Comparative Functional Analysis of the Caenorhabditis elegans and Drosophila melanogaster Proteomes

1The rapid lifecycle, small size, reproducible development, and ease of cultivation in the laboratory have made Caenorhabditis elegans an important experimental system for biological studies. Numerous human disease-related genes (e.g., related to cancer or neurological diseases) have orthologs in the worm. Sequencing and annotation of its genome has revealed more than 19,000 genes coding for more than 22,000 proteins, including splice variants. Extensive systematic studies of gene function have been performed. However, to completely understand complex biological processes such as development, aging, or disease, the analysis of the proteome—i.e., the entire set of the expressed proteins—is becoming increasingly important. Knowledge of the complete sequence of a genome is a necessary prerequisite for proteomics, but the DNA sequence itself does not reveal which proteins are actually expressed when, where, and to what level. Furthermore, in contrast to the genome, the proteome is changing under different biological conditions. Although for many years, transcriptome data (i.e., the collection of transcribed mRNAs) has been used to approximate the proteome, a number of studies have demonstrated that the correlation between mRNA and protein abundance is surprisingly low because of posttranscriptional regulation and variable protein half-lives. The analysis of the proteome is therefore a key method to provide systems-level information about protein function in time and space, and to obtain a concise view of biological processes. In the case of C. elegans, previous analyses of the proteome were either limited in scope and coverage, or largely focused on improving genome annotation, with the biggest C. elegans proteome dataset published so far encompassing 6,779 proteins.

2To generate a comprehensive, deeply sampled C. elegans proteome database that can be used for quantitative proteome analysis, we applied subcellular and biochemical fractionation methods to the worm proteins, performed tryptic digests, separated the resulting peptides using a variety of techniques, and identified the peptides by mass spectrometry (MS). This resulted in a unique global view on the expression status of the C. elegans proteome. We identified a number of protein features and functions that are underrepresented in the expressed proteome, likely representing specialized functional systems expressed only in a small subset of cells and/or developmental stages. We demonstrate the importance of proteomics data towards improved genome annotation. Finally, we compared the proteome data with similar data from the fruit fly Drosophila melanogaster. The latter comparison provided—for the first time to our knowledge—an overview of the expressed “core animal proteome,” which should arguably become the initial focus for monitoring the basic metazoan cellular machinery in the future.

3To identify C. elegans proteins, we collected worms at various developmental stages and homogenized whole animals and eggs to isolate the proteins. Their tryptic peptides were separated using strong cation exchange chromatography (SCX), in several cases after labeling them with isotope-coded affinity tags (ICAT) to reduce sample complexity, or by isoelectric focusing (applying free-flow electrophoresis and immobilized pH gradient strips). The peptides were finally identified using microcapillary liquid chromatography–electrospray ionization–tandem MS (μLC-ESI-MS/MS). With this extensive shotgun proteomics approach, we identified 10,977 different proteins, including splice variants, via 84,962 nonredundant peptide identifications (Table S1; 759,320 peptide identifications were obtained in total). We identified 10,631 gene loci, corresponding to 54% of the gene loci in WormBase (WS140: 19,735 loci). Of these, 7,476 loci (38%) were detected via several distinct peptides, 580 (3%) were detected via the same peptide more than once, and 2,575 (13%) were detected only via a single peptide identification (Figure 1). When considering individual annotated exons (irrespective of their various splicing contexts), our peptide data covered 28.2% of the 129,047 exons contained in WormBase.

4Protein identification from MS peptide spectra is prone to false-positive assignments, and we employed strict search cutoffs using PeptideProphet (see Materials and Methods). To independently estimate our false discovery rate (FDR), in particular for identifications based on a single peptide spectrum (“single hits”), we first took advantage of one of our experiments that used isoelectric focusing to fractionate peptides. In each peptide fraction, true-positive identifications should scatter around a narrow range of isoelectric points (pIs), whereas false-positive identifications should follow the background distribution in the database. This analysis, using computational predictions of pIs to check all peptides, yielded an estimated FDR of 35% for single hits in this particular experiment. Independently, a newly developed model based on a robust decoy search strategy yielded an upper limit for the FDR of single-hit identifications at around 63% for all combined experiments (L. Reiter, M. Claassen, S. P. Schrimpf, J. M. Buhmann, M. O. Hengartner, et al., unpublished data). By the latter method, multi-hit identifications were found to be much more reliable, resulting in an FDR of 7% in our study. Since almost half of all single-hit identifications do represent bona fide protein identifications, we chose to include single-hit identifications in our subsequent analyses. A separate analysis focusing on just these proteins alone showed that they often belonged to groups that were underrepresented in the complete dataset and are therefore presumably of low abundance in C. elegans (short, uncharacterized proteins and in particular those with seven transmembrane domains; also see below). This means that they do represent valuable information about which proteins are expressed at low levels in C. elegans. It should also be stressed that all conclusions reported below remained valid when single-hit identifications were excluded.

5To assess whether proteins from sources other than C. elegans were present in our preparations, we focused on the bacteria on which the worms were feeding (Escherichia coli). We tested a single, representative experiment, encompassing 67 MS/MS analyses by searching the spectra against a combined C. elegans and E. coli database. A total of 1.3% of the protein identifications mapped to E. coli, among them 14 hits mapping to both organisms. However, for each of these 14 proteins, there was at least one additional C. elegans peptide identified, confirming that these overlapping detections did not influence the C. elegans results.

6In order to characterize C. elegans proteins that were not detected, and that are therefore most likely expressed at particularly low levels, or in specialized cells or developmental stages only, we classified the entire predicted C. elegans proteome with respect to several aspects (length, pI, hydrophobicity, transmembrane topology, and functional annotation). This should reveal the nature of underrepresented proteins (with potentially more peripheral, or even worm-specific functions), and separate them from abundant proteins involved in basic cellular processes such as growth, metabolism, and information processing. It should also reveal potential technical limitations (proteins/peptides difficult to detect using our procedure), which is important to assess for future systematic uses of MS.

7Our bias analyses revealed an underrepresentation of proteins shorter than 400 amino acids (Figure 2A) and of proteins with basic pIs (Figure 2B). A similar bias has previously been observed for D. melanogaster. The underrepresentation of basic proteins was partly to be expected, due to our isoelectric focusing experiments, which centered on the pH range 3–7. The underrepresentation of short (low molecular weight) proteins might be caused by a generally higher prevalence of spurious gene predictions among short genes, and also by a lower probability of detecting one of the few tryptic peptides generated by short proteins. We observed a bimodal distribution of hydrophobicity values within the annotated set of all C. elegans proteins, and a strong underrepresentation of proteins in the second, high hydrophobicity peak in our dataset (Figure 2C). This second peak consists mostly of multipass transmembrane proteins (∼64% of these proteins have seven or more predicted transmembrane domains). To better understand how membrane association relates to protein abundance and detectability, we globally characterized WormBase proteins with respect to their content of transmembrane segments, using Phobius. Overall, we found a notable underrepresentation of transmembrane proteins in our proteomics data, and decided to subdivide these proteins further according to the number of transmembrane sections and annotated functions as shown for other species (Figure 2D and 2E). Remarkably, we found that the strongest underrepresentation is observed for proteins with seven transmembrane regions, in particular those annotated with the function “receptor activity.” This may point to a biological (rather than technical) explanation for the relative paucity of transmembrane proteins in our data: Seven-transmembrane chemosensory receptors are widespread in the C. elegans genome, but many of these are known to be expressed only in a small number of neurons each. Because we assessed whole animals, those proteins might be too rare to be successfully detected. This general underrepresentation in our proteome data suggests similar sensory functions for other transmembrane proteins of hitherto unknown function that we also found to be of too low abundance to be detected.

8Finally, we globally analyzed the functional classifications of all the detected proteins. We observed a clear bias towards proteins with known functions. The same bias was also observed for the D. melanogaster proteome. A possible explanation could be that some of the undetected proteins with unknown functions are actually erroneous gene predictions or pseudogenes. It could also be a testament to the biases of previous studies: abundant proteins are easier to work with biochemically, and may therefore have obtained a functional annotation more easily. In total, our proteomics approach identified proteins belonging to 125 out of the 127 Gene Ontology (GO) slim categories defined for WormBase. The global GO slim analyses confirmed the underrepresentation of proteins with receptor activity mentioned above, and of “membrane” or “integral to membrane” proteins in general (Figure 2F).

9Large-scale proteome analyses (such as ours) represent an important cornerstone for an improved genome annotation. In WormBase (WS160), 4,987 gene loci were still listed with the gene status “predicted” only, i.e., without any supporting transcript data (expressed sequence tag [EST], mRNA). We experimentally confirmed the protein expression of 1,062 of these predicted genes (among them, more than 40% via multiple peptide detections). As was the case for the whole proteome, this subset was enriched for proteins with GO slim annotations (45% in our dataset, as compared to 38% expected for this subset in WormBase; p-value: 4.65E−08). Apart from these gene confirmations, our C. elegans proteomics dataset contains numerous spectra originating from nonannotated regions in the worm genome. In computationally intensive analyses, we are identifying these by searching our data against six-frame translations of the genome, and filtering the results for high confidence spectra that map to nonannotated regions. For example, from one particular experiment, we identified 78 likely novel peptides. Two of these are illustrated in Figure 3 (the corresponding MS/MS spectra are provided as Figure S1). These data suggest an alternative translational start site for the protein SYN-4 (T01B11.3; Figure 3A): the observed peptide is located upstream of the annotated translational start site, and only partially overlaps with the currently annotated protein sequence. The second example demonstrates a novel splice variant for the gene F47B7.7 (Figure 3B). In this case, we identified a peptide that extends an existing annotated exon downstream, in the correct frame. These and similar analyses, suggesting altered or new gene models, are computationally very intensive and were not yet completed at the time of submission. Furthermore, due to the increased search space when searching proteomics against the genome, extra scrutiny is needed when interpreting each reannotation instance, and additional experimental data should probably be taken into account before fully accepting these gene annotation changes.

10C. elegans and its relatives are unique among characterized metazoans in that a large number of their genes are organized into operons (multicistronic transcription units, containing up to eight genes that are strictly coexpressed). Following transcription, the primary transcript is split up through a unique trans-splicing mechanism, and the individual open reading frames are subsequently translated separately into distinct mature proteins. In order to assess the potential influence of operon structure on the regulation and abundance of proteins, we studied the expression status of genes in operons, compared to individually transcribed genes. Although an absolute quantification of protein levels is not possible with our shotgun approach, we performed a semiquantitative analysis based on spectral counting. Surprisingly, we observed that proteins encoded by operons are expressed far more strongly than those encoded by individually transcribed genes: we observed 84% of the former, with a median relative abundance of 20 ppm (parts per million of total protein molecules), but only 47% of the latter with a median relative abundance of 5 ppm (Figure 4A). The same tendency was found when analyzing publicly available transcript-abundance data (Figure 4B). This striking observation confirms that operons are preferentially made up of genes that are strongly transcribed, and we now establish that this is reflected also at the protein level: operon proteins, on average, are more than 3-fold more abundant than proteins from single-gene transcripts. Apart from grouping strongly expressed proteins, operons are also expected to facilitate coordinated regulation of their constituent genes. We assessed whether this is the case by searching for operons that were either fully expressed (i.e., all encoded proteins detectable) or silenced (none or very few of the encoded proteins detectable). Indeed, we found significantly more operons of both types than expected by chance, as illustrated for operons of lengths 4 to 6 (Figure S2). In principle, our observations could be stemming from a limited selection of tissues only, for example from the hermaphrodite germline, where operons are thought to be strongly expressed during oogenesis. However, we observed that operon proteins are more abundant even in dauer and L1-stage larvae, which both should have very little germline material (Figure S3). We further checked whether our observation could be explained by systematic differences in length or transmembrane segments of operon proteins. Although we did observe slight differences in length and transmembrane content—operon proteins are on average 11% longer, and transmembrane proteins are 40% less frequent—these differences were not sufficient to explain the increased abundance of operon proteins (unpublished data). Together, our observations indicate, for the first time, that operons in C. elegans ensure the coordinated regulation of highly expressed proteins.

11In this study, we had the unique opportunity to compare large-scale proteome datasets from two different animal species, owing to the recent publication of the D. melanogaster proteome (http://www.mop.uzh.ch/peptideatlas/; previous work in D. melanogaster had mainly focused on protein–protein interactions or subproteomes only). We performed spectral counting for both organisms to obtain semiquantitative measurements of protein abundance, and compared these to published mRNA expression data derived from Affymetrix and serial analysis of gene expression (SAGE) platforms. In the C. elegans and D. melanogaster proteomes, 2,695 pairs of orthologs were identified for which all three types of data were available. Surprisingly, we observed that orthologs showed a strong correlation in protein abundances across the two organisms, despite more than 600 million years of separate evolution (Spearman rank correlation RS = 0.79; Figure 5A). Notably, this biological correlation at the protein level between the two species is even higher than the within-species correlation between protein and transcript abundances (within C. elegans: RS = 0.59 and 0.44 for protein-Affymetrix and protein-SAGE, respectively; within D. melanogaster: RS = 0.66 and 0.36, respectively). In contrast to the protein-level correlation, the abundance correlations at the transcript level between the two species were also rather low (Figure 5B). Interestingly, the overall protein-abundance correlations are not equally tight across functional categories: the highest correlation was observed for the functional category “translation” (RS = 0.93) and the lowest for the category “regulation of biological process” (RS = 0.65).

12Despite the fact that it is difficult to compare tissues and developmental stages across organisms, our analysis provides a first insight into the evolutionary behavior of animal proteins over long time scales. It is important to point out that for all six data points, several developmental stages and/or tissues had been mixed, but that these were not, of course, always directly equivalent and comparable between the two organisms. However, many of the ancient animal orthologs that we studied here can be expected to be expressed similarly across many cell types and stages, and we thus attempted to capture an organism-wide “average” proteome for both animals. That notwithstanding, we also repeated the analysis for one set of samples that is arguably more directly comparable: mixed staged embryos sampled in both D. melanogaster and C. elegans at the proteome and at the transcript levels (Figure S4). Here again, we saw that protein abundances correlated far better (RS = 0.70) across organisms than transcript abundances (RS = 0.50).

13Another potential complication for our analysis lies in the technique of spectral counting. Individual tryptic peptides are known to ionize and be detected with widely differing efficiencies in mass spectrometers. Although protein conservation between C. elegans and D. melanogaster is low (∼40% sequence identity), a higher-than-expected abundance correlation might still result if equivalent peptides in both organisms were correlated in their suitability for MS. We assessed the extent of this effect by making the spectral counts independent: For any given section in the alignment of two orthologs, only one of the proteins was allowed to generate peptide counts; these sections were alternated across the length of the alignment, effectively reducing the data by half. As expected, this lowered the abundance correlation, but not by much (RS = 0.68). Importantly, the resulting correlation is still much higher than the correlation of transcript abundances across organisms (Figure S5).

14Since one of our original interests was to characterize the “core animal proteome,” we also analyzed lower-coverage datasets from two additional organisms: Saccharomyces cerevisiae and Mus musculus () (for the latter, we additionally included plasma data from PeptideAtlas; http://www.peptideatlas.org/). Comparative proteomics using multiple organisms has recently become popular, for example in bacteria, but has not yet been possible for animals. We searched for groups of orthologs that were detected in all four organisms; these would constitute the universally detectable eukaryotic proteome core. We found 847 such proteins, mostly from information-processing and metabolism genes. Conversely, we found 1,287 proteins to be detectable in all three animals, but not in yeast. This latter set might be considered the specific core of multicellular animal proteomes. However, it is clear that neither of these sets is complete, as of yet, mostly due to low coverage in mouse.

15Our protein-abundance estimates from two organisms also allowed us to study in more detail the fate of duplicated genes. Here, of particular interest, are cases in which a gene family has duplicated in one lineage (fly or worm), but not in the other. It is known that long-term retention of duplicated gene copies requires neo- or subfunctionalization, but it is unclear what consequences this has for overall protein-abundance levels. We found that when averaging over all cases of lineage-specific gene duplications (Figure S6), the abundance of duplicated genes is significantly lower than that of their nonduplicated counterparts in the other lineage. Strikingly, however, when all the duplicated genes of a given gene family are pooled, they tend to add up again to the original abundance of the nonduplicated counterpart (Figure S6).

16We describe here a comprehensive inventory of C. elegans proteins, the functional characterization of this inventory, and the first-ever comparison of two such inventories between two model animals (“comparative proteomics”). Although some subsets of the proteome are more difficult to analyze (e.g., the membrane compartment), we achieved a relatively thorough representation of the genome, where the major exceptions can be explained biologically. For example, the systematic underrepresentation of seven-transmembrane proteins appears to be caused mainly by G protein–coupled receptors. The putative chemoreceptor gene families in C. elegans encompass about 7% of its total genome, and many are thought to be expressed only in a few neurons each. Despite their generally low abundance, we did identify 172 seven-transmembrane receptor proteins, showing that they are, in principle, amenable to high-throughput MS analysis (this is relevant, for example, for screens of putative therapeutic targets).

17We also demonstrated that a whole-proteome analysis of a model organism can contribute to an improved genome annotation. First, we experimentally confirmed the expression of 1,062 predicted genes for which no transcript data were available, but for which our proteome data allowed the extraction of a first rough expression pattern. Second, we identified novel peptides from spectra that could not be matched to annotated gene models, suggesting a way to more precisely map open reading frames and splice isoforms to the genome.

18With respect to genome organization, we found that, in C. elegans, genes in operons are far more consistently and more strongly expressed than individually transcribed genes. In principle, this observation could be an artifact of genome annotation—if a disproportionally large number of annotated nonoperon genes were misannotations that are biologically meaningless. This is highly unlikely, however, since more than 6,000 such misannotations would be needed to reconcile the observed differences. Instead, it is likely that operons in C. elegans indeed serve to group strongly expressed genes into coregulated transcription units. Another question that arises is whether these genes were highly expressed even before they were grouped into operons, which would hint at a possible selective advantage for the grouping (e.g., to enable more efficient, more reliable, or more uniform transcription of genes whose products are in high demand). This is difficult to address conclusively, but our comparison to D. melanogaster provides some information: we observe that orthologs of operon genes are more strongly expressed even in the fly (Figure S7), where they are not arranged in operons nor are even neighbors on the genome. If one assumes that the operons in C. elegans are the derived state, then the corresponding genes were indeed already strongly expressed before they formed operons.

19The comparison of our data to the D. melanogaster proteome also sheds some light on an important evolutionary puzzle, namely the surprisingly low correlation between mRNA expression levels of orthologous genes across animal species, despite evidence for strong stabilizing selection against expression changes in experimental evolution. We found that the abundances of orthologous proteins from worm and fly correlate well (RS = 0.79), far better than the corresponding abundances of mRNA transcripts (RS < 0.50; Figure 5B).

