The genome of a species encodes genes and other functional elements, interspersed with non-functional nucleotides in a single uninterrupted string of DNA. Recognizing protein-coding genes relies on finding stretches of nucleotides free of stop codons (called Open Reading Frames, or ORFs) that are too long to have likely occurred by chance. Since stop codons occur at a frequency of roughly 1 in 20 in random sequence, ORFs of at least 60 amino acids will occur frequently by chance (5% under a simple Poisson model) and even ORFs of 150 amino acids will appear by chance in a large genome (0.05%). This poses a huge challenge for higher eukaryotes in which genes are typically broken into many, small exons (on average 125 nucleotides long for internal exons in mammals27).
The basic problem is distinguishing real genes those ORFs encoding a translated protein product from spurious ORFs the remaining ORFs whose presence is simply due to chance. The current public catalogue of yeast genes lists 6062 predicted ORFs that could theoretically encode proteins of at least 100 amino acids. Only two-thirds of these have been experimentally validated (known), and the remaining ~2000 ORFs are currently annotated as hypothetical. The total number of real protein-coding genes has been a subject of considerable debate, with estimates ranging from 4,800 to 6,400 genes (in mammalian genomes, estimates have ranged from 28,000 to more than 120,000 genes).
In this chapter, we use the comparative information to recognize real genes based on their conservation across evolutionary time. With the availability of genome-wide alignments across the four species, we first examined the different ways by which sequences change in known genes and in intergenic regions. The alignments of known genes revealed a clear pressure to preserve protein-coding potential. We constructed a computational test for reading frame conservation (RFC) and used it to revisit the annotation of yeast. We showed that more than 500 previously annotated ORFs are not meaningful and discovered 43 novel ORFs that were previously overlooked. We additionally refined the gene structure of hundreds of genes, including translation start, stop, and exon boundaries. We show that our method has high sensitivity and specificity, and suggest changes that affect nearly 15% of yeast genes.
We examined the different types of conservation in genes and intergenic regions. We used the 1-to-1 orthologous anchors (see Chapter 1) to construct a nucleotide-level alignment of the genomes. The strong conservation of local gene order and spacing (Figure 2.1) allowed us to construct genome-wide multiple alignments. We aligned each gene together with its flanking intergenic regions using CLUSTALW38 for the multiple alignments across the four species. When sequence gaps were present in one or more species, we constructed the alignment in multiple steps. We first aligned the gapless species creating a base alignment. Then we aligned each portion of a partially covered ortholog onto the base alignment, and constructed a consensus for each species based on the individually aligned portions. We marked missing sequence between contigs by a dot and disagreeing overlapping contigs by N. Finally, we constructed a multiple alignment of the four species by merging the piece-wise alignments. With sequence alignments at millions of positions across the four species, it is possible to obtain a precise estimate of the rate of evolutionary change, including substitutions and insertion-deletions (indels), in the tree connecting the species. We counted transitions, transversions, insertions and deletions within these alignments and used these to estimate the rate of evolutionary change between the species. We counted the rate of synonymous and non-synonymous substitutions for every protein coding gene to find evidence of positive selection. The detailed results will be described in chapter 6.
We compared the rate of sequence change at aligned sites across the four species in intergenic and genic (protein-coding) regions (Figure 2.2). We found radically different types of conservation. Intergenic regions typically showed short stretches between 8 and 10 bases of near-perfect conservation, surrounded by non-conserved bases, rich in isolated gaps. Protein-coding genes on the other hand were much more uniform in their conservation, and typically differed in the largely-degenerate third-codon position. The proportion of sites corresponding to a different nucleotide in at least one of the three species is 58% in intergenic regions but only 30% in genic regions a difference of ~2-fold. The difference becomes much greater when one considers the gapped positions in alignments, representing insertion and deletion events (indels). The proportion of indels is 14% in intergenic regions, but only 1.3% in genic regions. The contrast is even sharper for indels whose length is not a multiple of three. These would disrupt the reading frame of a functional protein-coding gene, and are detrimental when they occur in real genes, unless they are compensated by a nearby indel that restores the reading frame. Frame-shifting gaps are found in 10.2% of aligned positions in intergenic regions, but only in 0.14% of positions in genic regions, a 75-fold strong separation. We used these alignment properties to recognize real genes.
