Putting Pleiotropy and Selection into Context Defines a New Paradigm for Interpreting Genetic Data

Abstract

Background

Natural selection shapes many human genes, including some related to complex diseases. Understanding how selection affects genes, especially pleiotropic ones, may be important in evaluating disease associations and the role played by environmental variation. This may be of particular interest for genes with antagonistic roles that cause divergent patterns of selection. The lectin like low-density lipoprotein 1 receptor (LOX-1), encoded by OLR1, is exemplary. It has antagonistic functions in the cardiovascular and immune systems as the same protein domain binds oxidized LDL and bacterial cell wall proteins - the former contributing to atherosclerosis, the latter presumably protecting from infection. We studied patterns of selection in this gene, in humans and non-human primates, to determine whether variable selection can lead to conflicting results in CVD association studies.

Methods and Results

We analyzed sequences from 11 non-human primate species as well as SNP and sequence data from multiple human populations. Results indicate that the derived allele is favored across primate lineages (probably due to recent positive selection). However, both the derived and ancestral alleles were maintained in human populations, especially European ones (possibly due to balancing selection derived from LOX-1's dual roles). Balancing selection likely reflects response to diverse environmental pressures among humans.

Conclusions

These data indicate that differential selection patterns, within and between species, in OLR1 render association studies difficult to replicate even if the gene is etiologically connected to CVD. Selection analyses can identify genes exhibiting gene-environment interactions critical for unraveling disease association.

Introduction

Genes that associate with complex disease are likely to be shaped by natural selection. However, the underlying pattern of selection may be confounded by multiple functions of a single gene, or pleiotropy¹. Specifically, if a gene has two functions and a variant improves function for one phenotype but has the opposite effect on the second, a phenomenon known as antagonistic pleiotropy, it will influence our ability to detect associations in a context-dependent fashion, potentially masking important biology. One way to explore this phenomenon is to compare inter-specific and intra-specific patterns of genetic variation in genes with multiple known functions and to assess patterns of selection.

We tested this hypothesis in the lectin like oxidized low density lipoprotein receptor 1 gene (OLR1, OMIM 602601), which encodes LOX-1, thought to increase cardiovascular disease risk by acting as a scavenger receptor for ox-LDL in endothelial cells². LOX-1 also binds other ligands, including bacterial cell wall proteins, thus playing an important protective role in immune function³. These two biological functions likely operate independently and antagonistically. Although it is impossible to know with certainty what the dominant function(s) was and is for OLR1, the immune function role is likely the ancestral one since ancient humans had a short life span and were not likely exposed to the selection pressure of atherosclerosis-based cardiovascular disease.

Regarding the possible functional role of genetic variation in OLR1, studies have demonstrated that a haplotype containing single nucleotide polymorphisms (SNP) in and around intron 4 causes alternative splicing of exon 5 and leads to increased synthesis of a truncated isoform known as LOXIN^{4, 5}. The SNPs in this haplotype (rs3736232, rs3736234, rs3736233, rs3736235, rs3816844 and rs1050283) are in strong linkage disequilibrium (LD) with each other in all populations, and the region is therefore often studied using only one representative SNP⁴. Since exon 5 encodes a C-type lectin-binding domain (CTLD), crucial for LOX-1 to bind its ligands, LOXIN is dysfunctional (Fig. 1)^5-7. Despite compelling experimental data indicating loss of function due to this variation⁵, studies investigating association with disease have often yielded inconsistent and even contradictory results (Supplemental Table 1)⁸. We hypothesize that this is because LOX-1 is favored in some environments while LOXIN is favored in others. Variation in the relative importance of the two functions and the subsequent divergent selection may confound results of studies addressing association with CVD. For example, validation of epidemiological results using different populations or different endpoints might lead to contrasting results if the two functions of the gene are both involved in the pathogenesis of the disease, as is the case for LOX-1 and LOXIN in atherosclerosis. In addition, since both immunity and lipid oxidation are important players in atherogenesis, it is reasonable to expect classical genetic association studies to be influenced by this effect. To determine if this is likely, we performed analyses of OLR1 with respect to its evolutionary history both inter- and intra-species.

Figure 1.

A) Structure of OLR1 with location of variants (in order from left to right: rs3736232, rs3736234, rs3736233, rs3736235, rs3816844 and rs1050283, r² > 0.95) that modulate the ratio between the LOX-1 and LOXIN mRNA. Bracket identifies region of strong LD.

Results

Evidence for ancient positive selection in intron 4 of the OLR1 locus

To examine if and how ancient selection has shaped the OLR1 locus, we conducted two selection tests on coding and non-coding regions in primate lineages. First, we evaluated the ζ ratio of non-synonymous to synonymous substitution rate in coding regions and the ω ratio of non-coding to synonymous substitution rate for the intron 4 region across the 12 primate species (human plus 11 non-human). We found strong evidence of selection in intron 4 (p = 0.001), but not in the coding sequence of the OLR1 locus (p = 0.519) (Table 1). Furthermore, using either the Naive Empirical Bayes algorithm (which uses maximum likelihood estimates of model parameters to identify sites undergoing positive selection) or the Bayes Empirical Bayes algorithm (which assigns priors for model parameters and estimates averages by numerical integration over these priors to identify sites under positive selection) ⁹ we identified three sites (sites 112, 207 and 224 from the beginning of the amplicon) with posterior probability of positive selection ≥92% (Supplemental Table 2).

