Key points are not available for this paper at this time.
Here we report a systematic approach for predicting subcellular localization (cytoplasm, mitochondrial, nuclear, and plasma membrane) of human proteins. First, support vector machine (SVM)-based modules for predicting subcellular localization using traditional amino acid and dipeptide (i + 1) composition achieved overall accuracy of 76.6 and 77.8%, respectively. PSI-BLAST, when carried out using a similarity-based search against a nonredundant data base of experimentally annotated proteins, yielded 73.3% accuracy. To gain further insight, a hybrid module (hybrid1) was developed based on amino acid composition, dipeptide composition, and similarity information and attained better accuracy of 84.9%. In addition, SVM modules based on a different higher order dipeptide i.e. i + 2, i + 3, and i + 4 were also constructed for the prediction of subcellular localization of human proteins, and overall accuracy of 79.7, 77.5, and 77.1% was accomplished, respectively. Furthermore, another SVM module hybrid2 was developed using traditional dipeptide (i + 1) and higher order dipeptide (i + 2, i + 3, and i + 4) compositions, which gave an overall accuracy of 81.3%. We also developed SVM module hybrid3 based on amino acid composition, traditional and higher order dipeptide compositions, and PSI-BLAST output and achieved an overall accuracy of 84.4%. A Web server HSLPred (www.imtech.res.in/raghava/hslpred/ or bioinformatics.uams.edu/raghava/hslpred/) has been designed to predict subcellular localization of human proteins using the above approaches. Here we report a systematic approach for predicting subcellular localization (cytoplasm, mitochondrial, nuclear, and plasma membrane) of human proteins. First, support vector machine (SVM)-based modules for predicting subcellular localization using traditional amino acid and dipeptide (i + 1) composition achieved overall accuracy of 76.6 and 77.8%, respectively. PSI-BLAST, when carried out using a similarity-based search against a nonredundant data base of experimentally annotated proteins, yielded 73.3% accuracy. To gain further insight, a hybrid module (hybrid1) was developed based on amino acid composition, dipeptide composition, and similarity information and attained better accuracy of 84.9%. In addition, SVM modules based on a different higher order dipeptide i.e. i + 2, i + 3, and i + 4 were also constructed for the prediction of subcellular localization of human proteins, and overall accuracy of 79.7, 77.5, and 77.1% was accomplished, respectively. Furthermore, another SVM module hybrid2 was developed using traditional dipeptide (i + 1) and higher order dipeptide (i + 2, i + 3, and i + 4) compositions, which gave an overall accuracy of 81.3%. We also developed SVM module hybrid3 based on amino acid composition, traditional and higher order dipeptide compositions, and PSI-BLAST output and achieved an overall accuracy of 84.4%. A Web server HSLPred (www.imtech.res.in/raghava/hslpred/ or bioinformatics.uams.edu/raghava/hslpred/) has been designed to predict subcellular localization of human proteins using the above approaches. The successful completion of a human genome project has yielded huge amount of sequence data. Analysis of this data to extract the biological information can have profound implications on biomedical research. Therefore, mining of biological information or functional annotation of piled up sequence data is a major challenge to the modern scientific community. Determination of functions of all of these proteins using experimental approaches is a difficult and time-consuming task. Traditionally, the similarity search-based tools has been used for functional annotations of proteins (1Reinhardt A. Hubbard T. Nucleic Acids Res. 1998; 26: 2230-2236Crossref PubMed Scopus (533) Google Scholar). This approach fails when unknown query protein does not have significant homology to proteins of known functions. The functions of the proteins are closely related to its cellular attributes, such as subcellular localization and its association with the lipid bilayer (subcellular localization) (2Alberts B. Bray D. Lewis J. Raff M. Roberts K. Watson J.D. Molecular Biology of the Cell. 3rd Ed. Garland Publishing, New York1994: 1255-1272Google Scholar, 3Lodish H. Baltimore D. Berk A. Zipursky S.L. Matsudaira P. Darnell J. Molecular Cell Biology. 