We developed a Reading Frame Conservation (RFC) test to classify each ORF in S. cerevisiae as biologically meaningful or not, based on the proportion of the ORF over which reading frame is locally conserved in each of the other three species. Each species with an orthologous alignment cast a vote for accepting or rejecting the ORF, and the votes were tallied to reach a decision for that ORF.
We evaluated the percent of nucleotides that are in the same frame within overlapping windows of the alignment. For every such window, we labeled each nucleotide of the first sequence by its position within a codon, as 1, 2 or 3 in order, starting at codon offset 1. We similarly labeled the nucleotides of the second sequence, but once for every start offset (1, 2, or 3). We then counted the percentage of gapless positions in the alignment that contained the same label in both aligned species, and selected the maximum percentage found in each of the three offsets of the second sequence (Figure 2.3). The final RFC value for the ORF was calculated by averaging the percentages obtained at overlapping windows of 100 nucleotides starting every 50 nucleotides. For overlapping ORFs in the S. cerevisiae genome (n = 948), the RFC was calculated only for the portion unique to each overlapping ORF. For spliced genes (n = 240), the RFC was calculated only on the largest exon.
We found that the distribution of frame conservation within each species is bimodal, and we chose a simple cutoff for each species, 80% for S.paradoxus, 75% for S.mikatae and 70% for S.bayanus. If the RFC of the best hit was above the cutoff, a species voted for keeping the ORF tested. If the RFC was below the cutoff and the hit was trusted as orthologous, the species voted for rejecting the tested ORF. Finally, if no orthologous hit could be found due to coverage, a species abstained from voting. We calculated a score between 3 and +3 for every ORF based on the number of species that accepted it (+1) and the number of species that rejected it (-1). We kept all ORFs with a score of 1 or greater, and rejected all ORFs with a score of 1 or smaller. We manually inspected the remaining ORFs.
We also applied this test to 3966 annotated ORFs with associated gene names (Table 2.4). These have been studied and named in at least one peer-reviewed publication, and are likely to be represent real genes. Only 15 of these (0.38%) were rejected (KRE20, KRE21, KRE23, KRE24, VPS61, VPS65, VPS69, BUD19, FYV1, FYV2, FYV12, API2, AUA1, ICS3, UTR5, YIM2). We inspected these manually and concluded that all were indeed likely to be spurious. Most lack experimental evidence. For the remainder, reported phenotypes associated with deletion of the ORF seems likely to be explained by fact that the ORF overlaps the promoters of other known genes.
To investigate the power of the approach to reject spurious ORFs, we also applied it to a set of controls sequences consisting of 340 intergenic sequences in S. cerevisiae with lengths similar to the ORFs tested (Table 2.4). About 96% were rejected as having conservation properties incompatible with a biologically meaningful ORF, showing that the test has high sensitivity. Of the remaining 4% that were not rejected, close inspection shows that three-quarters appear to contain true ORFs. Some define short ORFs with conserved start and stop codons in all four species and others extend S. cerevisiae ORFs in the 5- or 3-direction in each of the other three species. Thus, at most 1% of true intergenic regions failed to be rejected by the RFC test.
The conservation-based gene identification algorithm we proposed has thus high sensitivity and specificity. In the next section, we apply it systematically for de-novo gene identification in S. cerevisiae.
When the yeast genome sequence was completed23, 6275 ORFs were identified in the nuclear genome that could theoretically encode proteins encoding at least 100 amino acids and that do not overlap a longer ORF by more than half of their length (Figure 2.5). SGD has since updated the catalog based on complete resequencing and re-annotation of chromosome III, re-analysis of other chromosomes and reports in the scientific literature. This resulted in a current version (as of May 2002) with 6062 ORFs ≥ 100 amino acids, consisting of 3966 named genes (described in at least one publication) and 2096 uncharacterized ORFs. SGD also includes a small collection of ORFs < 100 amino acids (see below).