20There are several possible explanations for this finding: First, sweeping changes within the transcriptional machineries in one or both organisms could have resulted in global differences in transcript abundance, whereas selection would have kept protein abundances at least partially stable. One candidate for such a mechanistic change could be, for example, the unique trans-splicing mechanism of nematodes. A second possible explanation might be that posttranslational regulation may have changed systematically, for example due to differences in developmental strategies, physiology, or life styles of the two animals. Here, possibly relevant changes include the fixed cell lineage of nematodes, differences in reproductive strategies, increased endurance in nematodes (dauer stage), or the constraints imposed on D. melanogaster because of its need for metamorphosis and its higher motility (flight).

21However, in our view, the most parsimonious explanation might be that many changes in the transcriptome might be neutral, or at least nearly neutral. Ultimately, it is the protein levels that are under selection. Protein levels are not only determined by mRNA abundance, but are equally affected by translation efficiencies, protein half-lives, and other factors. Genetic mutations resulting in small changes on any of these levels might persist for some time in a population, as long as their fitness effects are small (around 1/[2Ne] or less). This might be sufficient time to allow for compensatory mutations either in the same gene or elsewhere in the genome, which would reconstitute optimal protein abundance through action on the same or another factor that influences protein abundance. Thus, changes in mRNA expression could be offset by opposite changes in translation rate or protein half-life, and vice versa. Over evolutionary time scales, such small changes may accumulate, resulting in appreciable changes of mRNA abundance, whereas protein abundance would remain roughly constant. This model is a generalization of the concept of compensatory mutations that explains the rapid divergence of some cis-regulatory nucleotide sequences despite the maintenance of stable transcript levels, or the conserved expression of assembled protein complexes despite variable expression patterns of their individual components.

22The presence of several interacting levels of protein-abundance regulation also would explain another two of our observations: a wide variance of the number of mature proteins per transcript, and a correspondingly low correlation between protein and transcript abundance within an organism (interestingly, the latter correlation is quite similar between our C. elegans data and data published in yeast [RS = 0.57]). Our data, in principle, provides an opportunity to study transcript features that would directly influence the ratio of proteins per transcript (and thereby potentially uncover novel mechanisms of translational regulation). However, when checking the influence of transcript length, GC content, or UTR length, we failed to detect correlations with protein/transcript ratios (unpublished data). We did observe a weak, but significant, positive correlation of our protein/transcript ratios and experimental protein half-life measurements of orthologous proteins in yeast (unpublished data), suggesting that protein stability is indeed one of the factors determining the steady-state protein/transcript ratio.

23We note that the most abundant proteins (often found in central pathways like energy metabolism or protein synthesis) also tend to be the ones that show the best abundance correlation between species. This may simply be the case because of a greater relative measurement accuracy for abundant proteins. However, highly expressed genes are also more likely to be housekeeping genes, and may thus be more likely to be under the same evolutionary pressures in different organisms. Strong and constant stabilizing selection is also consistent with our observation that amino acid sequences of more highly expressed proteins evolve more slowly (Figure 5C), mirroring the analogous observation for mRNA expression data.

24When we stratify proteins by functional categories, we find that those involved in translation and in core metabolism are those with the most highly correlated abundances across species. These functional groups are also those where the coexpression between pairs of transcripts is most highly conserved across species. Furthermore, the same categories also tend to show the best correlation within each organism, with respect to rank-correlation between transcripts and proteins (Table S2). We also find that the correlation between transcript and protein levels is particularly poor for genes that are presumably heavily regulated (the categories “signal transduction” or “transcriptional regulation”), arguing for abundant posttranscriptional regulation in these functional classes.

25Proteins differ not only in their mean abundance, but also in the variance of this abundance among individuals (“noise”). Interestingly, whereas yeast proteins involved in translation also show low levels of noise, other groups of proteins found here to be conserved in their abundance between species (e.g., protein metabolism) are characterized by high protein expression noise. Thus, it appears that abundance fluctuations on short time scales (within populations) are partially decoupled from fluctuations on long time scales (between species). However, as natural variation is the substrate of evolutionary change, we expect that changes in mRNA levels via compensatory mutations may occur faster in proteins that exhibit higher levels of noise; this remains to be tested in future studies.

26Our comparative analysis underlines clearly the necessity and usefulness of quantitative proteome analyses, since these better reflect the abundance of the actual effectors of biological processes. Most likely, the actual conservation of protein levels is even higher than what we report here, due to the shortcomings of a simple spectral-counting procedure. In fact, comparisons across organisms might generally provide a good test scenario to improve spectral-counting algorithms or other proteomics algorithms: the higher the abundance correlation, the more precise the measurements (due to the high number of data points, and due to the quickly changing positions of tryptic cleavages, this is difficult to “over-train” by choosing biased parameters). With respect to the transcriptomics datasets that we used, the above test argued for a better quality of the Affymetrix data, as compared to SAGE, because the latter were seen to correlate less well across organisms. This is intriguing, and it may point to additional biases in the SAGE procedure (for example, due to the added molecular biology steps of cleavage and ligation).