3rd Ed. Scientific American Books, New York1995: 739-777Google Scholar); hence, the related proteins must be localized in the same cellular compartment to cooperate toward a common function (4Nair R. Rost B. Nucleic Acids Res. 2003; 13: 3337-3340Crossref Scopus (34) Google Scholar). In addition, information on the localization of proteins with known function may provide insight about its involvement in specific metabolic pathways (5Himmelreich R. Hilbert H. Plagens H. Pirkl E. Li B.C. Herrmann R. Nucleic Acids Res. 1996; 24: 4420-4449Crossref PubMed Scopus (961) Google Scholar, 6Nakai K. Kanehisa M. Proteins. 1991; 11: 95-110Crossref PubMed Scopus (642) Google Scholar, 7Nakai K. Kanehisa M. Genomics. 1992; 14: 897-911Crossref PubMed Scopus (1376) Google Scholar). Therefore, an attempt has been made to predict subcellular localization of proteins to elucidate the function. Several methods have been devised earlier to predict the subcellular localization of the eukaryotic and prokaryotic proteins using different approaches and data sets (8Hua S. Sun Z. Bioinformatics. 2001; 17: 721-728Crossref PubMed Scopus (763) Google Scholar). The most commonly used approach utilizes alignment or similarity search against an experimentally annotated data base. But this approach fails in the absence of significant similarity between the query and target protein sequences (1Reinhardt A. Hubbard T. Nucleic Acids Res. 1998; 26: 2230-2236Crossref PubMed Scopus (533) Google Scholar). Another popular approach is based on identification of sequence motifs such as signal peptide or nuclear localization signal (9Fujiwara Y. Asogawa M. Genome Inform. Ser. Workshop Genome Inform. 2001; 12: 103-112PubMed Google Scholar). This approach has been limited by the observation that all of the proteins residing in a compartment do not have universal motif. To overcome these limitations, several machine learning technique-based methods, such as artificial neural networks and support vector machines (SVMs), 1The abbreviations used are: SVM, support vector machine; MARS, multivariate adaptive regression splines; RI, reliability index. have been developed to predict the subcellular localization of proteins. These methods are based on the several features of protein sequences such as recognition of N-terminal sorting signals or the composition of amino acids. These methods predict subcellular localization either for prokaryotic or eukaryotic proteins, such as PSORT (10Bannai H. Tamada Y. Maruyama O. Nakai K. Miyana S. Bioinformatics. 2002; 18: 298-305Crossref PubMed Scopus (601) Google Scholar) and TargetP (11Emanuelsson O. Nielsen H. Brunak S. von Heijne G. J. Mol. Biol. 2000; 300: 1005-1016Crossref PubMed Scopus (3666) Google Scholar) for eukaryotes and SubLoc (8Hua S. Sun Z. Bioinformatics. 2001; 17: 721-728Crossref PubMed Scopus (763) Google Scholar) and NNPSL (1Reinhardt A. Hubbard T. Nucleic Acids Res. 1998; 26: 2230-2236Crossref PubMed Scopus (533) Google Scholar) for both prokaryotes and eukaryotes, with good accuracy (>70%). Recently, our group has also developed a new hybrid approach-based method, ESLPred, which predicts the four major subcellular localizations (nuclear, cytoplasmic, mitochondrial, and extracellular) of eukaryotic proteins with an overall accuracy of 88% (12Bhasin M. Raghava G.P.S. Nucleic Acids Res. 2004; 32: 415-419Crossref PubMed Scopus (62) Google Scholar). To the best of our knowledge, there is no method for the prediction of subcellular localization of human proteins. Availability of sequence data of human genes in recent years demands a reliable and accurate method for prediction of subcellular localization of human proteins. In the present study, a systematic attempt has been made to develop a method for the subcellular localization of human proteins. The SVM modules based on different features of the proteins such as amino acid composition and dipeptide composition of proteins have been constructed. In addition, a similarity search-based module, HuPSI-BLAST, has also been developed, using PSI-BLAST to predict the localization of human proteins. Further, SVM module “hybrid1” has been developed using amino acid composition, traditional dipeptide composition, and results of PSI-BLAST prediction. The SVM modules based on higher order dipeptide compositions (i + 2, i + 3, and i + 4) and combinations of various feature-based modules have also been constructed. Here we have also compared the performance of the present organism-specific method (HSLPred) with ESLPred (12Bhasin M. Raghava G.P.S. Nucleic Acids Res. 2004; 32: 415-419Crossref PubMed Scopus (62) Google Scholar), a general method for prediction of subcellular localization of eukaryotic proteins. In addition, the