Prediction of Mitochondrial Proteins Using Support Vector Machine and Hidden Markov Model

Mitochondria are considered as one of the core organelles of eukaryotic cells hence prediction of mitochondrial proteins is one of the major challenges in the field of genome annotation. This study describes a method, MitPred, developed for predicting mitochondrial proteins with high accuracy. The data set used in this study was obtained from Guda, C., Fahy, E. & Subramaniam, S. (2004) Bioinformatics 20, 1785–1794. First support vector machine-based modules/methods were developed using amino acid and dipeptide composition of proteins and achieved accuracy of 78.37 and 79.38%, respectively. The accuracy of prediction further improved to 83.74% when split amino acid composition (25 N-terminal, 25 C-terminal, and remaining residues) of proteins was used. Then BLAST search and support vector machine-based method were combined to get 88.22% accuracy. Finally we developed a hybrid approach that combined hidden Markov model profiles of domains (exclusively found in mitochondrial proteins) and the support vector machine-based method. We were able to predict mitochondrial protein with 100% specificity at a 56.36% sensitivity rate and with 80.50% specificity at 98.95% sensitivity. The method estimated 9.01, 6.35, 4.84, 3.95, and 4.25% of proteins as mitochondrial in Saccharomyces cerevisiae, Drosophila melanogaster, Caenorhabditis elegans, mouse, and human proteomes, respectively. MitPred was developed on the above hybrid approach. Mitochondria are considered as one of the core organelles of eukaryotic cells hence prediction of mitochondrial proteins is one of the major challenges in the field of genome annotation. This study describes a method, MitPred, developed for predicting mitochondrial proteins with high accuracy. The data set used in this study was obtained from Guda, C., Fahy, E. & Subramaniam, S. (2004) Bioinformatics 20, 1785–1794. First support vector machine-based modules/methods were developed using amino acid and dipeptide composition of proteins and achieved accuracy of 78.37 and 79.38%, respectively. The accuracy of prediction further improved to 83.74% when split amino acid composition (25 N-terminal, 25 C-terminal, and remaining residues) of proteins was used. Then BLAST search and support vector machine-based method were combined to get 88.22% accuracy. Finally we developed a hybrid approach that combined hidden Markov model profiles of domains (exclusively found in mitochondrial proteins) and the support vector machine-based method. We were able to predict mitochondrial protein with 100% specificity at a 56.36% sensitivity rate and with 80.50% specificity at 98.95% sensitivity. The method estimated 9.01, 6.35, 4.84, 3.95, and 4.25% of proteins as mitochondrial in Saccharomyces cerevisiae, Drosophila melanogaster, Caenorhabditis elegans, mouse, and human proteomes, respectively. MitPred was developed on the above hybrid approach. The mitochondrion, popularly known as the power house of the cell, is the central unit of eukaryotic cells. 2MitPred was developed on a hybrid approach to predict mitochondrial proteins, and additional data are available upon request. It is a double membrane-bounded organelle with two spaces, the outer intermembrane space and inner matrix. Due to the presence of an explicit mitochondrial genome, unlike other organelles, its function is regulated by two genomes. It performs a plethora of biochemical reactions like oxidative phosphorylation, Krebs cycle, β-oxidation of fatty acids, DNA replication, transcription, translation, etc., some of which occur in mitochondria only. In addition mitochondria are also involved in apoptosis (1Gottlieb R.A. Crit. Rev. Eukaryot. Gene Expr. 2000; 10: 231-239Crossref PubMed Scopus (37) Google Scholar) and ionic homeostasis (2Jassem W. Fuggle S.V. Rela M. Koo D.D. Heaton N.D. Transplantation. 2002; 73: 493-499Crossref PubMed Scopus (204) Google Scholar). Because of their multidimensional utility, mitochondrial proteins are associated with several human diseases, including Alzheimer disease (3Hutchin T. Cortopassi G.A. Proc. Natl. Acad. Sci. U. S. A. 1995; 92: 6892-6895Crossref PubMed Scopus (177) Google Scholar), Type II diabetes (4Gerbitz K.D. Gempel K. Brdiczka D. Diabetes. 