PredSL: a tool for the N-terminal sequence-based prediction of protein subcellular localization.

The ability to predict the subcellular localization of a protein from its sequence is of great importance, as it provides information about the protein's function. We present a computational tool, PredSL, which utilizes neural networks, Markov chains, profile hidden Markov models, and scoring matrices for the prediction of the subcellular localization of proteins in eukaryotic cells from the N-terminal amino acid sequence. It aims to classify proteins into five groups: chloroplast, thylakoid, mitochondrion, secretory pathway, and "other". When tested in a five-fold cross-validation procedure, PredSL demonstrates 86.7% and 87.1% overall accuracy for the plant and non-plant datasets, respectively. Compared with TargetP, which is the most widely used method to date, and LumenP, the results of PredSL are comparable in most cases. When tested on the experimentally verified proteins of the Saccharomyces cerevisiae genome, PredSL performs comparably if not better than any available algorithm for the same task. Furthermore, PredSL is the only method capable for the prediction of these subcellular localizations that is available as a stand-alone application through the URL:http://bioinformatics.biol.uoa.gr/PredSL/.

Introduction

The subcellular localization of a protein is indicative of its function. Most proteins in eukaryotic cells are synthesized in the cytosol and are translocated to various subcellular compartments with the aid of a targeting sequence (. In most cases of mitochondrial (mTP), chloroplast (cTP), thylakoid lumen (lTP), and secreted proteins (SP), this targeting sequence is in the N-terminal. There have been efforts to use several types of information towards predicting the subcellular localization of proteins in recent years (, including the use of protein homology (, the whole amino acid composition (, quasi-sequence-order effect (, evolutionary and structural information (, and the N-terminal targeting sequence (. The techniques used for this purpose include statistics (, weight matrices (, text mining (, and machine learning methods such as support vector machines (SVMs; ref. ), hidden Markov models (HMMs; ref. 12., 13., 14.), feed-forward neural networks (, and recurrent neural networks (. Currently, the most widely used methods for the prediction of protein subcellular localization are SignalP (, ChloroP (, and TargetP (, which use the N-terminal residues of a sequence and two layers of neural networks for prediction.

Here, we present a novel approach named PredSL for the prediction of protein subcellular localization based on the presence of an N-terminal targeting sequence, combining several of the techniques mentioned above. We use the N-terminal targeting sequence as it is responsible for the specific targeting of most proteins belonging to the particular subcellular localizations. Therefore, we consider it to be sufficient to provide a reliable prediction. Other approaches, using for instance the whole amino acid sequence for prediction, have been shown to produce higher accuracy in some circumstances, but we did not consider them here. This is because the N-terminal sequence-based prediction also offers a useful biological insight into the mechanism of protein targeting. For instance, a predictor that uses the whole amino acid sequence as input may be biased against particular protein families, reflecting the distinct biological functions of the organelles that they are targeted to. While using the N-terminal sequence, we are in accordance (at least for the vast majority of the proteins targeted to the particular subcellular localizations) with the experimentally derived evidence, supporting the idea that the N-terminal sequence is sufficient for the appropriate targeting. The tool developed here classifies proteins into five subcellular localizations (chloroplast, thylakoid, mitochondrion, secretory pathway, and “other”), and is freely available for download, unlike other methods that perform a similar task. It is also freely available for academics on the web.

Results and Discussion

The prediction tool was tested using a five-fold cross-validation procedure (Table 1). The comparison of PredSL with TargetP cannot be made in an objective manner, as the datasets used for the training are different. However, comparing the results from the cross-validation of PredSL and the results of TargetP in Emanuelsson et al. (, we found that in most cases the sensitivity, specificity, and Matthew’s correlation coefficient (MCC) achieved by PredSL are comparable, if not better. Meanwhile, PredSL has the advantage of offering a prediction for the lumen targeting peptide (lTP). The results of the lTP prediction, compared to that of LumenP (, a module not included in the TargetP web server, are significantly better. The lTP prediction (Table 2) was tested on the whole dataset, resulting to 91.9% accuracy, compared to 88.8% provided by the LumenP predictor. The cleavage site prediction is also better: 88.7% of the sequences have their lTP cleavage sites predicted within ±2 residues, compared to 75.1% that the LumenP predictor estimates correctly. These results were based on the complete dataset of 259 sequences. On a redundancy reduced dataset of 109 sequences, which was obtained using cd-hit ( by removing sequences that had more than 40% sequence homology, we obtained 85.3% accuracy of prediction concerning the existence of the lTP, and 82.4% accuracy concerning the prediction of the cleavage site, which are better than the 82.4% and 70.1% accuracy of the LumenP predictor, respectively. When testing with a five-fold cross-validation, we obtained 87.3% accuracy of prediction concerning the existence of the lTP, and 66.1% accuracy concerning the prediction of the cleavage site, compared to the 87.0% and 54.8% accuracy of the LumenP predictor, respectively. Using a negative test set of 2,400 sequences of the proteins located in the chloroplast but not in the thylakoid lumen or the membrane, only 5.3% were identified with the lTP.

Table 1

Comparison of the Localization Performance of PredSL and TargetP Tested by Five-fold Cross-validation and Self-consistency*

Predictor set	Overall accuracy (%)	Category	Sensitivity	Specificity	MCC
A. PredSL (cross-validation/self-consistency)
Plant	86.7/88.3	cTP	0.90/0.90	0.80/0.91	0.82/0.88
		mTP	0.89/0.96	0.87/0.81	0.84/0.85
		SP	0.96/0.95	0.92/0.89	0.91/0.90
		other	0.70/0.72	0.86/0.95	0.74/0.79
Non-plant	87.1/92.5	mTP	0.88/0.91	0.84/0.96	0.80/90.5
		SP	0.94/0.95	0.91/0.91	0.89/0.90
		other	0.80/0.92	0.86/0.91	0.77/0.88
B. TargetP (cross-validation/self-consistency)
Plant	85.3/90.4	cTP	0.85/0.96	0.69/0.78	0.72/0.84
		mTP	0.82/0.88	0.90/0.95	0.77/0.88
		SP	0.91/0.94	0.95/0.94	0.90/0.92
		other	0.85/0.85	0.78/0.87	0.77/0.84
Non-plant	90.0/92.2	mTP	0.89/0.92	0.67/0.72	0.73/0.79
		SP	0.96/0.97	0.92/0.95	0.92/0.95
		other	0.88/0.90	0.97/0.97	0.82/0.86

The PredSL datasets for plant proteins consist of 249 chloroplast sequences, 250 mitochondrial sequences, and 253 secreted proteins’ sequences, whereas for non-plant proteins the datasets consist of 366 mitochondrial sequences and 370 secreted proteins’ sequences. The TargetP datasets for plant proteins consist of 141 chloroplast sequences, 368 mitochondrial sequences, and 269 secreted proteins’ sequences, whereas for non-plant proteins the datasets consist of 371 mitochondrial sequences and 715 secreted proteins’ sequences.

Table 2

Comparison of PredSL and LumenP on the Prediction Accuracy of the lTP and Its Cleavage Site*

Dataset	lTP prediction (%)		Cleavage site prediction (±2 residues) (%)
	PredSL	LumenP	PredSL	LumenP
Complete set (259 sequences)	91.9	88.8	88.7	75.1
Reduced set (40% similarity)	85.3	82.4	82.4	70.1
Cross-validation (259 sequences)	87.3	87.0	66.1	54.8

Tested by the five-fold cross-validation on the complete dataset (259 sequences) and on a 40% redundancy reduced dataset by cd-hit (109 sequences), respectively.

In an effort to compare PredSL with the already available prediction tools in a real dataset, we used the information from the YeastGFP database (http://yeastgfp.ucsf.edu). In this publicly available resource, the results of the global analysis of protein localization studies in the budding yeast Saccharomyces cerevisiae are deposited, which were obtained with the use of green fluorescent protein (GFP; ref. ). The comparison was made against TargetP, iPSORT, and Predotar, mainly because they are the most commonly used and successful prediction methods and are readily available to the public. TargetP is considered to be the most reliable tool currently available. It consists of two layers of feed-forward artificial neural networks. By using the N-terminal sequence of a protein, it classifies the protein into mitochondrion, chloroplast, secretory pathway, or “other” localization. iPSORT is one of the oldest available prediction methods (, which is based similarly to TargetP on identifying N-terminal sorting signals. However, it uses simple empirical rules (charge, hydrophobicity, etc.) that are amenable to human intervention whereas at the same time are interpretable to visual inspection. iPSORT is available to the public also as a web server and as a stand-alone executable application. Lastly, we used Predotar, which is also based on feed-forward neural networks for identifying the N-terminal targeting signals. Predotar is also available as a stand-alone application and its use was also dictated by the need to compare our method with other available stand-alone applications.

Table 3 shows that PredSL significantly outperforms the iPSORT algorithm and it also performs similarly to TargetP and Predotar. Based on the above results tested on an experimentally characterized genome, it can be assumed that the same will also hold, approximately, for other real-life applications of these software tools. Towards this end, we applied the PredSL algorithm in various completely sequenced genomes of organisms belonging to different taxonomic groups (plants, fungi, mammals, protozoa, arthropoda, and fishes). The results are presented in Table 4 and are generally in agreement with previously published estimates for the particular organisms. For instance, MITOPRED predicts 4.8%, 4.3%, 10%, 6.3%, and 4% mitochondrial proteins for Homo sapiens, Caenorhabditis elegans, S. cerevisiae, Drosophila melanogaster, and Arabidopsis thaliana, respectively. These estimates are clearly lower than the ones predicted by PredSL, but it is known that MITOPRED, which also uses profile HMMs derived from domains with known localization to the mitochondria, shows greater specificity in predicting proteins targeted to mitochondria (though without predicting any other localization). This fact was also acknowledged in the respective paper (, where the work of Kumar et al. ( was also cited, pointing to an overall estimate of 13% mitochondrial proteins in the yeast genome. Similar results were also obtained by the recent development of MitPred (, which also uses profile HMMs coupled with SVM. Chloroplast located proteins in A. thaliana predicted by TargetP were of the same magnitude (approximately 15% of the genome), whereas proteins targeted to thylakoid were also predicted by LumenP to be more rare (<1%) (. In another recent work (, among the 28,952 protein sequences of A. thaliana that were tested, TargetP and Predotar respectively predicted 4,780 and 4,582 proteins to be chloroplast localized. Among these proteins, 1,947 were predicted to be chloroplast localized by both predictors. These figures are also in agreement with the results obtained by PredSL and provide further evidence that, with the combination of the results of many independently developed and reliable predictors, we may have more specific estimates. PredSL thus, could be used in conjunction with the already established methods, and it would be interesting to perform large-scale analyses in order to discover the degree of concordance of the various predictors. In a completely new genome (that is, a genome of an organism with few homologous sequences to those already used to train the methods), it would be interesting to have the option of using various predictors and obtain different sets of proteins “predicted” to the various subcellular localizations. In this situation, the proteins predicted by all the algorithms would be the most “reliable”.

Table 3

Comparison of PredSL with Other Three Prediction Tools on the Subcellular Localization Prediction of the S. cerevisiae Proteins

Subcellular localization	PredSL	iPSORT	TargetP	Predotar
Total (unknown=2,164)	2,621/3,554 (73.7%)	2,404/3,554 (67.6%)	2,616/3,554 (71.6%)	2,475/3,554 (69.6%)
Mitochondrion	301/499 (60.3%)	304/499 (60.9%)	306/499 (61.3%)	315/499 (63.1%)
Secretory pathway	224/850 (26.4%)	206/850 (24.2%)	257/850 (26.4%)	204/850 (24.0%)
Other	2,096/2,305 (90.9%)	1,894/2,305 (82.2%)	2,053/2,305 (89.1%)	1,956/2,305 (84.9%)

Table 4

Prediction Performance of PredSL on Various Completely Sequenced Genomes from Different Taxonomic Groups

Group	Organism	cTP	lTP	mTP	SP	other	Total
Plants	Arabidopsis thaliana	4,596 (13.8%)*	184 (5.5%)	5,326 (16.0%)	8,191 (24.6%)	15,160 (45.6%)	33,273
	Thalassiosira pseudonana	813 (7.1%)	21 (0.2%)	1,406 (12.3%)	2,493 (21.9%)	6,686 (58.7%)	11,397
Fungi	Schizosaccharomyces pombe	–	–	586 (11.8%)	511 (10.3%)	3,890 (78.0%)	4,987
	Saccharomyces cerevisiae	–	–	566 (13.0%)	635 (14.5%)	3,167 (72.5%)	4,368
	Magnaporthe grisea	–	–	1,314 (11.8%)	2,364 (21.3%)	7,431 (66.9%)	11,109
Mammals	Homo sapiens	–	–	2,727 (9.4%)	7,221 (24.8%)	19,159 (65.8%)	29,107
	Mus musculus	–	–	3,353 (9.4%)	9,099 (25.5%)	23,274 (65.2%)	35,726
Protozoa	Plasmodium falciparum	–	–	314 (6.2%)	706 (14.0%)	4,029 (79.8%)	5,049
	Dictyostelium discoideum	–	–	644 (4.7%)	2,158 (15.8%)	10,878 (79.5%)	13,680
Arhthropoda	Drosophila melanogaster	–	–	1,949 (10.5%)	3,973 (21.5%)	12,576 (68.0%)	18,498
	Bombyx mori	–	–	1,627 (7.6%)	2,648 (12.4%)	17,027 (79.9%)	21,302
Fishes	Ciona intestinalis	–	–	1,383 (8.7%)	2,370 (15.0%)	12,099 (76.3%)	15,852
	Takifugu rubripes	–	–	1,617 (4.3%)	4,478 (12.0%)	31,344 (83.7%)	37,439