performance of HSLPred has also been assessed on various mammalian and nonmammalian genomes and on an independent data set. It was observed that this method can predict the subcellular localization of human proteins and proteins from related genomes with high accuracy. In other words, our method can also be used for the prediction of subcellular localization of mammalian proteins. The Data Set—The data set of human proteins with experimentally annotated subcellular localization has been derived from release 44.1 of the SWISSPROT data base (13Bairoch A. Apweiler R. Nucleic Acids Res. 2000; 28: 45-48Crossref PubMed Scopus (2445) Google Scholar). Of 10,777 human proteins available in the data base, subcellular localization information was available for 7910 sequences. These 7910 sequences were screened strictly in order to develop a high quality data set for predicting subcellular localization of human proteins. The sequences annotated as “fragments,” “isoforms,” “potential,” “by similarity,” or “probable” were filtered out from the data set. Further, sequences residing in more than one subcellular location (such as a protein sequence labeled with “nuclear and cytoplasmic” or “mitochondrial and cytoplasmic”) were also excluded from the data set. The sequence redundancy of data set was further reduced by using PROSET software (14Brendel V. Math. Comput. Modelling. 1991; 16: 37-43Crossref Scopus (33) Google Scholar) such that no two sequences had >90% sequence identity in the data set. The final data set consists of 3780 protein sequences that belong to 11 subcellular locations as shown in Table I. The number of sequences for the last seven subcellular locations was not sufficient for developing a prediction method. Therefore, a method was developed for only four major subcellular locations of human proteins (840 cytoplasmic, 315 mitochondrial, 858 nuclear, and 1519 plasma membrane).Table INumber of sequences within each subcellular location groupSubcellular locationNumber of in a new machine learning support vector machine has been used for the prediction of subcellular localization of human proteins. SVM has been used for the of and protein prediction D. M. D. S. A. 2000; PubMed Scopus Google Scholar, P. A. Bioinformatics. 2002; PubMed Scopus Google Scholar, 2003; Scholar). In the present study, a of SVM, has been used to predict the subcellular localization of proteins. The prediction of subcellular localization is a Therefore, for have been constructed. the number was to four for human proteins. The SVM was with all of the in the with and for proteins of subcellular This of SVM is known as one SVM (8Hua S. Sun Z. Bioinformatics. 2001; 17: 721-728Crossref PubMed Scopus (763) Google Scholar). In this four were constructed for the subcellular localization of human proteins. unknown was the that to the SVM with output We have different approaches based on different features of a such as amino acid composition and dipeptide composition, in the acid composition is the of each amino acid in a This the order of amino acids. The of all amino was using of amino acid number of amino acid number of amino in 1) i can be amino (i + composition was used to the information about each protein which a of This the information of the amino acid composition with the order of amino acids. The of each dipeptide was using of number of number of all (i + 1) is one of In addition, to the of the with the and in the higher order such as i + 2, i + 3, and i + were using of number number of all 2, 3, or and (i + is one of this study, we also made an attempt to a and reliable machine learning multivariate adaptive regression for predicting subcellular localization of human proteins. It has been shown that as as other machine learning such as neural In addition, also information about the of different for the or 1991; Google Scholar, Comput. Scopus Google Scholar). In the present study, we have the regression software that the of for the on the Web This of a of we have used compositions of amino acid for the prediction amino acid and dipeptide of have used commonly used of amino amino and amino and amino and amino and and amino and In order to the composition of a we compositions of its compositions of be compositions of and module was designed to predict subcellular localization of human proteins, in which the query sequence was against data base of human proteins using The data base consists of sequences to four major subcellular locations mitochondrial, nuclear, and plasma The subcellular localization of these proteins has been