IsoSVM--distinguishing isoforms and paralogs on the protein level

Abstract

Background: 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.

Results: 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.

Conclusion: 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.

Figure 1

Visualization of a part of an alignment of (A) two paralogous sequences (the human ABCB4 and ABCB1 protein) and (B) two isoforms (the human ABCB4 protein and its isoform c), representing an ideal case. Positions with matches between the two sequences are indicated by "|", mismatches by "#" and amino acids vs. gap characters by ":". The values of the three features (cf. Methods, section Features) for the full-length sequences compared in panel (A) are (i) sequence similarity 75.76%, (ii) inverse CBIN count 0.0027, (iii) fraction of consecutive matches and mismatches 0.7111. For the full-length sequences compared in panel (B) we have (i) sequence similarity 96.33%, (ii) inverse CBIN count 0.3333, (iii) fraction of consecutive matches and mismatches 0.9969.

Figure 2

Features displayed by the samples in the canonical training dataset. Panels (A) to (C) illustrate combinations of two of the three features. Panel (D) illustrates all three features at the same time. Samples arising from the comparison of paralogous sequences are shown in blue, whereas isoforms are shown in red. An inverse CBIN count of 1/n arises if n CBINs are featured by a given sample. Though the samples of both classes separate well in general, some samples of one class "overlap" into the other class.

Figure 3

Illustration of the different cases of consecutive blocks of identities or non-identities (CBINs). (A) CBIN of matches, (B) CBIN of gaps (counted as mismatches), (C) CBIN of mismatches, (D) example of a comparison of two sequences with an alignment length of 32. Matches are denoted by "|", mismatches by "#" and amino acids aligned to gaps by ":". The example alignment of length 32 features eight CBINs. The values of the three features are: (i) sequence similarity 0.594, (ii) inverse CBIN count 0.125, (iii) fraction of consecutive matches and mismatches 0.75.

Figure 4

Accuracy of classifiers measured by jackknife resampling, employing all three features. Performance of the SVM classifier is compared to classifiers based on an RBF network as well as a linear classifier. Mean accuracy and standard error of the mean were assessed by 100-fold jackknife resampling using 7604 samples resulting from a visual inspection process of protein sequences taken from Genbank.

Figure 5

Visual inspection process. Matches in the alignments are colored in blue and mismatches in red. Amino acids aligned to gaps are indicated in green. Panels (A) to (D) illustrate alignments of two protein sequences classified as isoforms (panels (A) and (B)) or as paralogs (panels (C) and (D)). The sequences shown in panel (A) feature a shared subsequence (a putative constitutive exon), marked in blue. The upper sequence features an additional exon at the beginning (marked in green) that is missing in the lower sequence. In contrast, a putative exon at the end (also shown in green) is found in the lower sequence only. Comparison of the two putative isoforms shown in panel (B) reveals two constitutive exons in the middle and towards the end of the alignment, colored in blue (the only mismatch is interpreted as a sequencing error, or a polymorphism). These are separated by a stretch of amino acids aligned to gaps, interpreted as an exon skipped in the lower sequence. At the beginning of the alignment, the upper sequence features a long stretch of amino acids aligned to gaps and a few mismatches; two mutually exclusive exons are a plausible interpretation, since the lower sequence (starting with G and not with M) is incomplete and its first exon is probably much longer. At the end of the alignment both sequences feature a stretch of mismatches and gaps (colored in red), interpreted as mutually exclusive exons (indicated by a black frame). The sequences compared in panel (C) give rise to a sample of the paralog class. In general, the alignment features many mismatches, interpreted as substitutions, and six stretches of amino acids aligned to gaps (putative deletions). Panel (D) illustrates another putative paralog. Besides a shared stretch (featuring numerous substitutions) in the middle of the alignment, the upper sequence features putative deletions, or missing exons. It may thus be a case of an isoform of a paralog.

Figure 6

SVM training process. The complete dataset generated by visual inspection was split into two parts, yielding a canonical training dataset of 3,802 samples and a canonical testing dataset of 3,802 samples, each consisting of an equal number of isoform and paralog instances. The canonical training dataset was again split into four subsets (denoted by numbers in circles) and submitted to the grid-search procedure. The resulting classifier was then tested on the canonical testing dataset.