1996; 45: 113-126Crossref PubMed Google Scholar) and Parkinson disease (5Orth M. Schapira A.H. Neurochem. Int. 2002; 40: 533-541Crossref PubMed Scopus (179) Google Scholar). A majority of mitochondrial proteins are synthesized in cytoplasm from where they are transported inside mitochondria. But a small number of mitochondria-resident proteins are also synthesized inside mitochondria by the mitochondrial genome. Proteins that are imported to mitochondria contain a leader sequence at the N terminus that contains all the information needed to localize to mitochondria (6Alberts B. Johnson A. Lewis J. Raff M. Roberts K. Walter P. Molecular Biology of the Cell.4th Ed. Garland Science, New York2000: 678Google Scholar). But this is not true for all mitochondrial proteins. In many cases the leader sequence is altogether absent. This poses a major challenge in predicting mitochondrial proteins in silico. In the past, a number of methods have been developed to predict the mitochondrial proteins, although most were not intended exclusively for mitochondrial proteins. Existing prediction methods can be divided into four categories. The similarity search-based techniques fall under the first category in which the query sequence is searched against experimentally annotated proteins. If the query protein has significant sequence similarity with any mitochondrial protein then it is predicted as a mitochondrial protein. But this method fails to predict new/novel proteins if these proteins do not have similarity with known proteins. Although the similarity-based method is very informative and considered to be best, it becomes severely handicapped when no apparent homology is found (7Reinhardt A. Hubbard T. Nucleic Acids Res. 1998; 26: 2230-2236Crossref PubMed Scopus (532) Google Scholar). In the second category, the methods are based on predicting signal sequences in proteins. A number of methods fall under the second category where sorting signals, present on the protein itself, are used for prediction. This category includes TargetP (8Emanuelsson O. Nielsen H. Brunak S. Heijne G. J. Mol. Biol. 2000; 300: 1005-1016Crossref PubMed Scopus (3638) Google Scholar), SignalP (9Bannai H. Tamada Y. Maruyama O. Nakai K. Miyano S. Bioinformatics. 2002; 18: 298-305Crossref PubMed Scopus (590) Google Scholar), and PSORT II (10Gardy J.L. Spencer C. Wang K. Ester M. Tusnady G.E. Simon I. Hua S. deFays K. Lambert C. Nakai K. Brinkman F.S. Nucleic Acids Res. 2003; 31: 3613-3617Crossref PubMed Scopus (362) Google Scholar). Although these methods are quite popular, their major limitation is that not all proteins have signals; for example, only around 25% of yeast mitochondrial proteins have “matrix-targeting signals” particularly at the N terminus (11Marcotte E.M. Xenarios I. van Der Bliek A.M. Eisenberg D. Proc. Natl. Acad. Sci. U. S. A. 2000; 97: 12115-12120Crossref PubMed Scopus (191) Google Scholar). Because of this, these methods fail to predict the proteins that do not have a signal. In the third category, methods attempt to predict subcellular localization on the basis of sequence composition. Some popular methods in this category are ESLpred (12Bhasin M. Raghava G.P.S. Nucleic Acids Res. 2004; 32: W414-W419Crossref PubMed Scopus (274) Google Scholar), HSLpred (13Garg A. Bhasin M. Raghava G.P.S. J. Biol. Chem. 2005; 280: 14427-14432Abstract Full Text Full Text PDF PubMed Scopus (204) Google Scholar), NNPSL (7Reinhardt A. Hubbard T. Nucleic Acids Res. 1998; 26: 2230-2236Crossref PubMed Scopus (532) Google Scholar), and LOCSVMPSI (14Xie D. Li A. Wang M. Fan Z. Feng H. Nucleic Acids Res. 2005; 33: W105-110Crossref PubMed Scopus (151) Google Scholar). Although their overall performance is very good, accuracy of prediction of mitochondrial proteins is much lower than for proteins in other locations. It shows that mitochondrial protein localization is much more complex than seen otherwise. Hence prediction of mitochondrial proteins warrants special attention. Recently Guda et al. (15Guda C. Fahy E. Subramaniam S. Bioinformatics. 2004; 20: 1785-1794Crossref PubMed Scopus (115) Google Scholar) developed a method, MITOPRED, which falls under the fourth category of prediction methods. This method is exclusively developed for predicting mitochondrial proteins with a maximum accuracy of 92.3% (Mathews’ correlation coefficient (MCC), 3The abbreviations used are: MCC, Mathews' correlation coefficient; SVM, support vector machine; HMM, hidden Markov model; SAAC, split amino acid composition. 