The PSI-BLAST was used of the to search the data base, has the to J. Mol. Biol. PubMed Scopus Google Scholar). It out an search in which the sequences in one of search are used to a for the of of PSI-BLAST were carried out a of This module predict of the four localizations mitochondrial, nuclear, or plasma the similarity of the query protein to the proteins present in the data base. The module subcellular no significant similarity was SVM our group has the of a hybrid SVM module for the prediction of subcellular localization of eukaryotic proteins (12Bhasin M. Raghava G.P.S. Nucleic Acids Res. 2004; 32: 415-419Crossref PubMed Scopus (62) Google Scholar). In the present study, an attempt has been made to the of hybrid modules by hybrid modules based on different approaches. The of the approaches used to develop different hybrid modules is SVM SVM module the information of amino acid composition, traditional dipeptide composition, and PSI-BLAST output SVM was with an vector of that of for amino acid composition, for dipeptide composition, and for PSI-BLAST The PSI-BLAST output was to using the shown in 4) SVM hybrid2 SVM module was constructed using all higher order dipeptide compositions (i + 2, i + 3, and i + 4) with traditional dipeptide composition (i + This hybrid2 module was with an vector of from each dipeptide composition SVM hybrid3 SVM module was constructed using amino acid composition, traditional dipeptide composition (i + higher order dipeptide compositions (i + 2, i + 3, and i + and similarity search-based results The module was with vector of for amino acid compositions, for the above four of dipeptide compositions, and for PSI-BLAST of performance of SVM modules constructed in this report was using a In this a data set was The and was carried out each using one set for and the four sets for To the the accuracy and were as by and Sun (8Hua S. Sun Z. Bioinformatics. 2001; 17: 721-728Crossref PubMed Scopus (763) Google Scholar), using and can be subcellular location mitochondrial, nuclear, or plasma is the number of sequences observed in location is the number of sequences of location is the number of sequences not of location is the number of and is the number of sequences. reliability is a commonly used of prediction that about the to the The is a of the of in the for the The used for the is to that used in the by our group (12Bhasin M. Raghava G.P.S. Nucleic Acids Res. 2004; 32: 415-419Crossref PubMed Scopus (62) Google Scholar). The was to the between the and SVM output The reliability for the approach-based module was using In order to the performance of HSLPred and to with other methods such as ESLPred (12Bhasin M. Raghava G.P.S. Nucleic Acids Res. 2004; 32: 415-419Crossref PubMed Scopus (62) Google Scholar), two other data sets were also A is as Data such as and are used for the performance of method. the best of the performance of a developed method is to on an independent data set that the used of the method. data were derived from the of the SWISSPROT data base (13Bairoch A. Apweiler R. Nucleic Acids Res. 2000; 28: 45-48Crossref PubMed Scopus (2445) Google Scholar). This data set human proteins cytoplasmic, 11 mitochondrial, nuclear, and plasma membrane) and was not used in the and of the HSLPred method. ESLPred Data the performance of the present method (HSLPred) with ESLPred, another method developed by our group for subcellular localizations of eukaryotic proteins (12Bhasin M. Raghava G.P.S. Nucleic Acids Res. 2004; 32: 415-419Crossref PubMed Scopus (62) Google Scholar), the data set of ESLPred was ESLPred was on eukaryotic proteins nuclear, cytoplasmic, mitochondrial, and This data set was further two mammalian and nonmammalian proteins other proteins, to the performance of HSLPred on these two different In addition, the data sets of other mammalian genomes such as and have also been from the of the SWISSPROT data base (13Bairoch A. Apweiler R. Nucleic Acids Res. 2000; 28: 45-48Crossref PubMed Scopus (2445) Google Scholar), to the of HSLPred on other closely related The data set used is shown in Table of the genome has sequences of more than genes are of four of and and more proteins that are made up of different of amino acids. The four of in various different information for the specific proteins that the up of each human other proteins human and and provide to In order to its each protein must be to its or subcellular localization is a of each functional have been developed to predict subcellular localization of proteins, based on amino acid compositions Proteins. PubMed Scopus Google Scholar), neural (1Reinhardt A. Hubbard T. Nucleic Acids Res. 1998; 26: 2230-2236Crossref PubMed Scopus (533) Google Scholar), D. 12: PubMed Scopus Google Scholar), Z. PubMed Scopus Google Scholar), and support vector machines (8Hua S. Sun Z. Bioinformatics. 