In the first step of analysis we searched for peptides that unambiguously indicated the presence of two or more splice isoforms from the same gene from the more than 42,000 peptides identified in the two proteomics studies. Peptides were matched to the proteins in FlyBase (release 5.4). Since FlyBase was used to identify the peptides from both studies, the peptides could be mapped back directly to proteins in FlyBase using a simple Perl script. Each of the more than 42,000 peptides mapped to at least one gene in FlyBase.

FlyBase release 5.4 contains 15,181 genes, of which 14,141 are predicted to be protein coding. The release contains a total of 20,823 protein coding transcripts and 17,961 of the polypeptide gene products are unique. The 2,762 transcripts that are alternatively spliced in the 3' or 5' untranslatable regions were not considered in this study, since they produce identical translated products and cannot be distinguished with peptide data alone. A total of 2,406 protein-coding genes (17.01%) code for more than one gene product.

It is important to note that the peptides in the Brunner and Bodenmiller experiments were identified using FlyBase. This means that neither of the studies could characterize peptides that did not match FlyBase sequences. As a result, our analysis of splice isoforms had to be limited to the alternative splice isoforms annotated in FlyBase. Therefore, the highest possible detectable alternative splicing rate from the two experiments was 17.01%.

After matching the Brunner and Bodenmiller peptides to the unique proteins in FlyBase, we searched for genes that had two or more alternative isoforms confirmed by the peptide evidence. Genes with confirmed alternative protein isoforms were those for which it was possible to map peptides to regions unique to two or more alternative isoforms. In other words, where the detected peptides could unequivocally demonstrate the presence of two gene products with distinct protein sequences (for example, Figure 1).

The peptide evidence from the Brunner set confirmed multiple alternative isoforms for 76 genes and the evidence from the Bodenmiller set confirmed multiple alternative isoforms for 60 genes. There was a certain amount of overlap between the two experiments - 19 genes had multiple isoforms confirmed by the peptides from both analyses. In addition, when the two sets of peptides were combined there was evidence for alternative gene products from another 13 genes. In total, we were able to demonstrate that 130 separate Drosophila genes expressed at least two alternative isoforms (Additional data file 1). While this is only a small proportion of the genes that are supposed to express alternative protein isoforms, the figure is considerably higher than any previous study we know.

Those genes for which it was possible to show the presence of three or more distinct gene products were particularly interesting. Five genes - SNF4A-gamma, Akap200, 14-3-3-epsilon, mod(mdg4), and LOLA - each expressed at least four distinct gene products. By combining the peptide data it was possible to show that one gene, LOLA, expressed at least seven different isoforms (Figure 1). The 26 splice isoforms of LOLA annotated in Flybase have very different carboxyl termini, but two-thirds of them include two zinc-finger domains. All seven confirmed LOLA splice isoforms have both carboxy-terminal zinc-finger domains (Figure 2). Of the five genes shown to have more than two distinct gene products, there are individual studies for LOLA 24, SNF4A-gamma 25 and mod(mdg4) 26 that predict the presence of multiple splice isoforms.

There was no evidence of the expression of alternative splice isoforms of the Dscam protein that was mentioned in the introduction, but the two studies did show the expression of different isoforms of Sex lethal (for the tandem mass spectra of the phosphopeptides for the two isoforms see Additional data files 2-4). Other genes that are predicted to have differently functioning splice isoforms include CTBP, thought to be important in development and to have two variants that are conserved across all insect species 27, Eif2 and Su(var)3-9, which fuse in Drosophila and are expected to create two different gene products 28, and Polychaetoid 29 and Thioredoxin reductase 30, both of which are predicted to have gene products whose cellular locations are controlled by alternative splicing.

While the two studies present irrefutable evidence that alternatively spliced transcripts are expressed as proteins, the total number of genes with confirmed alternative products from the two experiments is small. We could confirm the expression of alternative protein isoforms for just 1.1% of the 6,980 unambiguously identified genes in the Brunner set, while the Bodenmiller study showed that 1.8% of the 3,472 identified genes express more than one protein isoform.

However, this apparently low level of alternative splicing at the protein level has to be seen in the context of the relatively low coverage of the Drosophila proteome by the two experiments. Although the peptides detected by the Brunner study did confirm the presence of gene products for more than half of the Drosophila genes, the identified peptides covered just 5% of the amino acid residues in the Drosophila proteome. This obviously decreases the chances of finding peptides that unambiguously correspond to alternatively spliced regions.

