results) contact Technical Support. Please include the name of the service and version (e.g. NetPhos-4.0) and the options you have selected. If the error occurs after the job has started running, please include the JOB ID (the long code that you see while the job is running).
If you have scientific questions (e.g. how the method works or how to interpret results), contact Correspondence.
Correspondence: Technical Support: