- Research article
- Open Access
IsoSVM – Distinguishing isoforms and paralogs on the protein level
© Spitzer et al; licensee BioMed Central Ltd. 2006
- Received: 18 July 2005
- Accepted: 06 March 2006
- Published: 06 March 2006
Recent progress in cDNA and EST sequencing is yielding a deluge of sequence data. Like database search results and proteome databases, this data gives rise to inferred protein sequences without ready access to the underlying genomic data. Analysis of this information (e.g. for EST clustering or phylogenetic reconstruction from proteome data) is hampered because it is not known if two protein sequences are isoforms (splice variants) or not (i.e. paralogs/orthologs). However, even without knowing the intron/exon structure, visual analysis of the pattern of similarity across the alignment of the two protein sequences is usually helpful since paralogs and orthologs feature substitutions with respect to each other, as opposed to isoforms, which do not.
The IsoSVM tool introduces an automated approach to identifying isoforms on the protein level using a support vector machine (SVM) classifier. Based on three specific features used as input of the SVM classifier, it is possible to automatically identify isoforms with little effort and with an accuracy of more than 97%. We show that the SVM is superior to a radial basis function network and to a linear classifier. As an example application we use IsoSVM to estimate that a set of Xenopus laevis EST clusters consists of approximately 81% cases where sequences are each other's paralogs and 19% cases where sequences are each other's isoforms. The number of isoforms and paralogs in this allotetraploid species is of interest in the study of evolution.
We developed an SVM classifier that can be used to distinguish isoforms from paralogs with high accuracy and without access to the genomic data. It can be used to analyze, for example, EST data and database search results. Our software is freely available on the Web, under the name IsoSVM.
- Support Vector Machine
- Alternative Splice
- Radial Basis Function
- Testing Dataset
- Support Vector Machine Classifier
Typical eukaryotic genes are composed of several relatively short exons that are interrupted by long introns. The primary transcripts of most eukaryotic genes are composed of introns and exons separated by canonical splice sites. These mRNA precursors are shortened by a process called RNA splicing in which the intron sequences are removed yielding the mature transcript consisting of exons only . However, cells can splice the primary transcript in different ways and thereby generate different polypeptides from the same gene (reviewed in ). This process is called alternative splicing. The different polypeptides are termed alternatively spliced gene products, splice variants or protein isoforms .
To generate correctly spliced, mature mRNAs, the exons must be identified and joined together precisely and efficiently by a complex process that requires the coordinated action of five small nuclear RNAs (termed U1, U2 and U4 to U6) and more than 60 polypeptides . According to , five common modes of alternative splicing are known: (i) exon skipping or inclusion, (ii) alternative 3' splice sites, (iii) alternative 5' splice sites, (iv) mutually exclusive exons, and (v) intron retention which corresponds to no splicing. In complex pre-mRNAs, more than one of these modes of alternative splicing can apply to different regions of the transcript, and extra mRNA isoforms can be generated through the use of alternative promoters or polyadenylation sites .
Alternative splicing is a frequent process in eukaryotes. It is estimated that up to 60 percent of human genes are subjected to alternative splicing . Thus, alternative splicing is probably an important source of protein diversity in higher eukaryotes. For example, the fruitfly Drosophila melanogaster contains fewer genes than Caenorhabditis elegans while exhibiting significantly higher protein diversity . Furthermore, alternative splicing of primary transcripts is often tissue- or stage-specific (cf. the expression of different alternatively spliced transcripts during different stages of the development of an organism ), and is thus an important regulatory mechanism.
Available databases of proteins and their isoforms consider only a small number of protein families and species (see e.g. [6–8]). We wanted to identify isoforms without knowledge of genomic information and independently of specific protein families or species, in a fashion well suited for high-throughput genomics and proteomics.
For automation the approach of supervised learning using a Support Vector Machine (SVM) [9–11] was chosen. SVMs are gaining popularity in Bioinformatics [12–15] and are often superior to Neural Networks and Bayesian Learning . SVM classifiers distinguish two classes of input data by calculating separating hyperplanes (decision surfaces) in a vector space V that is endowed with a dot product. The dot product is used as a measure of similarity. Data samples from the input space are mapped to the vector space V (usually of dimensionality higher than the input space), making it easier to find a separating hyperplane. The position and margin of the hyperplane are optimized in V, maximizing the distance of the hyperplane to instances of both classes. The kernel function used to measure similarity behaves in input space like the dot product in space V. Thus, similarity of input data can be measured easily in V. Without a kernel function, computation of the dot products in V would be necessary, consuming a large amount of time, depending on the structure of V. For an in-depth description of properties and theory of SVMs, please see . The Support Vector Machine implementation SVMLight  was used. In this paper, we introduce a highly accurate SVM-based method to distinguish between isoforms and paralogs on the protein level (that is, without the need for genomic information). Our software is freely available on the Web (see Conclusions).
Importance of maximizing accuracy in distinguishing isoforms and paralogs
Why does isoform detection require such a high degree of accuracy? Why do we want to use an SVM even though this approach is usually employed in case the input space has dimensionality (much) larger than three? For example, when performing 2,000 sequence comparisons, even a 0.2% improvement in accuracy results in 4 fewer misclassifications. Such numbers are typical, for example, in applications of our automated phylogeny pipeline RiPE [18, 19]. Analyzing a large protein family with RiPE, few misclassifications make a difference since paralogs misidentified as isoforms (false positives) are deleted from the dataset, which may result in the loss of key members of the protein family, compromising the interpretation of the evolution of sequence, domain structure and function. (In this specific application, isoforms misidentified as paralogs (false negatives) do not pose a major problem.)
Performance statistics of different classifiers based on three features
Mean accuracy and standard error of the mean of various classifiers, using three features derived from the alignment of the sequences to be compared. 100-fold jackknife resampling was employed. "± " denotes the standard error of the mean.
99.55% ± 0.008
99.31% ± 0.015
1897.1 ± 0.21
1887.9 ± 0.28
13.1 ± 0.28
3.9 ± 0.21
RBF network classifier
99.33% ± 0.011
98.91% ± 0.019
1896.5 ± 0.22
1880.1 ± 0.38
20.9 ± 0.38
4.6 ± 0.22
3-feature linear classifier
99.42% ± 0.011
99.22% ± 0.020
1893.8 ± 0.35
1886.0 ± 0.39
15.0 ± 0.39
7.2 ± 0.35
Performance of different classifiers using a canonical training/testing dataset
In the following, we report results that are not supported by resampling but derived from a specific ("canonical") training and testing dataset (cf. Methods, section Canonical training and testing dataset). In this way, we were able to explore, on a large (3802 samples) dataset, a wide variety of classifiers in reasonable time.
Performance of the SVM classifier (accuracy/precision) on four testing scenarios.
Full-length-sequence (canonical testing dataset)
Selected Xenopus EST data
ABC protein homologous-regions-only
Performance comparison of the three-feature SVM classifier to linear classifiers, an RBF network classifier and other SVM classifiers, using canonical training and testing datasets.
Canonical testing dataset
Homologous-regions-only testing dataset
3-feature SVM classifier
Sequence similarity, inverse CBIN count, match/mismatch fraction (cf. Table 2)
2-feature SVM classifiers
Match/mismatch fraction, sequence similarity
Inverse CBIN count, sequence similarity
Match/mismatch fraction, inverse CBIN count
RBF Network classifier
Sequence similarity, inverse CBIN count, match/mismatch fraction
3-feature linear classifier
Sequence similarity, inverse CBIN count, match/mismatch fraction
2-feature linear classifiers
Match/mismatch fraction, sequence similarity
Inverse CBIN count, sequence similarity
Match/mismatch fraction, inverse CBIN count
1-feature linear classifiers
Inverse CBIN count
A linear classifier that was calculated using all three features of the samples in the canonical training dataset was found to classify the canonical testing dataset with an accuracy of 99.42%. Linear classifiers that were trained using all possible combinations of only two features showed at least slightly inferior results compared to the linear classifier based on all three features. Not surprisingly, the best-performing classifier based on two features does not use the weakest feature that is sequence similarity. Classifiers based on sequence similarity alone appear to be weak in distinguishing between isoforms and paralogs and perform much worse than any other of the tested classifiers; a linear classifier derived by line-sweeping using the feature sequence similarity alone results in an accuracy of approximately 82%. Linear classifiers based on one of the other features perform surprisingly well, however (cf. Table 3).
Finally, the radial basis function (RBF) network classifier  (cf. Methods , section Training of the radial basis function network) applied to the canonical testing dataset using all three features results in an accuracy of 99.32%.
Application of the SVM classifier to EST data
Summary of Xenopus EST cleanup and clustering.
Total number of ESTs and cDNAs
Number of good sequences
Average trimmed EST length (bp)
Number of clusters
Number of singletons
Number of CAP3 contigs
Number of CAP3 singletons
Average CAP3 contig length (bp)
Max. cluster size (no. of ESTs)
Average cluster size (no. of ESTs)
4,097 – 8,192
2,049 – 4,096
1,025 – 2,048
513 – 1,024
257 – 512
129 – 256
65 – 128
33 – 64
17 – 32
9 – 16
5 – 8
3 – 4
To assess whether the splitting of clusters by CAP3 into several contigs was caused by grouping isoforms into the same cluster, or whether the splitting was due to paralogs, we extracted 722 clusters that have multiple contigs (2,243 contigs total), and for which each contig has a full length protein match in the protein NR database . Most of the 722 clusters consist of only two contigs and only a fraction features three or more contigs. Treating each contig consensus as a sequence, 5,459 sequence pairs were compared by IsoSVM within clusters; 986 of these samples (19.3%) were classified as isoforms and 4,125 as paralogs (80.7%). 348 samples were left out, representing contigs with almost no overlap, i.e. sequence pairs of low (<1%) similarity. As a further check, to assess the accuracy of this analysis, 290 randomly chosen samples were reviewed manually and the result was noted (cf. Table 2); an accuracy of 97.93% and a precision of 99.23% was found. (In a few cases, early EST sequencing termination events produce a block of amino acids aligned with gaps at the end of the two sequences compared, causing classification of such cases as isoforms, and they were counted as such.)
Application of the SVM classifier to an automated phylogeny pipeline
As a second application, the classifier was incorporated into a pipeline for automatic generation of protein phylogenies called RiPE [18, 19], with the aim to further reduce the redundancy of the RiPE-retrieved protein data by recognizing and deleting sequences that are isoforms. Isoforms are usually considered irrelevant data in phylogenetic tree inference and analysis. RiPE data are generated by homology search (PSIBLAST, ), retrieving hits with putative homology to a search profile and assembling HSP-based homologous-regions-only data as described in Methods , section Homologous regions only. The pipeline already features a redundancy minimization stage, sorting out hits that are similar to other hits (95% identity or more). The IsoSVM classifier was incorporated, enabling the detection and deletion of isoforms, thus decreasing dataset size and redundancy while simultaneously increasing computational speed and legibility of the phylogenetic tree. We first tested the ability of our classifier to deal with homologous-regions-only data (using the testing dataset described in Methods, section Homologous regions only), noting an accuracy of 98.98% and a precision of 97.57% (cf. Table 2). Training on homologous-regions-only data did not improve classifier performance (data not shown).
Following our interest in ABC (A TP-b inding c assette) proteins, which are found in a wide variety of species and are of major biomedical importance, a dataset of 1,349 ABC protein hits was then retrieved by RiPE from 20 model proteomes (12 eukaryotes, 6 bacteria and 2 archaea) using 48 known human ABC proteins  as search profile. 115 hits were identified as isoforms of another hit by the SVM classifier. As a further check, all 115 putative isoforms were inspected visually, the automatic classification (isoform or paralog) was checked, and a precision of 95.65% was found. The accuracy of the classifier was not calculated in this case since RiPE reports only samples classified as positives (i.e. isoforms). While the precision reported is based on the number of false positives (i.e. sequences representing paralogous sequences being reported as isoforms), assessment of accuracy would require the visual inspection of tens of thousands of samples of (putative) paralogs, i.e. putative false negatives. Removal of isoforms resulted in a reduction of dataset size by about 8%, rendering the eukaryotic parts of the tree much more legible.
Limitations of the classifier
Despite showing reliable performance, the SVM classifier is not perfect. It may misleadingly classify a small portion of paralogs with high similarity as isoforms, since they feature long stretches of identical amino acid sequence. Further, sequences that are fragments of other sequences will be classified as isoforms.