0.638). Its output is the combined score of two scoring methods that assign an arbitrary score on the basis of existence of Pfam domains, differences in the amino acid composition, and the pI value of the protein. In the present study we tried to improve the prediction accuracy of mitochondrial proteins. First we carried out systematic analysis of amino acid composition of both mitochondrial and non-mitochondrial proteins, and then on the basis of the conclusion drawn we developed the prediction method. In our study we used a powerful machine learning technique, support vector machine (SVM), for classifying proteins instead of the pI score used in MITOPRED. Data Sets and Evaluation—We used the same data set that was constructed to develop MITOPRED by Guda et. al (15Guda C. Fahy E. Subramaniam S. Bioinformatics. 2004; 20: 1785-1794Crossref PubMed Scopus (115) Google Scholar). This data set contain 1432 mitochondrial and 8940 non-mitochondrial sequences. First both mitochondrial and non-mitochondrial proteins were randomly divided into five parts. Each of these five sets consists of one-fifth of mitochondrial (∼ 287) and one-fifth of non-mitochondrial (∼ 1788) proteins. For training and for testing and evaluating our methods, we used a 5-fold cross-validation technique in which four sets were used for training and the remaining set was used for testing. This process was repeated five times so that each set was used once for testing (16Kaur H. Raghava G.P.S. Protein Sci. 2003; 12: 627-634Crossref PubMed Scopus (124) Google Scholar). Amino Acid and Dipeptide of the composition of proteins is to the of protein sequences to This is an and most of proteins using machine learning techniques they In addition the of a protein sequence to a vector of using amino acid composition the of the protein into the In addition to amino acid composition, dipeptide composition was also used for that a of The of dipeptide composition amino acid composition is that it information the of amino as as their The amino acid as as dipeptide composition was as by et al. (13Garg A. Bhasin M. Raghava G.P.S. J. Biol. Chem. 2005; 280: 14427-14432Abstract Full Text Full Text PDF PubMed Scopus (204) Google Scholar). were used as to mitochondrial and non-mitochondrial proteins using Amino Acid the of SAAC, protein sequences were by a of instead of in the of amino acid composition. In each protein is divided into 25 amino of the N 25 amino of the and the these The using this is the that composition of the sequence not to which is known to be present in mitochondrial protein Hence the of amino acid composition is that it to proteins that have a signal at the N is a that mitochondrial proteins contain signal sequence at their N and this is used in TargetP for prediction (8Emanuelsson O. Nielsen H. Brunak S. Heijne G. J. Mol. Biol. 2000; 300: 1005-1016Crossref PubMed Scopus (3638) Google Scholar). In this we a to model signal. It was by a vector of (25 the 25 This vector was used to the proteins using of and BLAST W. W. J. Mol. Biol. PubMed Scopus Google Scholar) is the most used for similarity In this study we the performance of BLAST in mitochondrial proteins using at an value of In this, proteins of the sets were used as the query against a data the proteins of training The performance of BLAST was in of for mitochondrial proteins. In the the search was combined with the method. Although BLAST the were used only for proteins where BLAST not any significant the most significant was a non-mitochondrial protein and We also repeated the same at If a protein has a with value than the then the BLAST prediction was the prediction was of Pfam Markov are a of that are to sequences. It describes a number of sequences. it is a more and of the Markov In this was using The sequence search was using the Pfam data A. S. A. M. S. C. Nucleic Acids Res. 2004; 32: PubMed Google Scholar), which is a data of sequence of proteins to the same It contains each a Pfam Each mitochondrial and non-mitochondrial protein in our data set was searched against the Pfam data using the search at an value of were to of exclusively mitochondrial domains only in mitochondrial proteins, exclusively non-mitochondrial domains only in non-mitochondrial proteins, and domains in both of proteins. A protein was as a mitochondrial protein if it contains