In order to demonstrate whether the low rates of detection of alternative isoforms are significant, we carried out simulated in silico peptide identification experiments. These in silico experiments determine the expected rates of detection assuming that all peptides are equally detectable, even though some proteins may be more abundant or more easily detectable than others. The simulations cannot tell us what the real rate of alternative splicing at the protein level is, since the maximum detectable rate is limited by the rate of alternative splicing found in FlyBase (in this particular case 17%; see above). However, they do provide an estimation of the number of alternative isoforms that we would have expected the two experiments to detect.

For the comparison with the Brunner analysis we drew 37,279 peptides at random from an in silico trypsin digest of the D. melangaster proteome based on Flybase release 5.4. The simulation was performed 1,000 times. The results showed that a random selection of 37,279 peptides from the in silico digest would be expected to confirm the expression of alternatively spliced isoforms for a mean of 242.75 genes (standard deviation of 9.23). By way of contrast, the peptides identified in the Brunner analysis confirmed multiple alternatively spliced isoforms for just 76 genes.

For the Bodenmiller study we drew 10,118 peptides at random from the in silico trypsin digest and again the simulation was performed 1,000 times. From the random drawing we were able to show that 10,118 peptides would be expected to confirm expression of distinct splice isoforms for a mean of 56.24 genes (standard deviation of 5.05). The Bodenmiller analysis confirmed multiple alternatively spliced gene products for 60 genes.

The simulations allowed us to show that the Brunner analysis detects a little under a third of the number of genes that would be expected to produce alternative isoforms. There are several possible explanations for this. One reason may be that many isoforms are only expressed in certain tissues or certain stages of development. Given the wide range of cell types and developmental states that were used in this experiment, this explanation seems less likely. A second possibility is that the transcripts predicted from cDNA and EST evidence are simply not all transcribed and the transcript evidence is an overestimation of the real number of proteins expressed in the cell. Another explanation might be that many alternative isoforms may only be expressed in very low quantities and are less easily detected.

By way of contrast with the results from the Brunner simulations, the 60 genes with confirmed multiple alternatively spliced gene products in the Bodenmiller analysis is almost exactly what would be expected from the Bodenmiller simulations. It is somewhat surprising that it was the Bodenmiller analysis, and not the Brunner analysis, that detected the rate of alternative splicing expected from the random simulations. The prevailing ideology of alternative splicing assumes that alternative isoforms are expressed in distinct tissues and developmental stages; therefore, we would expect to confirm a higher rate of alternative splicing with the Brunner analysis, where many cell types and developmental stages were interrogated, than in the Bodenmiller analysis, where only a single cell type was tested.

The difference in the frequency of alternative splicing detected by the two studies is enlightening. The Bodenmiller analysis identified proportionally more alternative isoforms than the Brunner analysis (1.8% of genes detected in the Bodenmiller study had alternative protein isoforms, compared to 1.1% detected in the Brunner experiments) and this is even more clear from the simulations. The contrasting results from the two analyses strongly suggest the possibility that methods such as those used in the Bodenmiller analysis are more sensitive when it comes to detecting certain alternative isoforms.

While some peptides were identified by both studies, on the whole the peptides recognized by the two analyses are different. This is clear from the residue composition of the peptides detected by the two analyses. Although the residue composition of the peptides found in the Brunner analysis is similar to the residue composition of the Drosophila proteome - except that there are fewer basic residues and more acidic residues (Figure 3b) - the peptides detected in the Bodenmiller analysis have a quite distinctive composition (Figure 3a). The Bodenmiller analysis specifically selected peptides that were phosphorylated. The peptides detected in these experiments have substantially more serine residues (unsurprising because almost 90% of the phosphorylated residues are serines), but also many more proline (5.5-7.6%), asparagine (4.7-5.7%), glycine (6.2-7.2%) and aspartate residues (5.2-5.9%). All these values are significantly higher than expected according to Chi-squared tests (p-values < 0.005).

In addition, all hydrophobic residues are markedly under-represented: the values for cysteine (1.83% of the residues in the Drosophila proteome and just 0.63% of the Bodenmiller peptides), phenylalanine (3.47 and 2.18%), isoleucine (4.9 and 3.71%), leucine (8.94 and 6.69%), methionine (2.33 and 1.14%), valine (5.9 and 5.34%), tryptophan (0.98 and 0.35%) and tyrosine (2.92 and 1.57%) are significantly less than expected (all Chi-squared p-values < 0.005). The peptides from the Bodenmiller set are considerably less hydrophobic than normal - just one in five of the Bodenmiller residues are hydrophobic, compared to the one in three residues in the Drosophila proteome (Figure 3c).

This residue composition of the Bodenmiller peptides is typical of regions that are disordered in solution. It is well known that proteins with few hydrophobic residues and more polar residues are likely to correspond to disordered regions of the structure 3132. Studies also suggest that phosphorylated residues tend to be more frequent in flexible, unstructured segments and linkers 3334. Taken together, this information strongly suggests that many of the Bodenmiller peptides, as well as being in exposed regions on the surface of proteins, will be disordered when in solution. Indeed, where the Bodenmiller peptides can be mapped to known structures, most map to regions on the surface or regions known to be disordered in solution (Figure 4).

It is widely recognized that a certain proportion of peptides identified in proteomics techniques can be false positives. However, both the Brunner and Bodenmiller studies have low rates of false positives. The Brunner analysis has a false positive rate of approximately 5% and the Bodenmiller analysis has a false positive rate of 1-4%. Even if 10% of these peptides were to be false positives (twice the determined value), there would still be considerably more than 100 genes with evidence of alternative splicing at the protein level from the two studies. In any case, a small number of false positives will not affect the main conclusions of this study. Most alternative transcripts do seem to produce alternative gene products and many of these alternative isoforms may have regions that are disordered in solution.

Given the depth and range of peptides detected in the two studies, we might also have expected to be able to uncover the expression of peptides not in FlyBase, such as those from predicted genes and transcripts (isoforms), translated pseudogenes and small RNAs, or in principal any other peptide produced by the 6-frame translation of the fly genome. A complete re-analysis of the spectra from the Bodenmiller study was beyond the scope of this paper, but we were able to carry out an initial re-analysis against a locally generated database that contained 903,842 peptides from translated transcripts from predicted gene models, translated pseudogenes and translated miscellaneous functional RNA. The re-analysis identified seven peptides that mapped exclusively to predicted gene models and two peptides that were linked to the miscellaneous RNA. There was no evidence for the expression of any of the pseudogenes as peptides.

Three of the predicted gene models (genscan_masked:gene254366, genscan_masked:gene247065, and genscan_masked:gene245985) were not similar to any sequences in the UniProt database. One ab initio predicted gene model (genscan_masked:gene264127) did match a unique sequence in UniProt, but only because the prediction itself had been erroneously included in the UniProt sequence database.

One peptide mapped to four different predictions (genscan_masked:gene266459, genie_masked:gene1703010, genie_masked:gene1402427 and genscan_masked:gene1391762) that were 40% identical to a putative gag-pol protein (Drosophila ananassae). The remaining two predicted gene models identified by the spectra might be alternative variants of vav (both genie_masked:gene1736185 and genscan_masked:gene267148) and lethal (2) 05510 (genscan_masked:gene263593) but have yet to be annotated as variants in FlyBase.

Of the two miscellaneous transcripts identified by the re-analysis of the spectra, one is a piece of rRNA (FBtr0114214) that is not similar to anything in the UniProt database and the other maps to a piece of small nucleolar RNA (snoRNA; Or-aca1, FBtr0113530) that, when translated, is 71% identical (but over just 20 residues) to the hypothetical protein SNOG_09564 (Phaeosphaeria nodorum SN15). Or-aca1 is located inside an intron of ribosomal protein S16.

We have been able to demonstrate conclusive evidence for the genome-wide expression of alternative splice variants at the protein level and have shown that distinct proteins are indeed produced from alternative splice variants. The results from the two large-scale proteomics studies on which our analysis is based showed that the expression of alternative gene products is extensive. These studies confirmed the presence of multiple alternative isoforms for over a hundred genes. Moreover, the alternative isoforms detected in these two studies were sufficiently stable in vivo and produced in sufficient quantities to be detectable in proteomics workflows. Even though the current technical limitations of proteomics studies allowed the recovery of just a small fraction of the potential alternative isoforms (less than 2% of Drosophila genes were identified with alternative protein isoforms), the results were enough to estimate the presence of alternative splicing in the genome and to propose that most, if not all, recorded alternative variants are likely to be expressed at the protein level in some form.