The SVM classifier, trained using visually classified cases of isoform and paralog relationships, proved to be reliable in all tests, exhibiting an accuracy of over 97% and a precision of over 95%. We are thus able to distinguish isoforms and paralogs in a satisfactory way, no matter whether full-length, homologous-regions-only or EST cluster sequences are handled. In particular, for species such as Xenopus laevis, for which few detailed analyses of the evolution of genes and proteins exist, the analysis of paralogs and isoforms can improve statistical models of sequence evolution, e.g. regarding the likelihood of gene duplication and alternative splicing. Overall, the IsoSVM tool should be useful for researchers in several fields of genomic research and EST analysis as a reliable method of automatic isoform identification. Our software is freely available at the IsoSVM Website , under an open source license.
To automatically determine if one protein sequence is an isoform of another, we first derive three features, characterizing the degree and pattern of matches and mismatches in a pairwise alignment of the two sequences as detailed in the paragraphs below. The three features depend on the length of the alignment of the two sequences and c onsecutive b locks of i dentities or n on-identities (CBINs).
Length of the alignment (l)
The length of the alignment of two protein sequences a and b is used in two of the features described below to normalize their values to a range from 0 to 1. This was done in order to avoid numerical problems that may affect classification performance and to exclude features of large absolute amount that may numerically dominate smaller ones during training of the SVM (cf. [29, 30]).
Consecutive blocks of identities or non-identities (CBIN)
A CBIN is a block in which the alignment features consecutive matches or mismatches (cf. Figure 3). Few large CBINs are characteristic for comparisons of isoforms whereas many short CBINs are typically found in comparisons of paralogs (cf. Figure 1, illustrating the comparison of two isoforms and two paralogs).
There are two possible cases of a CBIN. First, if sequence a features a subsequence of length c starting at position i (with c between 1 and l-i) that is a maximum run of exact matches (that cannot be extended any further) to its aligned counterpart of sequence b, then this block of consecutive matches is a CBIN of length c. Second, if sequence a features a subsequence of length c starting at position i (with c between 1 and l-i) that is a maximum run of mismatches to its aligned counterpart of sequence b, then this block of consecutive mismatches is a CBIN of length c. Formally, for internal CBINs that are not located at the beginning or at the end of the alignment, we have
a k = b k for all k,k = i,...,i+c and ai-1≠ bi-1 and ai+c+1≠bi+c+1
a k ≠ b k for all k,k = i,...,i+c and ai-1= bi-1 and ai+c+1= bi+c+1 (1)
where i is the start coordinate and i+c the end coordinate of the maximum block of matches or mismatches. For CBINs that are not internal, the definition can be generalized in an obvious way. Amino acids aligned with gaps are considered mismatches.
Sequence similarity is the overall number of matches in the alignment of the sequences a and b, divided by its length l:
where |M| denotes the number of elements in a set M.
Inverse CBIN count
As the second feature we us the reciprocal value of the number of CBINs n in the pair of aligned sequences:
Fraction of consecutive matches and mismatches
This feature describes the overall number of consecutive matches and mismatches (not counting the match or mismatch at the first position of a CBIN). In other words, it is the sum of the lengths c j minus one, of all n CBINs (with j = 1..n), divided by l:
The feature fraction of consecutive matches and mismatches is abbreviated as match-mismatch fraction in all figures and tables. In the following we describe the procedure of the generation of the training and testing datasets, the learning pipeline and the validation of classifier performance.
Generation of the training and testing datasets
Sequence retrieval, homology search and visual classification
The NCBI non-redundant (NR) database  was used as the source for retrieving protein sequences and was downloaded from the NCBI FTP server on March 8, 2004. The NR database was then searched for sequences annotated as "isoform" or "splice variant". 13,061 sequences featuring at least one of the two keywords were found and retrieved from the NR database, establishing a set of unrelated sequences that are from any species for which isoforms can be expected to exist. From this set, 250 sequences were randomly selected to give rise to the canonical training and testing datasets, as follows (for a complete list of taxa included in this set please consult the supplementary material [see Additional file 1]).
We observe large blocks of (almost) identical sequence with no (or few) mismatches that can be interpreted as common exons, except for a few sequencing errors or polymorphisms.
Additionally, we observe either one or both of the following:
We observe one or more sequence blocks that do not match (interspersed with a few random matches) which can be interpreted as mutually exclusive exons of similar size that are spuriously aligned and which are embedded in blocks of (almost) identical sequence.