2001; 17: 721-728Crossref PubMed Scopus (763) Google Scholar, J. Biol. 2002; PubMed Scopus Google Scholar). Recently, K. M. I. S. K. Nakai K. Nucleic Acids Res. 2003; 13: Scopus Google Scholar) have developed a that several methods for the prediction of subcellular localization for proteins. In artificial such as SVM and artificial neural networks are as approaches for the prediction of subcellular localization of proteins. The performance of all of the SVM modules developed in this has been a The SVM has been carried out by the of various function and the of the The results using various function have been shown in the Table It has been observed that the better than and in the of the amino acid SVM for all of the SVM modules developed in the present study, the has been The amino acid SVM module 2, 1) has been to an overall accuracy of for all of the four subcellular localizations Further, to information about as as the order of an SVM module based on traditional dipeptide compositions has been constructed. The traditional dipeptide (i + 1) SVM module has achieved the best results with the This accuracy is better than the amino acid SVM The performance of amino and traditional dipeptide SVM modules in different subcellular localizations has been shown in Table performance of various SVM modules developed using different features of a protein and i + + + in a new The homology of a protein with other related sequences a of information about the similarity search-based module has been constructed to information of the proteins. no significant have been for of proteins. Therefore, the performance of this module is in with amino acid as as dipeptide This module has cytoplasmic, mitochondrial, nuclear, and plasma subcellular localizations with and and achieved an overall accuracy of 73.3% It that compositions acid and can the data more in with the similarity search-based To the prediction such as have been devised to more information of the proteins. The hybrid module, has been constructed using amino acid composition, traditional dipeptide composition, and PSI-BLAST The module with the 2, 1) has achieved overall accuracy of which is better of the modules developed in this These results that prediction accuracy of subcellular localization of proteins can be using a of information about a In addition, higher order dipeptide (i + 2, i + 3, and i + 4) SVM modules have been constructed to the of different of amino on the subcellular The overall performance of higher order dipeptide compositions in predicting subcellular localization is shown in The (i + dipeptide SVM module has achieved an overall accuracy of higher in with the traditional and other higher order dipeptide SVM It has also been observed that an accuracy of i + dipeptide modules is more for proteins, and for the subcellular localizations nuclear, and plasma is with traditional dipeptide Further, the performance of i + and i + 4 dipeptide modules has been to be to the traditional dipeptide SVM module the i + dipeptide module has achieved better accuracy in with traditional dipeptide composition, different hybrid modules have been constructed with an to the overall accuracy. The SVM module hybrid2 has been constructed using all higher order dipeptide compositions (i + 2, i + 3, and i + 4) with traditional dipeptide The overall accuracy of the hybrid2 SVM module is than the module, is higher in with traditional dipeptide This that is to more which is in the proteins of different subcellular Furthermore, another SVM module hybrid3 has been constructed using amino acid compositions, traditional dipeptide compositions (i + higher order dipeptide compositions (i + 2, i + 3, and i + and PSI-BLAST the hybrid3 SVM module has been with an overall accuracy of which is to the in accuracy be achieved to the of the hybrid3 module has been with an vector of In to hybrid a approach has also been to the human proteins with better accuracy. The SVM consists of two of SVM The consists of based on traditional and higher order dipeptide compositions (i + i + 2, i + 3, and i + and the consists of an SVM that the output of the and a final The SVM module has been to an accuracy of with the performance of the hybrid2 A of the of all of the SVM modules developed on the of different approaches is