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.
Usage 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:
Eukaryotes, Gram-negative bacteria or Gram-positive bacteria. - Method:
Neural networks, hidden Markov models or both. - Graphics output:
No graphics, in line GIF or in line GIF and EPS as links. See the Output format for examples. - Text output:
Standard, full or short output format. See the Output format for examples. - Sequence truncation:
Signal peptides occurr at the N-terminal end of protein sequences; they are seldom longer than 45 amino acids. It is normally not meaningful to submit more than 60-70 amino acids per sequence. Therefore, the default truncation has been set to 70.
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
Description of the scores
Examples of standard output
Examples of short output
The C-score is the ``cleavage site'' score. For eachposition in the submitted sequence, a C-score is reported, whichshould only be significantly high at the cleavage site. Confusionis often seen with the position numbering of the cleavage site.When a cleavage site position is referred to by a single number,the number indicates the first residue in the mature protein,meaning that a reported cleavage site between amino acid 26-27corresponds to that the mature protein starts at (and include)position 27.
Y-max is a derivative of the C-score combined with theS-score resulting in a better cleavage site prediction than theraw C-score alone. This is due to the fact that multiplehigh-peaking C-scores can be found in one sequence, where only oneis the true cleavage site. The cleavage site is assigned from theY-score where the slope of the S-score is steep and a significantC-score is found.
The S-mean is the average of the S-score, ranging from theN-terminal amino acid to the amino acid assigned with the highestY-max score, thus the S-mean score is calculated for the length ofthe predicted signal peptide. The S-mean score was in SignalPversion 2.0 used as the criteria for discrimination of secretoryand non-secretory proteins.
The D-score is introduced in SignalP version 3.0 and is asimple average of the S-mean and Y-max score. The score showssuperior discrimination performance of secretory and non-secretoryproteins to that of the S-mean score which was used in SignalPversion 1 and 2.
For non-secretory proteins all the scores represented in theSignalP3-NN output should ideally be very low.
The hidden Markov model calculates the probability of whether thesubmitted sequence contains a signal peptide or not. Theeukaryotic HMM model also reports the probability of a signalanchor, previously named uncleaved signal peptides. Furthermore,the cleavage site is assigned by a probability score together withscores for the n-region, h-region, and c-region of the signalpeptide, if such one is found.
EXAMPLES OF STANDARD OUTPUT
By default the server produces the following output for each input sequence:Example 1: secretory protein
The example below shows the output for thioredoxin domain containingprotein 4 precursor (endoplasmic reticulum protein ERp44), taken from theSwiss-ProtentryTXN4_HUMAN.The signal peptide prediction is consistent with the database annotation. # data # data>TXN4_HUMANSignalP-NN result:
>Sequence length = 70# Measure Position Value Cutoff signal peptide? max. C 30 0.565 0.32 YES max. Y 30 0.690 0.33 YES max. S 12 0.989 0.87 YES mean S 1-29 0.852 0.48 YES D 1-29 0.771 0.43 YES# Most likely cleavage site between pos. 29 and 30: VTT-EI
SignalP-HMM result:>TXN4_HUMANPrediction: Signal peptideSignal peptide probability: 0.984Signal anchor probability: 0.015Max cleavage site probability: 0.962 between pos. 29 and 30
# gnuplot scriptfor making the plot(s)
Example 2: non-secretory protein
The example below shows the output for BMP-2 inducible protein kinase(EC 2.7.1.37), a nuclear protein taken from theSwiss-ProtentryBM2K_HUMAN.No signal peptide is predicted.>BM2K_HUMANSignalP-NN result:
# data
>BM2K_HUMAN length = 70# Measure Position Value Cutoff signal peptide? max. C 20 0.035 0.32 NO max. Y 20 0.034 0.33 NO max. S 12 0.263 0.87 NO mean S 1-19 0.063 0.48 NO D 1-19 0.049 0.43 NOSignalP-HMM result:# data
>BM2K_HUMANPrediction: Non-secretory proteinSignal peptide probability: 0.157Signal anchor probability: 0.023Max cleavage site probability: 0.027 between pos. 28 and 29# gnuplot scriptfor making the plot(s)
EXAMPLE OF SHORT OUTPUT
When selecting the short output format, the prediction for each submittedsequence (in a multisequence FASTA file) are reported on a single line,one for each fasta entry. A two line header is included, showing theinformation of the different columns.# SignalP-NN euk predictions # SignalP-HMM euk predictions# name Cmax pos ? Ymax pos ? Smax pos ? Smean ? D ? # name ! Cmax pos ? Sprob ?TXN4_HUMAN 0.565 30 Y 0.690 30 Y 0.989 12 Y 0.852 Y 0.771 Y TXN4_HUMAN S 0.962 30 Y 0.984 Y BM2K_HUMAN 0.035 20 N 0.034 20 N 0.263 12 N 0.063 N 0.049 N BM2K_HUMAN Q 0.027 29 N 0.157 N
Scientific background
For a brief description of the SignalP method please consult the article abstracts.
Biological background
Interest in signal peptides has for a long time been one of thehot topics in bioinformatics. 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". He pointed out the importance of defined peptide motifs for targeting proteins to their site of function. The press release can be read here
For biological background of protein localization we refer to the followingpages.
Signal peptides
Signal anchors
Other secretory signals
Data sets and statictics
A very important task in machine learning methods is to obtain a clean and accurate dataset for trainingand testing. Bias and noise in the data set often lead to wrong predictions, which is undesirable.
Description of data sets
Dataset extraction
Dataset cleanup
Sequence logos
Length distributions
Characteristics of signal peptides
Download the training sets
Methods for prediction of signal peptides
With the current growth of sequence databases and speed of genome sequencing,accurate prediction methods have become increasingly important.For SignalP we have focused on neural networks as well as Hidden Markov Models.
Neural Networks
Hidden Markov Models
Performance and results
Any machine learning approach must be evaluated to test the predictive performance on unknown sequences.
Performance of the current prediction method
Five fold crossvalidation
Independent test set by Menne
Signal anchor prediction
Acknowledgements
The information on these pages are partly generated by the initial creator ofSignalP, Henrik Nielsen. The information provided have been updated with new knowledge, but most of the biological background text emerges from Henriks work.
References
Main references:
- Original method (SignalP v. 1.1)
- Update to SignalP v. 2.0
- Update to SignalP v. 3.0 (current method)
Other publications
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(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
Other publications
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.Predictions can be made on a publicly available WWW server:http://www.cbs.dtu.dk/services/SignalP/.
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(2):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.
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