We list the total number of sequences classified in each subcellular location and their percentage in the whole genome.

In conclusion, PredSL is currently the only method that performs classification of eukaryotic proteins to the five subcellular localizations: chloroplast, thylakoid, mitochondrion, secretory pathway, and other. It uses a combination of several methods in order to exploit different properties of the amino acid sequence and results to a reliable classification of the proteins. When compared with other available methods such as TargetP, iPSORT, and Predotar, PredSL offers a comparable if not better reliability concerning the translocations of proteins to the chloroplast, the mitochondrion, and the secretory pathway. Moreover, it offers the option of a reliable prediction of protein targeting to the thylakoids of the chloroplast, a feature not offered by any other publicly available tool. Finally, it is one of the very few tools for subcellular localization prediction that is available for download as a stand-alone application, and it is the only one performing such a classification.

Materials and Methods

Datasets

The training data were extracted from the release 3.5 of UniProt (. The datasets were initially extracted requiring the keyword “Eukaryota” in the “OC” field of their entry. If in the same field there was the keyword “phyta” or “planta”, the sequences were included in the plant dataset. Otherwise, they were included in the non-plant dataset. Sequences that did not have an N-terminal methionine (M) were excluded from the datasets, as well as those including either B, Z, or X. The datasets were split into chloroplast, mitochondrial, secreted, cytoplasmic, and nuclear proteins. For the first three categories, it was required that they contain the keywords “TRANSIT” or “SIGNAL”, and “Chloroplast” or “Mitochondrial” in the “FT” field. For the others, it was required that they contain the keyword “Cytoplasm” or “Nucleus” in their “CC” field. Sequences that had more than one subcellular localization assigned to them were excluded from the datasets. After a 40% redundancy reduction using the program cd-hit (, a plant set consisting of 1,309 sequences (249 chloroplast, 62 mitochondrial, 422 secreted, 171 cytoplasmic, and 405 nuclear sequences) was obtained and also a non-plant set of 10,559 sequences (366 mitochondrial, 5,247 secreted, 1,458 cytoplasmic, and 3,488 nuclear sequences). The training set construction was based on having a positive set and a negative set of the same size, consisting of equal numbers of the other categories. Because of the small size of the plant mitochondrial set, we used the mitochondrial sequences from the non-plant set as well.

Algorithm

PredSL employs a combination of neural networks, Markov chains, scoring matrices (PrediSi), and HMMs, in order to get a number of scores for each sequence. These scores are then fed into a neural network, which makes the final prediction of the subcellular localization. More specifically, the algorithm follows a number of steps (Figure 1):

Fig. 1

An schematic overview of the PredSL algorithm.

1. Initially we create a first layer of two parallel neural networks (cTP and mTP) that distinguish between chloroplast and mitochondrial targeting sequences. For the training of these networks, we use the NevProp4 package (http://brain.unr.edu/publications/), which utilizes a version of the Quickprop algorithm (, having 4 hidden units and the input window size is 55 and 35 amino acid positions for each network, respectively. Each amino acid is represented by a set of 20 nodes with one of them switched on. For example, alanine is represented by the vector “10000000000000000000”. It results in 55×20=1100 and 35×20=700 input nodes for each network, respectively. Each network produces 100 output scores for the 100 N-terminal residues of each sequence.

2. The output scores for the 100 N-terminal residues of each sequence from each network are fed as input into two integrated neural networks, again one for each category (cTP and mTP), and are trained using the same sets without hidden units, thus providing two scores. Since at this stage we are not interested in the prediction of the cleavage site, the length of 100 is sufficient for the discrimination even though in many cases the targeting sequence is significantly shorter, and in some others longer.

3. Moreover, from the output scores of the first neural network, we calculate an approximate cutoff, under which the residues no longer belong to a targeting peptide, and we take a window of 40 amino acid positions for which we estimate the cleavage site of the cTP or mTP correspondingly, using a set of neural networks. For the training of these networks, we only use the sequences that have definite annotations concerning the cleavage site as positive data. It results in a reduction of the numbers of sequences in the cTP and mTP datasets to 122 and 31 correspondingly for the plant set and the mTP dataset, and to 241 for the non-plant set. We use equal numbers of positive and negative sequences, and the negative sequences consist of equal representation of each of the negative datasets. Due to the small number of sequences used in the negative dataset, for each case (cTP and mTP) we train five neural networks with different sets of sequences as negative, in order to make the prediction more objective. For training, after experimenting initially using several window sizes, we choose the ones with the best performance; for the prediction of the cTP cleavage site it is 27 residues around the cleavage site (−20, +6), and for the prediction of the mTP cleavage site it is 21 residues around the cleavage site (−12, +8). The output score is the average of those of the five neural networks.

4. We calculate the average of the output scores of the first neural network from the first position to the predicted cleavage site for each set of networks and obtain two more scores.

5. Next we use the PrediSi algorithm ( available at http://www.predisi.de, which predicts signal peptides for secreted proteins using weight matrices, getting one more score.

6. Subsequently, we use a combination of positive and negative sets to train a simple first-order Markov chain to distinguish sequences between the two classes (. The combinations used for the plant set as positive/negative are chloroplast/cytoplasmic, mitochondrial/cytoplasmic, chloroplast/mitochondrial, mitochondrial & chloroplast/cytoplasmic, mitochondrial & chloroplast/secreted, chloroplast/mitochondrial & cytoplasmic, and for the non-plant set are mitochondrial/cytoplasmic, mitochondrial/secreted, mitochondrial/cytoplasmic & secreted. Thus, we acquire six scores for the plant proteins and three for the non-plant.

7. Finally, using the release 2.2 of HMMER (, we generate profiles for the cTP and mTP and search the 120 N-terminal residues of the sequences for them. The length of the 120 N-terminal residues is used to ensure that we take into account most of the targeting sequences, even if they are longer than the average length. The alignments used to generate the profiles are created as follows: For the cTP, the first 15 residues are aligned in the left, the next 8 residues are aligned in the middle, and the rest, up to the cleavage site plus 9 residues, are aligned to the right. For the mTP, the first 10 residues are aligned to the left, and the rest, up to the cleavage site plus 9 residues, are aligned to the right. In this manner we get two more scores for each sequence. Therefore, we have 13 scores for the plant sequences and 7 for the non-plant sequences.

8. We feed these scores to a final neural network, which has been trained with equal numbers of sequences from each category (chloroplastic, mitochondrial, secreted, and other) and using six hidden units for the plant set and three for the non-plant set. The output consists of one score corresponding to each subcellular localization (chloroplast, mitochondrion, and secretory pathway). If all the scores are less than a certain cutoff empirically determined, the sequence is classified as “other”. Otherwise the prediction is based on the largest score.

9. The proteins predicted as chloroplast are then searched for certain HMM-profiles from position 30 to 120 for the existence of the lTP. We search from position 30 of the precursor sequence to avoid false positives, since it is improbable that there will be an lTP within the 30 N-terminal residues, and we terminate the search in position 120 since the average length of a cTP is between 30–90 residues. Therefore, the lTP should in theory be detected within the 120 N-terminal residues. The profiles are created using the datasets from LumenP ( by aligning with ClustalW (, having 35 residues previous to the lTP cleavage site plus 6 positions after that. Three such profiles are used: one from the proteins following the Sec-pathway, one from those following the TAT-pathway, and one including all proteins. If even one output score from searching the sequence is above the cutoff (−4), the protein is then classified as “thylakoid”.

Cleavage site prediction

Depending on the prediction of the subcellular localization, we use the corresponding cleavage site prediction from the neural networks (cTP and mTP) or PrediSi ( as a final prediction of the cleavage site. For the lTP, we use the prediction obtained by aligning the significant hits to the model generated by HMMER (.