one mitochondrial the hybrid approach and the search were combined to the of both prediction by and the similarity search technique in a annotated and data Because domains are and of proteins, it is known that all proteins are of Hence can be a major the localization of a protein that a on function of a protein. But all the domains are not where not a is found can be a In these cases of prediction can be a In this the hybrid approach be a of prediction. First proteins in the set were searched against a data of mitochondrial and non-mitochondrial A protein was as a mitochondrial protein if it has an mitochondrial and was as a non-mitochondrial protein if it has an non-mitochondrial But if the protein not have any then the method was used for prediction. on Data is the most popular method to performance of a prediction method. But it has been in the that performance of cross-validation is not M. Raghava G.P.S. Bioinformatics. 2004; 20: PubMed Scopus Google Scholar). the performance of any method one to it on an data this in we the performance of our method on an data set from which is a data of mitochondrial proteins A. Nucleic Acids Res. 2005; 33: PubMed Scopus Google Scholar). It contains and mitochondrial proteins of Caenorhabditis elegans, and mouse, respectively. of eukaryotic were from the Bioinformatics cerevisiae, C. elegans, and Drosophila and two and these proteomes, by using the hybrid approach of the Pfam search and SVM, an of the number of mitochondrial proteins was carried this study we by using the T. in B. C. A. Scholar), which to a number of and basis and any that we have a number of with the value is a mitochondrial non-mitochondrial the into space with on the training The function by can be as The value of is by the of the to C. is the that the the and the training and is the for the The of is very in SVM, to in In this learning was carried out using and the performance of methods we used several used in these of and The is a of these The sensitivity of mitochondrial proteins is the of mitochondrial proteins predicted as mitochondrial proteins. The specificity of non-mitochondrial proteins is the of non-mitochondrial proteins predicted as non-mitochondrial proteins. The accuracy is the of predicted proteins. can be using where and are predicted and proteins, and are predicted mitochondrial and non-mitochondrial proteins, respectively. is considered to be the most of any prediction method. to is as a is for a the above have a that they the performance at a A known is which is a true and The under the a value to the performance of a method. of the for predicting the function of a protein is to a similarity search against a data of annotated proteins. In this study we used BLAST for predicting mitochondrial proteins using 5-fold cross-validation where four sets of mitochondrial and non-mitochondrial proteins were used to a BLAST data and mitochondrial proteins of the set were searched against this BLAST data This process was repeated five times so the BLAST search was once for each mitochondrial protein. in the performance of BLAST at from to with an of This that BLAST predict all mitochondrial of BLAST search on data set used for of mitochondrial of BLAST in a of Amino Acid and Dipeptide has been in the that amino acid composition can be to proteins. We amino acid composition of mitochondrial and non-mitochondrial proteins It was that amino acid composition of mitochondrial proteins was from that of non-mitochondrial proteins. it is to mitochondrial proteins from other proteins. an was developed using amino acid composition as vector of and of were tried and achieved accuracy of with an of using basis function where sensitivity and specificity is the same It has also been in the that dipeptide methods are more than amino acid methods in of proteins M. Raghava G.P.S. J. Biol. Chem. 2004; Full Text Full Text PDF PubMed Scopus Google Scholar). an was developed to predict mitochondrial proteins using dipeptide composition as vector of But to our the performance of the dipeptide method was only than the amino acid method Although the was also found to be above the of prediction are of We also under the for and achieved under the of and for amino acid and dipeptide composition, of with acid in a composition of the first amino acid of mitochondrial and non-mitochondrial of BLAST and the similarity search-based