We observe one or more sequence blocks that align to gap characters which can be interpreted as surplus amino acids that arise if mutually exclusive exons of different length are spuriously aligned, or if exon(s) are missing in one of the sequences, or if an exon has an alternative splice site such that it is observed in a short and in a long version, and which are again embedded in blocks of (almost) identical sequence.
In contrast, two sequences are classified as paralogs if there is a large sequence block that displays sufficient similarity to allow assumption of common evolutionary origin, interspersed with a sufficiently large number of mismatches that must be interpreted as substitutions and that cannot be interpreted as sequencing errors, etc. Paralogs may feature deletions that give rise to observations similar to the ones in (i) and (ii) which are however embedded in blocks of sufficient similarity with many mismatches.
Canonical training and testing dataset
The dataset resulting from visual inspection featured 3,802 samples of the isoform class and 8,757 of the paralog class. We started training with many more paralogs than isoforms, with inferior testing results (data not shown). Therefore, to prevent one class from outweighing the other during SVM training, the number of samples of the larger class was truncated to 3,802 samples. One half of the dataset, consisting of 1,901 isoform and 1,901 paralog samples, was designated the canonical training dataset, the other half is the canonical testing dataset. As can be seen from Figure 2, the two classes separate quite well, although close inspection reveals that the boundary between them is in fact quite complex.
Homologous regions only
Another testing dataset was generated directly from the database search reports obtained before. They were converted into FASTA-formatted alignments of merged HSPs (partial hits called high-scoring segment pairs) using MVIEW . These merged HSPs can be viewed as the concatenation of the homologous regions of the full hit sequences. Some of the queries contained internal repeats that do not give rise to a single concatenation; these sequences were left out. By automatically transferring the visual classification of the corresponding full-length-sequence-based samples above to the merged HSP data, a set of 8,066 classified samples was obtained (5,518 samples of the paralog and 2,548 samples of the isoform class).
Training of the SVM
To find an optimum SVM classifier for a given problem, a kernel has to be specified. As kernel function the radial basis function (RBF) kernel was used. For SVMs with RBF kernels, two parameters, C and g need to be determined. C describes a penalty for training errors and is part of the soft margin concept of SVMs. It allows for a number of (misclassified) training samples to be located within the margin. Thus, a certain amount of noise is tolerated in the training data. The parameter g describes the width of the Gaussian bells of the radial basis function of the RBF kernel
where x i , x j denote feature vectors of training samples. We scanned for best parameter values in a specific range using a so-called grid-search.
and precision (cf. )
Training of the radial basis function network
To compare the performance of the SVM classifier to another machine learning technique, a neural network classifier (more precisely a radial basis function (RBF) network ) was trained on the canonical training dataset. The implementation of RBF networks with adaptive centers by  was used with default values (number of centers 3; regularization 10-4; iterations for optimization 10).
Assessing performance of classifiers based on three features by jackknife resampling
To estimate the mean accuracy and standard error of the mean of a classifier, it was trained and tested on datasets derived from random splits of the canonical samples derived from Genbank using a 100-fold jackknife resampling process . More specifically, the canonical training and testing datasets described above were concatenated yielding a dataset of 7,604 samples, with 3,802 samples of each class. For each jackknife run, 1,901 samples of each class were chosen randomly from this dataset for training, while the remaining samples were used for testing. The mean accuracy and the standard error of the mean (σ/, where σ denotes the standard deviation and N the number of jackknife resamplings) were calculated.
Sequence similarity: 0.01...0.05
Inverse CBIN count: 0.01...0.03
Fraction of consecutive matches and mismatches: 0.90...0.94
Although line-sweeping is not exhaustive, the best combination of thresholds found in the reduced search space should represent the optimum; these are 0.01832 for sequence similarity, 0.01613 for inverse CBIN count and 0.92827 for the fraction of consecutive matches and mismatches.
Accuracy, precision and true positive/true negative (TP/TN) and false positive/false negative (FP/FN) ratios were averaged over all jackknife runs and the standard error of the mean of each of these properties was calculated (cf. Table 1 and Figure 4).