27For those instances where orthologs were not found to be of similar abundance, one can speculate that this difference reflects differing roles (or even molecular functions) of the orthologs. Thus, these proteins are of particular interest when studying the evolutionary differences between species. Alternatively, differences in technical aspects for particular proteins might occur, such as shifted or absent trypsin cleavage sites or differences in protein solubility. Interestingly, we did not lose the observed interspecies correlation even for quite low-abundance proteins such as those involved in signal transduction (our measurements have a dynamic range of more than three orders of magnitude). This means that low-abundance measurements are still quantitative, at least to some degree.

28In our analysis of gene families with lineage-specific duplications, we found that duplicated proteins generally have lower abundance than their nonduplicated counterparts, whereas the summed abundances per gene family remained roughly constant. This finding might be most parsimoniously explained by a prevalence of subfunctionalization among duplicated genes, although it is also consistent with other scenarios (e.g., complementarity of tissue expression domains, functional fine-tuning, or subfunctionalization followed by neofunctionalization). Of course, protein abundances alone cannot directly inform us about any changes in the functions of duplicated genes. However, our finding does suggest that cases where an increased demand for protein product would provide the sole driving force behind gene copy retention are probably rare.

29With our dataset, we established an inventory of where and how proteins of interest can be specifically accessed using MS. It enables the generation of a proteotypic peptide library (i.e., peptides in a protein sequence that are most likely to be consistently and confidently observed by current MS-based proteomics methods). This library in turn can be used for targeted analyses and comparative studies of expressed proteins by spiking the samples to be analyzed with chemically synthesized proteotypic peptides, or by selected reaction monitoring (SRM) MS. Our C. elegans proteome dataset will be made publicly available within WormBase and will thus be useful for the entire C. elegans research community. In general, proteomics data like ours is closer to the biologically active players than transcriptomics data. It should therefore be increasingly used to investigate biological phenomena and mechanisms underlying disease pathogenesis such as neuronal degeneration and cancer development, and for the identification of conserved therapeutic target proteins.

30C. elegans wild-type strain N2 (Bristol) was grown on 9-cm nematode growth medium (NGM) agar plates seeded with a lawn of the E. coli strain OP50 or in 100 ml of liquid cultures in S-basal buffer in beveled flasks. Worms were harvested from plates or liquid culture, and separated from the bacteria by washing with water or sucrose flotation. For the collection of embryos, the worms were synchronized, and eggs were removed from agar plates or obtained from the hermaphrodites by bleaching. Worm and egg pellets were homogenized with glass beads (diameter of 212–300 μm; Sigma-Aldrich) in the ratio of 1:1:2 (worms:beads:buffer) in a cell disrupter (FastPrep FP120, Thermo Savant; Qbiogene) at 4 °C three times for 45 s at level 6. The buffer used was 50 mM Tris/HCl (pH 8.3), 5 mM EDTA, 8 M urea. After glass bead treatment, 0.125% SDS was added, and the homogenate was incubated for 1 h at room temperature (RT) to solubilize proteins. For other experiments, the worms or eggs were homogenized with glass beads in 50 mM Tris/HCl (pH 8.3), 5 mM EDTA, then 0.75% or 1% Rapigest (Waters) was added, the homogenate was heated at 95 °C for 5 min, and incubated at RT for 30–60 min with gentle agitation. Cell debris was removed by centrifugation, and the protein concentration was determined using the Bradford reagent (Sigma-Aldrich).