We sought to apply the RFC test to all 6062 ORFs in SGD. A total of 117 could not be analyzed because they were almost completely contained within an overlapping ORF (99 cases, with average non-overlapping portion = 12 bp) or because an orthologous region could not be unambiguously defined in any of the species (18 cases). Of the 5945 ORFs tested, the analysis strongly validated 5550 ORFs. The vote was unanimous in 5458 (~98%) of cases. In the remaining cases, a valid gene appears to have degenerated in one of the four species. A total of 367 ORFs were strongly rejected. These rejections were unanimous in 63% of cases. In most of the remaining cases, S. paradoxus was too closely related to S. cerevisiae to have accumulated enough frameshifts to allow definitive rejection. The analysis deadlocked (one confirmation, one rejection, one abstention) for 28 ORFs (0.5%). We inspected these, together with the 117 cases that could not be analyzed due to overlaps and found convincing evidence (based on conservation of amino acids, start and stop codons, and presence of indels), that 20 are valid protein coding genes and 105 are spurious. We were unable to reach a judgment in the remaining 20 cases. Overall, a total of 5570 ORFs were accepted, 472 ORFs were rejected, and 20 remain ambiguous.
The vast majority of the rejections (96%) involve uncharacterized ORFs (for an example see Figure 2.6). SGD reports no compelling biological evidence (such as changes in mRNA expression) to suggest that these ORFs encode a true gene. Most of these overlap another well-conserved ORF, but show many insertions and deletions in the non-overlapping portion. The remainder tend to be small (median = 111 aa, with 93% ≤ 150 aa) and show atypical codon usage23,39,40. Figure 2.6 illustrates the case of an ORF of 333 bp that is clearly biologically meaningless. The orthologous sequence in all four species is laden with frameshifts (as well as stop codons). Only one rejected ORF, YBR184W, appears to represent a true gene that fails the RFC test because it is evolving very rapidly (see section 6.6).
In summary, the Reading Frame Conservation (RFC) test allowed a major revisiting of the yeast genome annotation. By observing the pattern of indels in the multiple alignment of predicted ORFs, it allowed us to automatically classify them as biologically meaningful or spurious. It reached a decision automatically in 98% of cases, accepting 99% of named ORFs and rejecting 99% of real intergenic regions, showing strong sensitivity and specificity. It resulted in a drastic reduction of the yeast gene count, rejecting nearly 500 ORFs. We next use the comparative information to refine the boundaries of ORFs.
Comparative genome analysis not only improves the recognition of true ORFs, it also yields much more accurate definitions of gene structure including translation start, translation stop and intron boundaries. We used the comparative data to identify sequencing errors and refine the boundaries of true genes. Previous annotation of S. cerevisiae has defined the start of translation as the first in-frame ATG codon. However, the actual start of translation could lie 3 to this point, and the earlier in-frame ATG may be due to chance. Alternatively, if sequencing errors or mutations have obscured an earlier in-frame ATG codon, the true translation start could lie 5 to this point. Similarly, the annotated stop codon could be erroneously annotated, due to sequencing errors. Identifying the correct gene boundaries is important for many reasons, both experimental (for example to construct gene probes), as well as computational (for example to search for regulatory motifs).
We examined the multiple alignment of unambiguous ORFs to identify discrepancies in the predicted start and stop codons across the four species. We searched for the first in-frame ATG in each species and compared it to the annotated ATG in S. cerevisiae. In the S. cerevisiae start was not conserved, we automatically suggested a changed translation start if a subsequent in-frame ATG was conserved in all species and was the first in-frame ATG in at least one species. Otherwise, we searched for a conserved ATG 5 to that point. Similarly, we suggested changes in stop codons when a common stop in all other species disagreed with the S. cerevisiae annotation. We manually inspected the alignments to confirm that the suggested start and stop boundary changes agreed with conservation boundaries. We identified merges of consecutive S. cerevisiae ORFs, when they unambiguously matched a single ORF in at least one other species, and when their lengths added up to the length of the matching ORF.
We identified 210 cases in which the presumed translational start in S. cerevisiae does not correspond to the first in-frame start codon in at least two of the three other species (Figure 2.7 panel 1). In the vast majority of these cases, inspection of the sequence alignments provides strong evidence for an alternative conserved position for the translational start, either 3 or 5 to the previous annotation. We observed a lower overall conservation as well as frame-shifting indels outside the new boundaries. Similarly, we identified 330 cases in which the presumed translational stop codon in S. cerevisiae does not correspond to the first in-frame stop codon in at least two of the three species. In ~25% of these cases, the other three species share a common stop codon and a single base change to the S. cerevisiae sequence would result in a stop codon in the corresponding location (Figure 2.7 panel 2). The remaining 75% of cases appear to represent true differences in the location of the translational stop across the species. Thus, stop codons appear to show more evolutionary variability in position than start codons.
We also developed methods for the automatic detection of frame-shifting sequencing errors. When regions of the multiple alignment shifted from one well-conserved reading frame to another well-conserved reading frame, we pinpointed regions of potential sequencing errors in each of the species. A number of these were detected in the reference sequence of S. cerevisiae. We confirmed 32 of these computational predictions by resequencing and found that in each case the published sequence was in error, and that the predicted erroneous nucleotide was always within a few base pairs from the experimentally confirmed sequencing error.
We identified 32 cases where two adjacent ORFs in S. cerevisiae are joined into a single ORF in all three other species. In every case, a single nucleotide change would suffice to join the ORFs in S. cerevisiae (either a substitution altering a stop codon or an indel altering the reading frame). In principle, these cases could represent errors in the genome sequence, mutations private to the sequenced strain S288C, or substitutions fixed in S. cerevisiae. We examined 19 cases by resequencing the relevant region in S288C. Our results revealed an error in the published sequence in 11 cases (establishing that there is a single ORF in S288C) and confirmed the published sequence in the remaining 7 cases. Sequencing of additional strains will be required to determine whether these remaining cases represent differences in S288C alone or in S. cerevisiae in general.
We also found two named ORFs (FYV5 and CWH36) that pass the RFC test and cause phenotypes when deleted, but show no significant protein similarity across the four species. In both cases, inspection reveals that the opposite strand encodes a protein that shows strong amino acid conservation. (The latter gene has two introns, increasing the count of doubly spliced genes to 8.) In each case, we postulate that the protein responsible for the reported deletion phenotype is encoded on the opposite strand.
All merges and boundary refinements suggested specific changes to the nucleotide sequence of S. cerevisiae (except 3 changes of translation start that required no change). To validate our predictions, we re-sequenced the sites of predicted sequence discrepancies. We used both forward and reverse reads in two different PCR reactions spanning the site. We examined 4 cases in which the comparative data suggested an earlier start codon and found, by resequencing, that all correspond to errors in the published sequence of S288C. We examined 17 such cases and found that 15 are explained by errors in the published sequence of S288C.
New Introns. We also examined the conservation of introns in the yeast genome. We studied 218 of the 240 ORFs reported in SGD to contain at least one intron (omitting the rest primarily due to lack of an orthologous alignment). In 92% of cases, the donor, branchpoint, and acceptor sites were all strongly conserved with respect to both location and sequence. Moreover, exon boundaries closely demarcated the domains of sequence conservation as measured by both nucleotide identity and absence of indels. Discrepancies were found in 17 cases, of which at least 9 strongly suggest that the previous annotation is incorrect. Five identify a new first exon (Figure 2.8) and four predict that a previously annotated intron is spurious.
We then sought to identify previously unrecognized introns by searching the S. cerevisiae genome for conserved splicing signals. We searched for conserved and proximal splice donor and branch signals and manually inspected the resulting alignments. Having constructed multiple alignments of ORFs and flanking intergenic regions, we searched for conserved splicing signals. We used 10 variants of splice donor signals (6-7bp) and 8 variants of branch site signals (7bp) that are found in experimentally validated S. cerevisiae introns41. We searched each species independently but required that orthologous signals appear within 10 bp from each other in the multiple alignment of the region. We also required that branch and donor be no more than 600bp apart, which is the case for 90% of known S. cerevisiae introns. We then inspected the multiple alignment surrounding the conserved signals for three properties: (1) a conserved acceptor signal, [CT]AG, 3 of the branch site (2) high RFC 5 of the donor signal and 3 of the acceptor signal. (3) low RFC within the intron. Roughly half of the conserved donor/branch pairs met our additional requirements.