Classifiers based on fewer features, thresholds and parameters; measuring performance
Performance of the classifiers based on three features was compared to the performance of classifiers based on a reduced set of two or only one feature(s), using the canonical training and testing datasets only. In contrast to the studies using resampling, all linear classifiers were derived by exhaustive line sweeping, that is, by an exhaustive search for the best combination of thresholds or the best single threshold in case of one feature. The thresholds for linear classifiers are listed in the supplementary data, Tables S1 and S2 [see Additional file 1]. The kernel parameters (cf. Methods, section Training of the SVM) for SVM classifiers based on canonical training datasets are listed in Table S3 of the supplementary data [see Additional file 1]. Performance (in terms of accuracy) of all classifiers was noted on canonical testing datasets and homologous-regions-only datasets and is given in Table 3.
We would like to thank the Interdisciplinary Center for Clinical Research, Münster, for partial funding of this work, Karl Grosse-Vogelsang, Integrated Functional Genomics, Münster, for maintaining and providing access to a 16-node x86-cluster, enabling the calculation of countless grid-searches in acceptable time, and Martin Eisenacher, Integrated Functional Genomics, Münster, for advice on statistics and linear classifiers.
- Alberts B, Johnson A, Lewis J, Raff M, Roberts K, Walter P: Molecular Biology of the Cell. 4th edition. Garland Publishing, New York; 2000.Google Scholar
- Graveley BR: Alternative splicing: increasing diversity in the proteomic world. Trends Genet 2001, 17(2):100–107.View ArticlePubMedGoogle Scholar
- Cartegni L, Chew SL, Krainer AR: Listening to silence and understanding nonsense: exonic mutations that affect splicing. Nature Reviews Genetics 2002, 3: 285–298.View ArticlePubMedGoogle Scholar
- Grabowski PJ, Black DL: Alternative RNA splicing in the nervous system. Prog Neurobiol 2001, 65(3):289–308.View ArticlePubMedGoogle Scholar
- Fitch WM: Distinguishing homologous from analogous proteins. Syst Zool 1970, 19(2):99–113.View ArticlePubMedGoogle Scholar
- Lee C, Atanelov L, Modrek B, Xing Y: ASAP: The Alternative Splicing Annotation Project. Nucl Acids Res 2003, 31: 101–105.PubMed CentralView ArticlePubMedGoogle Scholar
- Pospisil H, Herrmann A, Bortfeldt R, Reich J: EASED: Extended Alternatively Spliced EST Database. Nucl Acids Res 2004, 32: D70–74.PubMed CentralView ArticlePubMedGoogle Scholar
- Thanaraj TA, Stamm S, Clark F, Riethoven JJM, Le Texier V, Muilu J: ASD: the Alternative Splicing Database. Nucl Acids Res 2004, 32: D64-D69.PubMed CentralView ArticlePubMedGoogle Scholar
- Boser BE, Guyon IM, Vapnik VN: A training algorithm for optimal margin classifiers. 5th Annual ACM Workshop COLT 1992, 144–152.Google Scholar
- Cortes C, Vapnik V: Support vector networks. Machine Learning 1995, 20: 273–297.Google Scholar
- Schölkopf B, Smola AJ: Learning with Kernels. MIT Press, Cambridge, MA; 2002.Google Scholar
- Müller KR, Mika S, Rätsch G, Tsuda K, Schölkopf B: An Introduction to Kernel-based Learning Algorithms. IEEE Neural Networks 2001, 12(2):181–201.View ArticlePubMedGoogle Scholar
- Byvatov E, Schneider G: Support vector machine applications in bioinformatics. Appl Bioinformatics 2003, 2(2):67–77.PubMedGoogle Scholar
- Zhang XH, Heller KA, Hefter I, Leslie CS, Chasin LA: Sequence information for the splicing of human pre-mRNA identified by support vector machine classification. Genome Res 2003, 13(12):2637–2650.PubMed CentralView ArticlePubMedGoogle Scholar
- Leslie CS, Eskin E, Cohen A, Weston J, Noble WS: Mismatch string kernels for discriminative protein classification. Bioinformatics 2004, 20(4):467–476.View ArticlePubMedGoogle Scholar