31The peptides were subjected to reversed-phase capillary chromatography using a 75-μm × 8-cm self-packed C18 column (Magic C18; Michrom) at a flow rate of 250 nl/min. Peptides were eluted with a gradient between solvent A (5% ACN, 0.2% formic acid) and solvent B (80% ACN, 0.2% formic acid). The gradient was from 5% up to 45% solvent B within 69 min. The peptides were identified by CID (collision induced dissociation) on a Thermo-Finnigan ion trap mass spectrometer “LTQ”. Six dependent scans followed each survey scan. Raw data were converted into mzXML files and searched against a C. elegans database derived from the Wormpep database (http://www.wormbase.org, release WS140) using the Sequest program. The search parameters used were two missed cleavage sites, two tryptic termini, a mass tolerance of 3 Da for the parent ion and 0.95 Da for the fragment ion, optional oxidized methionine, and depending on the experiment, modified cysteine. Peptide assignments were statistically validated at peptide level using PeptideProphet, and peptides with a probability score of 0.9 or higher and the proteins they belong to were selected. For the qualitative analysis of the proteome (Figure 2), peptides matching to more than one protein (such as duplicated tubulins or histones), or matching to several splice variants of a protein, were counted only once (for the first entry of the search results). For the quantitative analysis, however, such peptides were assigned fractionally (see below). From a total of 18 different experiments (Table S3), we identified 10,977 proteins from 10,631 gene loci (Table S1). The comparative analysis of the different protein parameters was also based on WS140. For technical reasons, all the information for the other functional analyses was extracted from release WS160 using WormMart (http://www.wormbase.org/biomart/martview). The FDR for single hits was estimated first based on an experiment in which isoelectric focusing of peptides was performed on an immobilized pH gradient strip (pH range 3–5.6), followed by subsequent analysis of computationally predicted pIs for each peptide identification, and second by a new model based on a decoy search strategy (L. Reiter, M. Claassen, S. P. Schrimpf, J. M. Buhmann, M. O. Hengartner, et al., unpublished data). To evaluate potential bacterial contamination in our dataset, one experiment was searched against a combined C. elegans (WormBase WS140) and E. coli (SPproteomes at the European Bioinformatics Institute [EBI], release 2005-03-19, 4,338 entries) database using the same search parameters as for the searches against the C. elegans database.

32After redundancy analysis, 22,269 distinct proteins (including splice variants, WormBase WS140) and 10,977 proteins in our dataset were compared for the bias analysis with respect to different protein parameters. Tools from the ExPASy Web site (http://www.expasy.ch) were used to calculate the pIs of proteins (protein parameter tool “protparams”) and their hydrophobicity (gravy computation “grand average hydrophobicity”). The statistical analysis shown in Figure S8 was carried out as described before; the p-values for all parameter analyses were 1E−10 or better.

33The number and orientation of transmembrane domains of the proteins in WormBase (WS160) and in our dataset were predicted using Phobius. Only gene loci—not splice variants—were processed. Whenever transmembrane predictions differed for splice variants, the predictions for the longest splice variants were used. For the GO slim analysis, the GO terms listed in WormBase (WS160) were mapped onto higher-level terms using the GO slim guide (http://www.geneontology.org), with two exceptions: the terms “membrane” and “integral to membrane” were not mapped to the higher category term “cell,” but instead were retained. In Figures 2E and S9, we assigned the GO slim terms of the category “molecular function” to the predicted transmembrane proteins. For 412 proteins, there was more than one entry for molecular function. For the statistical analysis of the GO slim categories in Figure 2, we applied the Fisher exact test and included the Bonferroni correction for multiple testing. We plotted the log ratio of observed versus expected, using the proportions in WormBase as the expectation. The GO slim categories with a p-value better than 0.05 are shown (Figure 2E and 2F).

34We mined our dataset for nonannotated translated regions by preparing a whole-genome open reading frame database that was searched using the Sequest algorithm. To do this, WormBase release WS160 was used to translate each chromosome into all six reading frames. Open reading frames longer than 20 amino acids were assembled into a database with headers containing the coordinates of the sequences on the genome. The resulting database contains 3,136,258 open reading frames and 132,018,220 amino acids. A subset of our data (experiment 15) obtained by isoelectric focusing, comprising approximately 304,000 MS/MS spectra, was searched at the Functional Genomics Center Zurich. We allowed fully tryptic peptides with up to two missed cleavages, and specified oxidized methionine as variable modification and carbamylated cysteine as static modification. The results were further analyzed with PeptideProphet, and 27,940 search hits with a PeptideProphet score greater than or equal to 0.95 were selected. From these, we removed 26,952 scans that also generated a hit against the normal Wormpep140 protein database with a score greater than 0.8. Of the remaining 988 spectra, 789 were further observed to exist in Wormpep178 or an E. coli database and were therefore omitted, resulting in a final set of 199 spectra belonging to 173 different peptides. For the resulting peptides, a theoretical pI value was calculated and compared to the mean pI of all peptides in the corresponding fraction. Only peptides with a delta pI value smaller than or equal to 0.5 were selected. This resulted in 78 distinct peptides.

35WormMart (http://www.wormbase.org/biomart/martview) was used to extract operon architectures from WormBase release WS160. To test whether the coregulation of genes in operons would be detectable also at the level of translated proteins, operons were first divided into length classes (here, length is defined as the number of cotranscribed genes in each operon). For each length class, the fraction of operon genes was then determined for which at least one peptide was detected in at least one proteomics experiment. This fraction determines how many proteins should, on average, be detectable from a single operon if expression of the operon genes were truly independent (when assuming independence, the number of detections per operon should follow a binomial distribution, shown as grey lines in Figure S2). Applying the two-sided Kolmogorov-Smirnov test yielded p-values better than 1E−10. For the study of operons in specific stages (Figure S3), only the proteome data was analyzed, limited to the experiments done in these stages (with concomitantly reduced spectral counts).