The comparison of the two proteomics studies showed that the level of expression of alternatively spliced variants in the general proteomics analysis of Brunner was less than would be expected, but that the expression levels of alternative gene products in the Bodenmiller experiment, which specifically targeted and identified phosphopeptides, corresponded to the levels of expression predicted by FlyBase. The higher proportion of alternatively spliced gene products detected in the Bodenmiller analysis is most probably related to the sensitivity of the analysis of charged phosphopeptides. One of the effects of the sensitivity to phosphorylation sites is that the identified peptides have a significant compositional bias. The peptides have many fewer hydrophobic residues and markedly more polar residues, suggesting that many of these phosphorylation sites are in regions that are disordered in corresponding protein structures.

This observation is interesting since it may have consequences for our understanding of the structural and functional consequences of splicing. The detailed analysis of the potential effects of alternative splicing in proteins shows that alternative splicing would be expected to lead to substantial rearrangements in the corresponding structures 11 and it is unlikely that the large changes introduced by alternative splicing events will generate regions that fold with a stable hydrophobic core. It has previously been suggested that a substantial proportion of alternative gene products are unstructured in solution 35. A corollary to this is that there are only eight known pairs of alternative splice isoforms in the Protein Data Bank (PDB) structural database 36. In five of these pairs the regions resulting from alternative splicing events are disordered. It may be that many alternative splicing events result in proteins that are, at least in part, unstructured and flexible in solution. If alternative splicing events are related to disordered regions and phosphorylated residues are more frequent in these unstructured and flexible regions, then it follows that the disordered regions resulting from alternative splicing events will be more easily detected by methods that detect phosphorylated peptides. Therefore, it is not surprising that the Bodenmiller analysis was able to detect a higher proportion of splice isoforms.

The results of this analysis have shown that proteomics data can indeed be used to investigate the extent of alternative splicing at the protein level. The Bodenmiller analysis detected peptides that differed from those in the Brunner analysis because they specifically isolated phosphorylated peptides from whole cell lysate, suggesting a methodology for carrying out further experiments to detect alternative splicing at the protein level.

Perhaps surprisingly though, our initial re-analysis of the liquid chromatography tandem mass spectrometry data using databases of predicted transcripts, translated pseudogenes and small RNAs failed to reveal any significant new findings. Unfortunately, while we were able to detect some evidence for the expression of small RNAs and predicted gene models, the number of novel identified peptides fell within the estimated false positive rate. A complete re-analysis of the spectra might have produced more interesting results. However, our initial re-analysis made it clear that any proteomics study of the expression of functional aspects of the genome would be complicated by a series of logistical challenges. For example, six-frame translations of genome transcripts would create an enormous search space that would result in extensive and impracticable database search times. In addition, peptide detection sensitivity correlates with database size (especially if most added sequences are likely not to be real peptides) and is strongly reduced in the case of a 6-frame translated database.

Nevertheless, what is and is not expressed as protein is an interesting scientific question that needs to be addressed and our study points the way to further experiments of this nature. Future proteomics studies could address the question by searching against additional databases organized for the purpose, but specific methods to deal with the false positive problem still need to be developed and, ideally, the detected peptides would need to be confirmed using independent methods.

The fact that we can show that alternative transcripts are translated into proteins at the predicted rate is a great step forward and shows the importance of proteomics in validating predicted transcripts. Of course, showing that the alternative splice isoforms are indeed expressed as stable proteins is only the first step in assessing the functional role of alternative variants. In order to broaden our understanding of the role of genomic protein diversity, further experimental approaches are needed. We feel that these results will serve as an important point of reference for these experiments.

We searched 154,509 spectra from the Bodenmiller dataset against a database that contained 903,842 peptides derived from 17,868 translated transcripts: 10, 000 translated transcripts came from predicted gene models from FlyBase, 1,456 came from the 6-frame translation of pseudogenes from FlyBase, 3,818 from 6-frame translations of miscellaneous functional RNA (rRNA, small nuclear RNA (snRNA) and snoRNA) from FlyBase and 2,594 were generated from the 6-frame translation of transcripts predicted by Manak et al. 40 to be functional.