Prediction of signal peptides

The importance of signal peptides was shown in 1999 when Günter Blobel received the Nobel Prize in physiology or medicine for his discovery that "proteins have intrinsic signals that govern their transport and localization in the cell" [Blobel, G. (2000)]. He pointed out the importance of defined peptide motifs for targeting proteins to their site of function.

Performing a query to PubMed (http://www.ncbi.nlm.nih.gov/entrez/)reveals that thousands of papers have been published, regarding signal peptides, secretion and subcellular localization, including knowledge of using signal peptides as vehicles for chimeric proteins for biomedical and pharmaceutical industry. Many papers describe statistical or machine learning methods for prediction of signal peptides and prediction of subcellular localization in general. After the first published method for signal peptide prediction [von Heijne, 1986], more and more methods have surfaced, although not all methods have been made available publicly.

Different types of signal peptides

Figure: Schematic representation of various signal peptides. Red color indicates n-region, gray color indicates h-region, cyan indicates c-region. All white circles are part of the mature protein. indicates the first position of the mature protein. The length of the signal peptides is not drawn to scale.

Soon after Günter Blobel's initial discovery of signal peptides, more targeting signals were found. Most cell types and organisms employ several ways of targeting proteins to the extracellular environment or subcellular locations. Most of the proteins targeted for the extracellular space or subcellular locations carry specific sequence motifs (signal peptides) characterizing the type of secretion/targeting it undergoes.

Several new different signal peptides or targeting signals have been found during the later years, and papers often describe a small amino acid motif required for secretion of that particular protein. In most of the latter cases, the identified sequence motif is only found in this particular protein and as such cannot be described as a new group of signal peptides.

Describing the various types of signal peptides is beyond the scope of this text but several review papers on this topic can be found on PubMed. Targeting motifs can either be removed from, or retained in the mature protein after the protein has reached the correct and final destination. Some of the best characterized signal peptides are depicted in the figure above.

Numerous methods for prediction of protein targeting and signal peptides have been developed; some of them are mentioned and cited in the introduction of the SignalP research paper [Bateman et al., 2004]. However, no prediction method will be able to cover all the different types of signal peptides. Most methods predicts classical signal peptides targeting to the general secretory pathway in bacteria or classical secretory pathway in eukaryotes. Furthermore, a few methods for prediction of non-classically secreted proteins have emerged [Bendtsen et al., 2004a, Bendtsen et al., 2005].

Figure: Sequence logo of eukaryotic signal peptides, showing conservation of amino acids in bits [Schneider and Stephens, 1990]. Polar and hydrophobic residues are shown in green and black, respectively, while blue indicates positively charged residues and red negatively charged residues. The logo is based on an ungapped sequence alignement fixed at the position of the signal peptides.

Prediction of signal peptides and subcellular localization

In the search for accurate prediction of signal peptides, many approaches have been investigated. Almost 20 years ago, the first method for prediction of classical signal peptides was published [von Heijne, 1986]. Nowadays, more sophisticated machine learning methods, such as neural networks, support vector machines, and hidden Markov models have arrived along with the increasing computational power and they all perform superior to the old weight matrix based methods [Menne et al., 2000]. Also, many other "classical" statistical approaches have been carried out, often in conjunction with machine learning methods. In the following sections, a wide range of different signal peptide and subcellular prediction methods will be described.

Most signal peptide prediction methods require the presence of the correct N-terminal end of the preprotein for correct classification. As large scale genome sequencing projects sometimes assign the 5'-end of genes incorrectly, many proteins are annotated without the correct N-terminal [Reinhardt and Hubbard, 1998] leading to incorrect prediction of subcellular localization. These erroneous predictions can be ascribed directly to poor gene finding. Other methods for prediction of subcellular localization use information within the mature protein and therefore they are more robust to N-terminal truncation and gene finding errors.

The SignalP method

One of the most cited and best methods for prediction of classical signal peptides is the SignalP method. In contrast to other methods, SignalP also predicts the actual cleavage site; thus the peptide which is cleaved off during translocation over the membrane. Recently, an independent research paper has rated SignalP version 3.0 to be the best standalone tool for signal peptide prediction. It was shown that the D-score which is reported by the SignalP method is the best measure for discriminating secretory from non-secretory proteins [Klee and Ellis, 2005].

SignalP is located at http://www.cbs.dtu.dk/services/SignalP/

What do the SignalP scores mean?

Many bioinformatics approaches or prediction tools do not give a yes/no answer. Often the user is facing an interpretation of the output, which can be either numerical or graphical. Why is that? In clear-cut examples there are no doubt; yes: this is a signal peptide! But, in borderline cases it is often convenient to have more information than just a yes/no answer. Here a graphical output can aid to interpret the correct answer. An example is shown in the figure below.

The graphical output from SignalP (neural network) comprises three different scores, C, S and Y. Two additional scores are reported in the SignalP3-NN output, namely the S-mean and the D-score, but these are only reported as numerical values.

For each organism class in SignalP; Eukaryote, Gram-negative and Gram-positive, two different neural networks are used, one for predicting the actual signal peptide and one for predicting the position of the signal peptidase I (SPase I) cleavage site. The S-score for the signal peptide prediction is reported for every single amino acid position in the submitted sequence, with high scores indicating that the corresponding amino acid is part of a signal peptide, and low scores indicating that the amino acid is part of a mature protein.

The C-score is the `"cleavage site" score. For each position in the submitted sequence, a C-score is reported, which should only be significantly high at the cleavage site. Confusion is 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. This means that a reported cleavage site between amino acid 26-27 corresponds to the mature protein starting at (and include) position 27.

Y-max is a derivative of the C-score Main with the S-score resulting in a better cleavage site prediction than the raw C-score alone. This is due to the fact that multiple high-peaking C-scores can be found in one sequence, where only one is the true cleavage site. The cleavage site is assigned from the Y-score where the slope of the S-score is steep and a significant C-score is found.

The S-mean is the average of the S-score, ranging from the N-terminal amino acid to the amino acid assigned with the highest Y-max score, thus the S-mean score is calculated for the length of the predicted signal peptide. The S-mean score was in SignalP version 2.0 used as the criteria for discrimination of secretory and non-secretory proteins.

The D-score is introduced in SignalP version 3.0 and is a simple average of the S-mean and Y-max score. The score shows superior discrimination performance of secretory and non-secretory proteins to that of the S-mean score which was used in SignalP version 1 and 2.

For non-secretory proteins all the scores represented in the SignalP3-NN output should ideally be very low.

The hidden Markov model calculates the probability of whether the submitted sequence contains a signal peptide or not. The eukaryotic HMM model also reports the probability of a signal anchor, previously named uncleaved signal peptides. Furthermore, the cleavage site is assigned by a probability score together with scores for the n-region, h-region, and c-region of the signal peptide, if it is found.