shown in To the prediction has been carried out for the SVM It the of an approach in the prediction of subcellular localization of proteins. The is a of in the prediction. the accuracy and of a prediction with an of We have the prediction accuracy of proteins an than or to shown in Table of the HSLPred has been to predict of sequences with an prediction accuracy of This that a can predict a number of sequences with higher accuracy for HSLPred has been to predict sequences with an accuracy of for The of the present was to develop a method for the subcellular localization of human proteins. the present method has been on the specific proteins, be more accurate and better for the in with methods such as ESLPred, developed for all eukaryotic proteins. The has been to the of HSLPred methods such as First, the performance of HSLPred has been on proteins used to develop the ESLPred method. The approach of the HSLPred method has been to predict cytoplasmic, mitochondrial, and nuclear proteins with an accuracy of and and an overall accuracy of has been The have been in the Table in order to the performance of the ESLPred method on human proteins, we have the ESLPred method on proteins used to develop It has been observed that the approach of ESLPred cytoplasmic, mitochondrial, and nuclear proteins with an accuracy of and respectively. overall accuracy of has been Table of the These results that the performance of an organism-specific HSLPred method is better than ESLPred for predicting human proteins. Furthermore, in order to the performance of HSLPred in with ESLPred on eukaryotic proteins, the data set used to develop the ESLPred method has been two mammalian and nonmammalian eukaryotic proteins other than proteins. These two sets have been further using the HSLPred We that the HSLPred method has achieved an overall accuracy of and for mammalian and nonmammalian protein as shown in the Table It that HSLPred can predict mammalian proteins with good accuracy and proteins with accuracy. Further, the performance of both HSLPred and ESLPred has been assessed on an independent data set to the performance of a method. It has been observed that HSLPred has been to predict and proteins out of and mitochondrial, nuclear, and plasma using the overall accuracy of has been the ESLPred method has been to an overall accuracy of The results have been shown in Table of the In the performance of HSLPred has been to be better both and of an independent data that is not an We have also the of the HSLPred with other genomes such as and to the performance of HSLPred on other closely related It has been observed that HSLPred also predicts other mammalian proteins with high accuracy. The results have been shown in Table of the the HSLPred method can also be used for the prediction of subcellular localization of other closely related mammalian proteins. In other words, can as a method for various closely related mammalian SVM and artificial neural networks are for the of proteins, have limitations, these results that are difficult to subcellular localization has from a number of amino acid composition, homology to other localized proteins, and localization the of results can provide new protein subcellular In the present study, we also used the 1991; Google Scholar, Comput. Scopus Google Scholar) for the of subcellular localization of human proteins using of amino acids. It has been observed that for the of proteins, composition of amino and an for the of proteins, the of and and and amino has been In the of nuclear proteins, composition of amino and and for the plasma proteins composition of amino and has been to be The results have been shown in Table of the Further, cytoplasmic, mitochondrial, nuclear, and plasma proteins have achieved accuracy of and and an overall accuracy of has been In order to for this either to the of or the of we have developed an SVM module based on the used for We observed that the accuracy achieved by the SVM module was better than MARS, that is also a for the of proteins. we to that performance of can be further amino acid or dipeptide compositions are used as HSLPred of SVM modules constructed in the present have been on a Web server (HSLPred) using The HSLPred server is available on the Web or can a protein sequence in one of the such as or The server to various approaches for the prediction of the subcellular localization of a query In the of the the module for prediction. overall of the HSLPred server is shown in We for and D. for with
Garg et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: