Detection of Functional Overlapping Genes: Simulation and Case Studies

Abstract

As far as protein-coding genes are concerned, there is a non-zero probability that at least one of the five possible overlapping sequences of any gene will contain an open-reading frame (ORF) of a length that may be suitable for coding a functional protein. It is, however, very difficult to determine whether or not such an ORF is functional. Recently, we proposed a method that predicts functionality of an overlapping ORF if it can be shown that it has been subject to purifying selection during its evolution. Here, we use simulation to test this method under several conditions and compare it with the method of Firth and Brown. We found that under most conditions, our method detects functional overlapping genes with higher sensitivity than Firth and Brown’s method, while maintaining high specificity. Further, we tested the hypothesis that the two aminoacyl tRNA synthetase classes have originated from a pair of overlapping genes. A central piece of evidence ostensibly supporting this hypothesis is the assertion that an overlapping ORF of a heat-shock protein-70 gene, which exhibits some similarity to class 2 aminoacyl tRNA synthetases, is functional. We found signature of purifying selection only in highly divergent sequences, suggesting that the method yields false-positives in high sequence divergence and that the overlapping ORF is not a functional gene. Finally, we examined three cases of overlap in the human genome. We find varying signatures of purifying selection acting on these overlaps, raising the possibility that two of the overlapping genes may not be functional.

Introduction

Methods for prediction of protein-coding genes use three properties: (1) the presence of open-reading frame (ORF), (2) expression of mRNA, and (3) evolutionary conservation. These properties, however, are often uninformative in the case of overlapping genes because: (1) intact overlapping ORFs that are nonetheless non-functional are expected to be fairly common, (2) both same-strand and opposite-strand overlapping ORFs may be transcribed regardless of functionality (Lavorgna et al. 2004), and (3) non-functional overlapping ORFs are evolutionary conserved because their sequence is shared with functional genes. As a result, annotation programs often fail to correctly predict functional overlapping genes (Delcher et al. 1999).

Much work has been done in attempts to distinguish functional from spurious overlapping genes. Silke (1997) showed that the frequency of opposite-strand overlapping ORFs in vertebrate genomes is highly influenced by genomic GC content and codon usage, suggesting that a large proportion of these ORFs may be spurious. Nekrutenko and colleagues (Chung et al. 2007; Nekrutenko and He 2006; Nekrutenko et al. 2005; Szklarczyk et al. 2007) investigated overlapping genes in the human genome and introduced a set of methods for functionality prediction based on several overlap properties including overlap length, the probability of substitution to a stop codon, and mutation patterns. Liang and Landweber (2006) predicted that ~7% of human genes contain overlapping alternatively spliced exons, based on similarity to known proteins at the levels of sequence, and secondary and tertiary structures, as well as the existence of motifs known to exist in functional genes. Ribrioux et al. (2008) scanned alternative reading frames in orthologous human–mouse–rat genes for amino-acid conservation, length, repeats, and the probability of mutations to a stop codon to predict several new candidate overlaps. Palleja et al. (2008) examined the conservation of length between overlapping genes in different bacterial species and concluded that many of the long overlapping genes have been misannotated. Xu et al. (2010) combined experimental and computational analyses to explore how ORF length, the position of the first AUG codon, and the Kozak motif affect translation of overlapping transcripts.

Although all of the above studies have used the underlying principle that genomic features of gene overlap would be conserved by purifying selection when functional, none has estimated selection pressure directly. The reason is that estimation of selection intensity is complicated by the sequence interdependence between the two genes (Miyata and Yasunaga 1978). Firth and Brown (2005) were the first to use direct estimations of selection to detect functional overlapping genes. Their method (FB), which is suitable for sequence pairs, calculates several statistics for each particular pairwise sequence alignment and uses a Monte Carlo simulation to determine whether the sequence is single-coding or double-coding. This method was later applied to multiple sequences by choosing neighboring terminal pairs of taxa in a phylogenetic tree (Firth and Brown 2006). With the FB method, novel overlapping genes were discovered in many viral taxa (Chung et al. 2008; Firth 2008; Firth and Atkins 2008a, b,2009). We developed a different method (SLG) for the estimation of selection intensities in overlapping genes (Sabath et al. 2008b). SLG uses a maximum-likelihood framework to fit a Markov model of codon substitution to data from two aligned orthologous overlapping sequences. Later, we used SLG to detect signature of purifying selection in unannotated overlapping ORFs (Sabath et al. 2009). In order to detect a signature of purifying selection, we estimate the likelihood of two hierarchical models. In model 1, there is no selection on the ORF. In model 2, the ORF is assumed to be under selection. The likelihood-ratio test is used to test whether model 2 fits the data significantly better than model 1, in which case, the ORF is predicted to be under selection and most probably functional. With this method, we predicted the existence of a new functional overlapping gene in the genomes of four viruses within the Dicistroviridae family (Sabath et al. 2009). This prediction was later supported by Firth et al. (2009). Other methods that make use of selection signatures to detect functional overlapping genes were proposed (de Groot et al. 2007, 2008; McCauley et al. 2007), but these methods have seldom been used, most probably due to the lack of accessible implementation.

Here, we use simulation to test and compare the FB and SLG methods under several conditions. We discuss the influence of overlap type, selection pressure, sequence divergence, and sequence length on the performance of the methods. We also examine in detail a purported case of overlap between the gene encoding heat-shock protein-70 (HSP70) and an opposite-strand ORF (OS-ORF) (Carter and Duax 2002; Konstantopoulou et al. 1995; Monnerjahn et al. 2000; Rother et al. 1997; Silke 1997). This overlap was deemed the “Rosetta stone” for the origin of the two aminoacyl tRNA synthetase (aaRS) classes from opposite-strand overlapping genes, based on a perceived similarity between OS-ORF and class 1 aaRS, on the one hand, and between HSP70 and class 2 aaRS, on the other hand (Carter and Duax 2002; Rodin and Ohno 1995) (Fig. 1). Recently, the functionality of OS-ORF was questioned by Williams et al. (2009) most notably on the basis of the patchy phylogenetic distribution of OS-ORF and an observed association between the existence of intact OS-ORF sequences and genomic GC content. Finally, we examine three cases of predicted overlaps within the INK4a (Quelle et al. 1995), XBP1 (Yoshida et al. 2001), and GNAS1 (Klemke et al. 2001) genes in the human genome. In XBP1, the overlap is between two alternatively spliced exons whereas the overlaps in INK4a and GNAS1 are thought to produce distinct proteins. Although these genes were reported to have complex regulatory properties (Chung et al. 2007), questions regarding their functionality remain (Nekrutenko and He 2006; Szklarczyk et al. 2007).

Fig. 1.

The “Rosetta stone” hypothesis. The two aminoacyl tRNA synthetase (aaRS) classes were proposed to originate from opposite-strand overlapping genes, based on a perceived similarity between OS-ORF and class 1 aaRS, on the one hand, and between HSP70 and class 2 aaRS, on the other hand (Carter and Duax 2002; Rodin and Ohno 1995, images from Carter and Duax 2002)

Methods

We simulated the evolution of overlapping genes as in Sabath et al. (2008b). In each run of the simulation, one gene was designated as known and the second as hypothetical. We examined the effects of the following factors on the ability of the two methods to detect selection in the hypothetical gene: (1) nonsynonymous/synonymous rate ratios in the hypothetical gene and the known gene (ω_h and ω_k, respectively), (2) overlap types (same-strand (SS) phases 1 and 2 and opposite-strand (OS) phases 0, 1, and 2), (3) sequence divergence (t), (4) sequence length, and (5) GC content. For each combination of overlap type, sequence divergence, and sequence length, we set ω_k and vary ω_h between 0.2 (strong purifying selection) and 1 (no selection). For each set of parameters, we generated 100 random pairs of overlapping genes. In order to test the influence of GC content, we generated random sequences with 30, 50, and 70% GC content and simulated their evolution with standard parameters (same-strand phase 1, length of 300 codons, t = 0.4, ω_k set to 0.2, and ω_h varied between 0.2 and 1). Sensitivity is defined as the percent of hypothetical genes under selection that were identified correctly by the method. Specificity is defined as the fraction of hypothetical genes that were incorrectly identified to be under selection when ω_h was set to 1 (i.e., no selection).

HSP70–OS-ORF Overlap

In addition to the 29 bacterial HSP70 genes with an intact OS-ORF in Williams et al. (2009), we acquired 16 new sequences through BLASTing (Altschul et al. 1990) the nr database (limited to bacterial sequences) with the original sequences as queries. We constructed a multiple alignment of the sequences at the amino-acid level with CLUSTAW (Thompson et al. 2002) as implemented in the MEGA package (Kumar et al. 2008). Sites with gaps were removed resulting in a multiple-sequence alignment 561 codons in length. The multiple alignment is provided in the Supplementary Information. We then tested for selection on the OS-ORFs in all homologous pairs with OS-ORF sequence divergence of 50% or less.

Overlapping ORFs in the Human Genome

We identified the sequence of mouse INK4a (Accession: NM_001040654), based on Quelle et al. (1995) and obtained the orthologous gene sequences of 12 other species (human, rat, chimpanzee, orangutan, marmoset, macaque, cat, dog, cow, opossum, horse, and rabbit) through the UCSC Genome Browser (Karolchik et al. 2004). For the GNAS1 overlap, we used eight sequences from human, chimpanzee, orangutan, gorilla, macaque, gibbon, colobus, and squirrel monkey (Nekrutenko et al. 2005). In addition, we obtained the orthologous mouse GNAS1 CDS (accession AY519504) through a Blast search (Altschul et al. 1990). For the XBP1 overlap, we obtained eight sequences from human, mouse, rat, cow, frog, chicken, zebrafish, and tilapia (Nekrutenko and He 2006) as well as three additional sequence from orangutan, gorilla, and pufferfish) through a Blast search (Altschul et al. 1990).

We constructed multiple alignments of each gene at the amino-acid level (see, Supplementary Information) with CLUSTAW (Thompson et al. 2002) as implemented in the MEGA package (Kumar et al. 2008), and removed sites with gaps. The length of the multiple alignment after removal of gaps was noted. GC content was measured as the mean across all sequences in each alignment. For each overlap, we calculated the sequence identity between human and mouse genes, at both the nucleotide and aminoacid levels, as a reference for comparisons among the three cases. Finally, we tested for selection on all pairs with sequence divergence of 50% or less. Runs in which the program did not converge (usually in pairs of highly similar sequences) were excluded from further analysis.

Results and Discussion

Simulation

We initially set the sequence length to 300 codons and t = 0.4, which corresponds to a sequence divergence of ~12%. We set ω_k to 0.2 and varied ω_h between 0.2 and 1. The results are shown in Fig. 2a. Each square presents the results for SLG (solid blue: P < 0.01; dashed blue: P<0.05) and FB (red) methods against ω_h (X-axis). An ideal detector is exemplified by a dashed green line. Each data point is the percentage of runs in which the methods detected selection. The five overlap types are shown in each column. As expected, the sensitivity of both methods decreased with increase in ω_h. In all overlap types, SLG exhibits a higher sensitivity than FB, up to ~80% in same-strand for ω_h = 0.4. As expected, using SLG with P value of 0.05 (rather than 0.01), increase the method’s sensitivity at the cost of lower specificity. For opposite-strand phase 2, both methods perform similarly. This phase is unique in that the third codon position of both genes corresponds, and thus, most changes are either nonsynonymous in both genes, or synonymous in both (Sabath et al. 2008b). This overlap phase was also reported to generate high rate of false-positive results (Firth and Brown 2006).

Fig. 2.

Detection of selection by the SLG (solid blue: P < 0.01; dashed blue: P < 0.05) and FB (red) methods on simulated genes. An ideal detector is illustrated by a dashed green line. Each data point is the percentage of runs for which the methods detected selection. The five overlap types are shown in each column. Four sets of values for sequence divergence (t), ω_k, and sequence length are shown in each row (see text)

In the next three sets, we tested different values of t, ω_k, and sequence lengths, one parameter at a time. In Fig. 2b, we present the performance of the methods at high sequence divergence levels (t = 1, corresponding to a sequence divergence of ~24%). For both methods, the results are similar to those at low sequence divergence. In Fig. 2c, we present the results for stronger selection level on the known gene (ω_k = 0.1). The performance of SLG is similar to that in (a) and (b), whereas the sensitivity of FB is reduced in same-strand phases 1 and 2 and opposite-strand phases 0 and 1. In Fig. 2d, we present the results for short sequence length (60 codons). Under these conditions, SLG and FB perform similarly, with SLG showing reduced sensitivity compared to (a), (b), and (c). Finally, we tested whether the methods are influenced by GC content. We tested sequences with 30, 50, and 70% GC content and simulated their evolution with standard parameters (see Methods). We find that both methods are highly robust to differences in GC content (Supplementary Fig. S1).

Overall, the simulation demonstrates that, under most conditions, SLG performance is as good as FB or higher. The advantage of using SLG over FB is that it performs better when the known gene is under strong purifying selection, whereas both methods perform alike on short sequences. In addition, SLG was found to be more robust to overlap type in comparison to FB, whose performance is more variable, especially in the case of opposite-strand phase 2 overlaps. Similarly to FB, SLG can be applied to multiple sequences by choosing only neighboring terminal taxa in the phylogenetic tree (Firth and Brown 2006). This approach, while ingenious, only indirectly addresses the phylogenetic framework and may be biased for trees with non-uniform branch-length distribution. In future studies, it would be beneficial to take full advantage of the maximum-likelihood framework that allows testing hypotheses concerning variable selection pressures among lineages and sites (Nielsen and Yang 1998; Zhang et al. 2005). This might be of special significance for overlapping genes because they may exist as a non-functional ORFs before they become functional (Keese and Gibbs 1992).

Testing the Functionality of Functionality of OS-ORF

The functionality of HSP70 and its overlapping OS-ORF constitute a central tenet of the hypothesis concerning the origin of the two aaRS classes (Carter and Duax 2002; Rodin and Ohno 1995). A recent study by Williams et al. (2009) cast doubt on the functionality of OS-ORF. We used SLG to determine whether or not selection operates on OS-ORF. We identified 45 bacterial HSP70 genes with an intact OS-ORF and tested for selection on the OS-ORFs in all homologous pairs with OS-ORF divergence of up to 50%. The results are shown in Fig. 3. For each pair, the amino-acid sequence divergence of the OS-ORFs was plotted against that of HSP70. Pairs for which the method did not detect selection (at P values larger then 0.01) are marked in blue, and pairs for which a signature of selection was found are marked in red. First, in all pairs the divergence between the OS-ORF sequences is higher than the divergence of HSP70 (up to 30% higher), suggesting that HSP70 is under much stronger selection pressure. Second, pairs for which a signature of selection was found are concentrated in the upper divergence range of OS-ORF. The detection of selection signatures only in highly divergent pairs suggests that these are false-positive results and that OS-ORF is not a functional gene.

Fig. 3.

Testing for selection on OS-ORF. The amino-acid sequence divergence of the OS-ORF is plotted against that of HSP70 for all pairs of homologous sequences up to 50% divergence of the OS-ORF. (Red) Sequence pairs in which the method detected selection (P < 0.01). (Blue) Sequence pairs in which the method did not detect selection

The most likely reason for inaccuracy at high sequence divergence values is that the method estimates the probability of one codon to change to another by summing over all possible paths. With the increase in divergence, the number of possible paths raises and, consequently, the power of the method to recover to true path decreases. In addition, the reduced quality in alignment of distant sequences could also contribute to biased inference of selection (Schneider et al. 2009). These results imply that ~30% divergence between sequences should be the upper boundary for using the SLG method. This boundary is comparable to the boundary of 25% divergence suggested for exon detection using a single-coding genes analogous method (Nekrutenko et al. 2002).

Overlapping ORFs in the Human Genome

We chose INK4a, XBP1, and GNAS1 because they are considered to be the best examples of overlaps in the human genome (Chung et al. 2007). Since the aim of this study is to evaluate our method, we thought to concentrate on sequence-based evidence and not to review the full body of experimental evidence for and against the functionality of each gene. We, therefore, do not make any a priori assumption about the functionality of any of the genes and test for selection on both overlapping ORFs. The results (Fig. 4) are shown in the same format as in Fig. 3 where the amino-acid sequence divergence of an ORF was plotted against that of its overlapping ORF. Pairs for which the method did not detect selection (At P values larger then 0.01) are marked in blue, and pairs for which a signature of selection was found are marked in red. In each plot, the ORF under examination is highlighted in green. For example, in Fig. 4a, the top plot presents the results of testing ARF and the bottom plot presents the results of testing INK4a. In order to complement our method and to allow comparison among the three overlaps, we also note three benchmark genomic properties, GC content and sequence identity at both the nucleotide and amino-acid levels. The average GC content of human genes is 52%, and the median identity between human and mouse coding sequences is 85% at the DNA level and 78.5% at the amino-acid level (Waterston et al. 2002). High GC content, which could be the result of several mutational processes (Graur and Li 2000), leads to reduction in the frequencies of stop codons (which are AT rich) in overlapping ORFs and, hence, contributes to the conservation of overlapping ORFs even in the absence of selection (Sabath et al. 2008a; Silke 1997; Williams et al. 2009). In Table 1, we note: (1) overlap length after excluding gap-containing sites, (2) GC content, (3) percent identity between human and mouse genes at the DNA level, (4) percent identity between human and mouse genes at the amino-acid level, (5) number of pairs under 30% divergence, in which the method detected selection (P < 0.01), (6) total number of pairs under 30% divergence, and (7) percentage of positives out of the total number of pairs.

Fig. 4.

Testing for selection on INK4a/ARF (a), XBP1 (b), and GNAS1/ALEX (c). The amino-acid sequence divergence of each ORF is plotted against that of its overlapping ORF for all pairs. In each plot, the ORF, which is tested for selection is highlighted in green. For example, in a, the top plot presents the results of testing ARF and the bottom plot presents the results of testing INK4a. (Red dots) Sequence pairs in which the method detected selection (P < 0.01). (Blue dots) Sequence pairs in which the method did not detect selection

Table 1.

Testing for selection on overlaps within three human genes

Locus	Length (codons)	GC (%)	H-M NT (%)a	H-M AA (%)b	ORF	Positivec	Totald	%e
INK4a/ARF	64	74	77.6	50.8	ARF	1	40	2.5
				76.6	INK4a	57	71	80.3
XBP1	85	53	92.2	91.7	XBP1U	5	13	38.5
				82.1	XBP1S	1	12	8.3
GNAS1/ALEX	314	69	68.3	47.2	ALEX	0	26	0
				46.4	GNAS1 XLαs	2	26	7.7

^aPercent identity between human and mouse genes at the DNA level

^bPercent identity between human and mouse genes at the amino-acid level

^cNumber of pairs under 30% divergence, in which the method detected selection (P < 0.01)

^dTotal number of pairs under 30% divergence

^ePercentage of positives out of the total number of pairs

INK4a/ARF Overlap

The INK4a/ARF overlapping genes are known as regulators of tumor suppression pathways (reviewed in Szklarczyk et al. 2007). Using our method, we find considerable evidence of selection (80.3% of the pairs, Fig. 4a, bottom plot) acting on INK4a, whereas ARF exhibits very weak signature of purifying selection (2.5%, Fig. 4a, top plot). All sequence pairs exhibit equal or higher amino-acid divergence at the ARF frame than at the INK4a frame as seen by all of the points being located either on the diagonal or above it (Fig. 4a), similarly to the HSP70/OS-ORF overlap (Fig. 3). The overlap region has a very high GC content of 74% (Table 1). The human-mouse amino-acid sequence identity of INK4a is average (76.6%), while that of ARF is considerably low (50.8%). It is possible that our method failed to detect selection on ARF because of the short length of the region. However, together with the amino-acid divergence and high GC content, these results imply that the overlapping region of INK4a is functional whereas that of ARF is most likely spurious. This conclusion is in agreement with the absence of this overlap in chicken (Kim et al. 2003) and the lack of apparent function of ARF in other species. Arguing for functionality of ARF, Szklarczyk et al. (2007) have simulated random sequences corresponding to the amino-acid sequence of INK4a using the codon usage of all human genes. They found that only 4% of the simulated sequences contained full ORF in the ARF frame, suggesting that selection must be preserving the ORF. However, local deviation of codon usage (as reflected by the high GC content in the INK4a/ARF overlap) could be the result of both selection and mutation. Therefore, it is possible that the absence of stop codons disturbing the ARF sequence is simply the result of mutational bias rather than selection acting to conserve its sequence.

XBP1^U/XBP1^S overlap

XBP1 protein is associated with regulation of the expression of genes implicated in the unfolded protein response (Nekrutenko and He 2006). We find weak but significant evidence for selection (38.5% of the pairs) acting on XBP1^U, and a weaker signal (8.3%) on XBP1^S. Unlike HSP70/OS-ORF and INK4a/ARF overlaps, the divergence of sequence pairs do not have a distinct pattern and the points are scattered on both sides of the diagonal (Fig. 4b). The overlap region has a moderate GC content of 53% (Table 1), suggesting low mutational bias. The human–mouse amino-acid sequence identity of both ORFs is higher than average (91.7 and 82.1% for XBP1^U and XBP1^S, respectively), suggesting that the region is under strong purifying selection. Based on these results, we predict that XBP1^U is functional with a high degree of confidence. The functionality of XBP1^S is not so well supported by our method. However, given the short length of the overlap (which makes the detection of selection difficult), the phylogenetic-wide conservation of this ORF (Nekrutenko and He 2006), and the moderate GC content of this region, functionality cannot be ruled out.

GNAS1/ALEX Overlap

The GNAS1 locus encodes the alpha subunit of the stimulatory G protein of adenylyl cyclase (Kozasa et al. 1988; Levine et al. 1991). This protein is a member of the G proteins, a family of signal-coupling proteins that mediate numerous transmembrane hormonal and sensory transduction processes (Bourne et al. 1990). This gene is also known to have complex imprinting patterns (Hayward et al. 1998). We focused on the GNAS1 XLαs exon and its overlapping ORF, ALEX. We find very little evidence of selection (7.7% of the pairs) acting on GNAS1 XLαs exon, and no signal (0%) of selection on ALEX. The overlap region has a high GC content of 69% (Table 1). The human–mouse amino-acid sequence identity of both ORFs is lower than average (46.4 and 47.2% for GNAS1 XLαs and ALEX, respectively). Surprisingly, the sequence identity at the nucleotide level (68.3%) is much higher than the amino-acid sequence identity of both ORFs. This difference between the identities at the nucleotide and amino-acid levels (>20%) is unique to the GNAS1 overlap, whereas in the other two overlaps the minimum difference is ~1% (Table 1). The low amino-acid conservation of the GNAS1 XLαs exon is in striking contrast to the rest of the protein, which is nearly identical between the two species (Supplementary Fig. S2). Unlike INK4a/ARF and XBP1 overlaps, the XLαs/ALEX overlap is 314 long, a length at which the sensitivity of our method is high (Fig. 2). Based on these results, we do not find strong support for the functionality of neither XLαs nor ALEX. The only observation that remains in support of both overlapping ORFs being functional is that none of the multiple insertions and deletions in this region disturb the reading frame (Supplementary Fig. S2; Nekrutenko et al. 2005). In order to evaluate the significance of this observation, it would be necessary to apply an evolutionary model of insertions and deletions.

Nekrutenko et al. (2005) interpreted the high, but relatively similar, evolutionary rate of both ORFs as “constant adjustment between the two proteins aimed at maintaining mutual affinity.” Furthermore, Nekrutenko et al. (2005) suggested a model, in which purifying selection is acting on the two proteins to maintain both ORFs and their expression coupling. Using our method, we show that there is no strong purifying selection acting on these proteins (only a weak selection on XLαs exon if any). Once we exclude the scenario of purifying selection acting on both proteins, their equal rate of evolution could simply be explained by drift, which does not distinguish between two non-functional ORFs.

We studied the performance of our new method for the detection of functional overlapping genes. By simulation, we compared the method to the FB method and tested both methods across different overlap types. We found that under most conditions, our method predicts functionality with higher sensitivity while maintaining high specificity. Finally, we examined the overlaps within HSP-70, INK4a, XBP1, and GNAS1 genes. We conclude that long overlaps in non-viral genomes should be treated with caution since many of them may be spurious.