method and the machine learning method have their and methods are than any other method if the query sequence has significant similarity with any experimentally annotated protein. But it fails in the of the other the machine learning method is a method in which the performance not on similarity and query sequences. improve the performance both the BLAST search and were In this, the training set was as the data and the proteins from set were used as query proteins to search against this data Each protein from the set was used to query against the data all the only the was considered as If the to a mitochondrial protein then the query protein was as a mitochondrial protein. if the to a non-mitochondrial protein then the query protein was as a non-mitochondrial protein. Although the overall performance of the similarity search was quite good, it was that was no for a number of proteins. For these proteins was used to predict it is a mitochondrial non-mitochondrial protein. In this study from BLAST were used to assign The maximum accuracy using this was around with around which is than the performance of the BLAST search This that the of the similarity search and machine learning techniques can very high accuracy if used to each other M. A. Raghava G.P.S. Bioinformatics. 2005; PubMed Scopus Google of the combined approach of BLAST and value in a of and it has been that around 25% of proteins have an it is to the amino acid composition of mitochondrial and non-mitochondrial proteins. Recently Guda et al. (15Guda C. Fahy E. Subramaniam S. Bioinformatics. 2004; 20: 1785-1794Crossref PubMed Scopus (115) Google Scholar) the 25 of the N terminus and the remaining protein the 25 of the N for proteins of locations. a in amino acid composition It was also found that the composition of and remaining of mitochondrial proteins was We their study to the amino acid composition of mitochondrial proteins as 20, and remaining of the protein the 20, and residues) 20, and and remaining of the protein the 20, and residues) It was that the composition of from to shows the same This was also the when composition was into the these two were it was found that is a the composition at the N terminus of the protein and the remaining of the protein. Although at the terminus have a composition then the of the it is not as significant as for the We amino acid composition analysis for non-mitochondrial proteins. We not any significant amino acid composition of N-terminal, C-terminal, and remaining of non-mitochondrial proteins. We also the composition of mitochondrial and non-mitochondrial proteins. in of mitochondrial proteins are quite from of non-mitochondrial proteins. on on the above we developed a method using were by for for 25 a of 25 was used in which a at a be by vector of M. Raghava G.P.S. J. Biol. Chem. 2004; Full Text Full Text PDF PubMed Scopus Google Scholar). The based on 20, and were developed and achieved a maximum from to on of N-terminal, C-terminal, and Protein was that the amino acid composition of the N-terminal, C-terminal, and remaining mitochondrial protein was from that of non-mitochondrial proteins. we developed based on the composition of with vector 20, and remaining with vector and and remaining using vector The maximum of based on composition of 20, and from to which is much than the based on This was to as the sequence is to contain the composition has the number of It that of is not for mitochondrial signals, their presence is this is to the known This be all mitochondrial proteins do not have a leader which is the to the the sequence and their The performance of with composition and remaining composition vector from an of to that is than based on protein. Because the performance of using was very we it from further The performance of split amino acid composition C-terminal, and remaining residues) SAAC, where vector was an from to Pfam and the method developed by Guda et al. (15Guda C. Fahy E. Subramaniam S. Bioinformatics. 2004; 20: 1785-1794Crossref PubMed Scopus (115) Google Scholar) prediction of mitochondrial proteins was on the basis of of Pfam We a in this study and proteins in our data set were searched using against the Pfam data and a of domains was domains were found exclusively in were found exclusively in and were found in both of proteins. A data was that consists of all of predict a protein can be in mitochondria we an search against the data A protein was as mitochondrial if it has an mitochondrial as non-mitochondrial protein if it has an non-mitochondrial this approach mitochondrial and non-mitochondrial proteins were It was found that was no for a number of proteins to the of an mitochondrial non-mitochondrial to no present at we developed a hybrid method that the based on split amino acid composition and the of Pfam In the hybrid method a protein is predicted to be mitochondrial non-mitochondrial if it has an In the where a protein not have any the based on was used for prediction. The performance of this hybrid method was at of from to of Pfam search and in a on Data our study M. Raghava G.P.S. Bioinformatics. 2004; 20: PubMed Scopus Google Scholar), we a in the performance of a method on data used for testing and training testing 5-fold it is to a developed method on an data set for The data set used in this study consists of C. elegans, and mitochondrial proteins obtained from A. Nucleic Acids Res. 2005; 33: PubMed Scopus Google Scholar). method predicted and proteins to C. elegans, and mouse, of this data set as mitochondrial at the same sequences were for prediction on the MITOPRED at for C. elegans, and mouse, and proteins, were predicted as mitochondrial proteins It has been by Guda et al. (15Guda C. Fahy E. Subramaniam S. Bioinformatics. 2004; 20: 1785-1794Crossref PubMed Scopus (115) Google Scholar) that MITOPRED is than methods like PSORT and In their study they have that MITOPRED shows performance than other prediction methods like PSORT and But the performance on the data set shows that our method has than MITOPRED the that both are developed on the same data The this be the that we have used both and composition of proteins with the composition of the remaining protein which is a than amino acid composition as by the performance of of MitPred and MITOPRED on mitochondrial proteins from number of of proteins predicted as in a of proteomes, from yeast to more complex C. elegans, and mouse, for annotation. The number of proteins predicted as mitochondrial is in For MitPred predicted proteins as mitochondrial the which is of the This is than have estimated Guda et al. (15Guda C. Fahy E. Subramaniam S. Bioinformatics. 2004; 20: 1785-1794Crossref PubMed Scopus (115) Google Scholar) and et al. (11Marcotte E.M. Xenarios I. van Der Bliek A.M. Eisenberg D. Proc. Natl. Acad. Sci. U. S. A. 2000; 97: 12115-12120Crossref PubMed Scopus (191) Google Scholar) estimated of the proteins as experimentally A. J. Y. C. B. I. A. B. P. C. Proc. Natl. Acad. Sci. U. S. A. 2003; PubMed Scopus Google the is to the that we the most to out the For Drosophila and C. elegans, the number of mitochondrial proteins predicted was and from the of and proteins. In Drosophila MITOPRED has estimated that is mitochondrial this with that of our For C. although Guda et al. (15Guda C. Fahy E. Subramaniam S. Bioinformatics. 2004; 20: 1785-1794Crossref PubMed Scopus (115) Google Scholar) estimated et al. (11Marcotte E.M. Xenarios I. van Der Bliek A.M. Eisenberg D. Proc. Natl. Acad. Sci. U. S. A. 2000; 97: 12115-12120Crossref PubMed Scopus (191) Google Scholar) the other for and human and proteins were predicted as mitochondrial proteins from the set of and proteins, respectively. In the of human our method is with the of proteins by et al. Fahy E. 2003; Full Text Full Text PDF PubMed Scopus Google Scholar), et al. E. A. A. M. 2000; PubMed Scopus (191) Google Scholar), and Guda et al. (15Guda C. Fahy E. Subramaniam S. Bioinformatics. 2004; 20: 1785-1794Crossref PubMed Scopus (115) Google Scholar). Recently et al. T. Bioinformatics. 2005; PubMed Scopus Google Scholar) have predicted the mitochondrial proteins in human by using a very and approach. they first predicted the proteins of yeast subcellular localization was to be mitochondria. these proteins as the by using they the human proteins that are to be In addition by using several they predicted human mitochondrial proteins that homology with yeast mitochondrial proteins. We that one of the the in the number of predicted mitochondrial proteins the method and MITOPRED is the the of proteins in the data from the Bioinformatics method is available on the in the of a The is available upon request. The can a protein sequence in any as The has the to any of the SVM, BLAST SVM, and Pfam search with