36For the semiquantitative comparison between C. elegans and D. melanogaster proteomes, we used the STRING database and the Smith-Waterman similarity relations stored therein to compute orthologous groups. This analysis retrieved 4,184 loci in C. elegans, and 4,302 in D. melanogaster. When working with orthology sets, each pair of orthologs was aligned with “muscle”, available from http://www.drive5.com/muscle/. The protein sequences used were extracted from WormBase WS160 and from FlyBase release 5.1 (http://flybase.org). Due to lineage-specific gene duplications, some proteins had several orthologs. For the interspecies abundance correlation comparison, we summed up the abundances in these cases.

37We independently tested another source of orthology information, InParanoid, which resulted in slightly more orthologs but also in a somewhat lower interspecies abundance correlation (RS = 0.76 versus 0.79). Conversely, we also tested a stricter set of orthologs, to test for and exclude artifacts caused by potentially undetected paralogy. To conclusively separate orthologs from paralogs can be difficult, and this is the subject of intense study. Therefore, we constructed a very strict set of orthologs by searching for reciprocal best matches between worm and fly, with the additional constraint that any extra homologs within these genomes had to exhibit no more than half the alignment score than the score between these organism (plus, the score between the organisms had to be 60 bits or higher). This strict set contained only 2,001 pairs of orthologs, and resulted in an interspecies abundance correlation of 0.80. This shows that our high correlation is not caused, or affected, by the presence of paralogs in the comparison.

38We calculated the relative abundance of a protein by counting how often any of its amino acids had been identified in any peptide, divided by the total number of amino acids of the protein sequence. A length restriction to peptides with ≥7 and ≤40 amino acids (modified from) was applied.

39where a = protein abundance, p = identified peptides, q = tryptic peptides (in silico digest), and f(q) = peptide length correction factor.

40The peptide-length correction factor takes into account the technical bias of the MS instrument, which resulted in certain peptide lengths being observed more often than others. This was learned from the data by comparing the observed peptide-length spectrum with the expected, and was corrected accordingly (similar to). In our hands, peptide length proved to be the most important determinant of peptide observability, since using the original APEX implementation (“absolution protein expression profiling”) or a retrained version of the same classifier, did not further improve the observed cross-organism abundance correlation between C. elegans and D. melanogaster (RS = 0.78).

41A relative protein abundance of 1 means that the total number of amino acids in the identified peptides equals the number of amino acids in the protein. Whenever a peptide could be assigned to several proteins (because of identical predicted tryptic peptides), the amino acids were assigned fractionally. Peptides specific for any of the splice isoforms originating from a given locus were pooled. This approach means that the unit of interest in our comparisons is the gene locus—not individual splice isoforms—consistent with the observed lack of conservation of alternative splicing at very large evolutionary distances. Finally, protein abundances were normalized to total amount of protein detected. To plot the data, orthologs were binned into eight groups of equal size (sorting for binning was x + y), and the means, as well as first and third quartiles, for each group were calculated. For the comparison of gene and protein expression, SAGE data for C. elegans were downloaded from http://tock.bcgsc.bc.ca/cgi-bin/sage160. In order to best reflect the developmental stages analyzed in our proteome data, we chose the stages SWN21, SWL12, SWL21, SWL32, SWL41, SWYA1, MIXED, SW022, and DAUER. Only entries with “source = coding_RNA” were considered, and the average of the nine columns was calculated. SAGE data for D. melanogaster were obtained from Professor San Ming Wang (Northwestern University, Evanston, Illinois). D. melanogaster SAGE tags were mapped to all transcripts from FlyBase release 5.3. The C. elegans Affymetrix GeneChip data were obtained from the Genome British Columbia C. elegans Gene Expression Consortium at http://elegans.bcgsc.bc.ca. The D. melanogaster Affymetrix GeneChip data were obtained from http://www.flyatlas.org. For 2,695 pairs of orthologs protein abundance, SAGE and Affymetrix data were compared (in case of several paralogs, only one of them had to have data from all three measurements). For the comparisons of different abundances, Spearman rank correlation coefficients were computed to avoid assumptions about the underlying distributions. Probabilities for the correlation coefficients were calculated as implemented in R; all corresponding p-values were better than 2.2E−16. Further supporting the validity of spectral counting as a semiquantitative measure is a comparison of C. elegans protein abundance data against protein abundance data in yeast. Importantly, the latter is not based on MS, but on immunodetection of tagged open reading frames. Orthologs correlate linearly in their abundance over two orders of magnitude (RS = 0.54; Figure S10). The correlation for sequence conservation (aligned to D. melanogaster) and protein abundance was calculated for 4,013 C. elegans proteins. Orthologs were binned into eight groups of equal size (Figure 5C).