- Dror G, Sorek R, Shamir R: Accurate identification of alternatively spliced exons using support vector machine. Bioinformatics 2005, 21(7):897–901.View ArticlePubMedGoogle Scholar
- Joachims T: Making large-Scale SVM Learning Practical. In Advances in Kernel Methods – Support Vector Learning. Edited by: Schölkopf B, Burges C, Smola A. MIT-Press; 1999.Google Scholar
- Fuellen G, Spitzer M, Cullen P, Lorkowski S: BLASTing proteomes, yielding phylogenies. In Silico Biol 2003, 3(3):313–319.PubMedGoogle Scholar
- Fuellen G, Spitzer M, Cullen P, Lorkowski S: Correspondence of function and phylogeny of ABC proteins based on an automated analysis of 20 model protein data sets. Proteins 2005, 61(4):888–899.View ArticlePubMedGoogle Scholar
- Moody J, Darken CJ: Fast learning in networks of locally-tuned processing units. Neural Computation 1989, 1(2):281–294.View ArticleGoogle Scholar
- Wheeler DL, Barrett T, Benson DA, Bryant SH, Canese K, Church DM, DiCuccio M, Edgar R, Federhen S, Helmberg W, Kenton DL, Khovayko O, Lipman DJ, Madden TL, Maglott DR, Ostell J, Pontius JU, Pruitt KD, Schuler GD, Schriml LM, Sequeira E, Sherry ST, Sirotkin K, Starchenko G, Suzek TO, Tatusov R, Tatusova TA, Wagner L, Yaschenko E: Database resources of the National Center for Biotechnology Information. Nucl Acids Res 2005, 33: D39-D45.PubMed CentralView ArticlePubMedGoogle Scholar
- Sczyrba A, Beckstette M, Brivanlou AH, Giegerich R, Altmann CR: XenDB: full length cDNA prediction and cross species mapping in Xenopus laevis. BMC Genomics 2005, 6: 123.PubMed CentralView ArticlePubMedGoogle Scholar
- Abouelhoda MI, Kurtz S, Ohlebusch E: Replacing Suffix Trees with Enhanced Suffix Arrays. Journal of Discrete Algorithms 2004, 2: 53–86.View ArticleGoogle Scholar
- Huang X, Madan A: CAP3: A DNA sequence assembly program. Genome Res 1999, 9(9):868–877.PubMed CentralView ArticlePubMedGoogle Scholar
- Altschul SF, Madden TL, Schäffer AA, Zhang J, Zhang Z, Miller W, Lipman DJ: Gapped BLAST and PSI-BLAST: a new generation of protein database search programs. Nucl Acids Res 1997, 25: 3389–3402.PubMed CentralView ArticlePubMedGoogle Scholar
- Dean M, Rzhetsky A, Allikmets R: The human ATP-binding cassette (ABC) transporter superfamily. Genome Res 2001, 11(7):1156–1166.View ArticlePubMedGoogle Scholar
- Hsu CW, Chang CC, Lin CJ: A practical guide to support vector classification.[http://www.csie.ntu.edu.tw/~cjlin/]
- Sarle WS: Neural Network FAQ. Periodic posting to the Usenet newsgroup comp.ai.neural-nets 1997.Google Scholar
- Katoh K, Misawa K, Kuma K, Miyata T: MAFFT: a novel method for rapid multiple sequence alignment based on fast Fourier transform. Nucl Acids Res 2002, 30: 3059–3066.PubMed CentralView ArticlePubMedGoogle Scholar
- Jones DT, Taylor WR, Thornton JM: The rapid generation of mutation data matrices from protein sequences. Comput Appl Biosci 1992, 8(3):275–282.PubMedGoogle Scholar
- Fuellen G: A Gentle Guide to Multiple Alignment. Complexity International 1997., 4: [http://journal-ci.csse.monash.edu.au/ci/vol04/mulali/]Google Scholar
- Brown NP, Leroy C, Sander C: MView: a web-compatible database search or multiple alignment viewer. Bioinformatics 1998, 14(4):380–381.View ArticlePubMedGoogle Scholar
- Qian J, Lin J, Luscombe NM, Yu H, Gerstein M: Prediction of regulatory networks: genome-wide identification of transcription factor targets from gene expression data. Bioinformatics 2003, 19(15):1917–1926.View ArticlePubMedGoogle Scholar
- Rätsch G, Onoda T, Müller K: Soft Margins for AdaBoost. Mach Learn 2001, 42(3):287–320.View ArticleGoogle Scholar
- Efron B, Gong G: A leisurely look at the bootstrap, the jackknife, and cross-validation. The American Statistician 1983, 37: 36–48.Google Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.