We predict 58 novel introns. Fifty cases affect the structure of known genes (defining new 5-exons in 42 cases, 3-exons in 7 cases and an internal splice in one case) and two indicate the presence of new genes. The relationship of the apparent splice signals to existing genes is unclear for the remaining six cases. We visually inspected our predictions and compared our results to experimental studies by Ares and colleagues that identified new introns using techniques such as microarray hybridization41. Of our 58 predicted introns, 20 were independently discovered by this group. Of the four annotated introns predicted to be spurious, all four show no experimental evidence of splicing. Our remaining predictions are currently being tested in collaboration with Ares and colleagues.
The power of our method was limited for small ORFs. Smaller regions may indeed show lack of indels due to chance, and hence a high reading frame conservation score may not be meaningful.
We tested 141 ORFs encoding 50-99 amino acids for which some biological evidence has been published and are reported in SGD. Applying the RFC test and inspecting the results, we conclude that 120 appear to be true genes, 18 appear to be spurious ORFs and 3 remain unresolved. SGD also lists 32 ORFs encoding < 50 aa. We did not undertake a systematic search for all such ORFs, because control experiments showed that the RFC test lacked sufficient power to prove the validity of such small ORFs (see below). However, it is able to reject 7 of the 32 ORFs as likely to be spurious. Our yeast gene catalogue thus contains 188 short genes (<100 aa), of which 43 are novel.
To evaluate the predictive power of the RFC test for small ORFs, we additionally tested for presence of in-frame stop codons in the other species. When a small ORF in S. cerevisiae showed a strong overall frame conservation, we measured the length of the longest ORF in the same orientation in each orthologous locus. We measured the percent of the S. cerevisiae length that was open in each species (no stop codons), and took the minimum of the three percentages (OPEN) across the three additional species. When the reading frame was open in each of the other species, the lengths found were identical to that of S. cerevisiae, and OPEN was 100%. When OPEN was below 80%, we concluded that stop codons appeared in the orthologous sequence, and therefore that the RFC test falsely accepted a segment that did not correspond to a true gene. We observed the distribution of OPEN for different values of RFC. For S. cerevisiae ORFs between 50 and 100 amino acids (aa), selecting for high RFC automatically selected for high OPEN, and we estimated the test has high specificity. For ORFs between 30 and 50 aa however, only a small portion of the ORFs with high RFC show a high OPEN, and we conclude that the lack of indels within the small interval considered is not due to selective pressure, but instead lack of evolutionary distance between the species aligned.
We further systematically searched the remainder of the S. cerevisiae genome and evaluated all ORFs in this size range. Control experiments demonstrated that the RFC test has high power to discriminate reliably between valid and spurious ORFs in this size range. The genome contains 3161 such ORFs, nearly all are readily rejected by the RFC test. However, 43 novel genes were identified. These ORFs not only pass the RFC test, but they also have orthologous start and stop codons. Five of these have been reported in the literature subsequent to the SGD release studied here
Based on the analysis above, we propose a revised yeast gene catalog consisting of 5538 ORFs ≥ 100 amino acids. This reflects the proposed elimination of 503 ORFs (366 from the RFC test, 105 by manual inspection and 32 through merger). A total of 20 ORFs in SGD remain unresolved. Complete information about the gene catalog is provided in 29 and will be discussed more fully in a subsequent manuscript in collaboration with SGD and other yeast investigators. The revised gene count is consistent with at least two recent predictions based on light shotgun coverage of related species4,5. We believe that this represents a reasonably accurate description of the yeast gene set, because the analysis examines all ORFs ≥ 100 amino acids, the methodology has high sensitivity and specificity and the evidence is unambiguous for the vast majority of ORFs. Nonetheless, some errors are likely to remain. The results could be confirmed and remaining uncertainties resolved by sequencing of additional related yeast species, as well as by other experimental methods.
Despite the intensive study of S. cerevisiae to date, comparative genome analysis points to the need for a major revision of the yeast gene catalog affecting more than 15% of all ORFs (Figure 2.9). The results suggest that comparative analysis of a modest collection of species can permit accurate definition of genes and their structure. Comparative analysis can complement the primary sequence of a species and provide general rules for gene discovery that do not rely solely on known splicing signals for gene discovery. Previous studies have shown that such methods are also applicable to the understanding of mammalian genes42. The ability to observe the evolutionary pressures that nucleotide sequences are subjected to radically changes our power for signal discovery.