Table 1. Results of ancient (interspecific) selection analyses on coding (ω) or non-coding (ζ) regions of OLR1.

ω / ζ Ratio Tests	Coding Region				Intron 4
Model (Data)	M1a		M2a		ncM1a		ncM2a
	Frequency	ω	Frequency	ω	Frequency	ζ	Frequency	ζ
Class 0 (0 < ω < 1 or 0 < ζ < 1)	0.427	0.000	0.450	0.024	0.988	0.379	0.843	0.379
Class 1 (ω=1 or ζ = 1)	0.573	1.000	0.503	1.000	0.014	1.000	0.012	1.000
Class 2 (ω> 1 or ζ > 1)	n/a	n/a	0.047	3.229	n/a	n/a	0.147	4.748
Ln Likelihood	-1770.666		-1770.009		-2796.810		-2787.700
Parameters	2.000				4.000
Likelihood Ratio Test	1.313				18.214
p-value	0.519				0.001

Evidence for intra-species balancing selection in intron 4 within humans and for intra-species positive selection within chimps

In humans, especially the HapMap Caucasian European (CEU) population, balancing selection is likely to be acting on OLR1 as indicated by high Tajima's D from HapMap 2 data (D = 3.54, p < 0.001, Figure 2, Table 2). When we divided the gene into two distinct regions defined by their LD patterns: 1) the region of high LD between rs12822177 and rs11615002 that contains intron 4 and is marked in Figure 1 as including the CTLD and 2) the rest of the gene, we observed a higher Tajima's D in the LD block region in Europeans (D_LD = 3.49; D_Non-LD = 2.88, Figure 2a; Table 2a). Similarly, F_st for the region of LD was higher than for the rest of the gene (F_st-LD = 0.27, F_{st-non LD} = 0.080), supporting differential selection patterns in different parts of the gene (Table 2a). Similar results were found for SNPs in the LD block derived from 1000 Genomes samples of majority European ancestry, indicating that for these continental populations common and rare SNPs share patterns of selection (Supplemental Table 3). In general, the European data provide the strongest evidence for selection in OLR1.

Figure 2.

Representation of LD of SNPs (r2) in HapMap A) CEU, B) CHB+JPT, C) YRI and D) chimps from the present study. Tajima's D values are shown by the colored bars above the exons (if D ≥ 3 then balancing selection, p <0.001; if 2 < D > 3 then balancing selection, p < 0.01; if D = 2 then balancing selection, p < 0.05; if D < - 2 then positive selection, p <0.001; if D = 2 then positive selection, p < 0.05; if -2 < D > 1.8 then neutral). Level of significance for each region is also indicated above the exons by *, p <0.05; **, p < 0.01; and ***, p <0.001 and darker shades represent stronger selection signals. The region that demarks the C-type Lectin Domain (CTLD) also demarks what we refer to as the LD block (between SNPs rs12822177 and rs11615002). Markers: 1 = rs2742110, 2 = rs11053654, 3 = rs2634162, 4 = rs2742113, 5 = rs2742114, 6 = rs2742115, 7 = rs16910917, 8 = rs3741860, 9 = rs3912640, 10 = rs11611453, 11 = rs11611438, 12 = rs2010655, 13 = rs11053649, 14 = rs6488265, 15 = rs11053648, 16 = rs12822177, 17 = rs11053646, 18 = rs3736232, 19 = rs3736233, 20 = rs3736234, 21 = 3736235, 22 = rs3816844, 23 = rs2634156, 24 = rs12309394, 25 = rs1050283, 26 = rs10505755, 27 = rs1050286, 28 = rs1050289 (A, B, C). Polymorphisms in chimpanzees have no rs numbers.

Table 2.

Intra-specific selection analyses in humans (a) and chimpanzees (b) by region of the gene OLR1. For human populations also population stratification indexes are reported (* p < 0.05, ** p < 0.02, *** p < 0.001, LD block refers to the SNPs between rs12822177 and rs11615002, non LD block refers to all the SNPs in OLR1 that are out of that region).

a.
Species	Population/region	Tajima's D	Fu and Li		Population differentiation (FST)
			D	F
Homo sapiens	CEU − overall	3.54***	1.84**	3.00**	0.117***
Homo sapiens	YRI − overall	1.77	1.40	1.84*
Homo sapiens	CHB + JPT − overall	3.50***	1.83**	3.01**
Homo sapiens	CEU − LD block	3.49***	1.41	1.84*	0.270***
Homo sapiens	YRI − LD block	0.89	0.81	1.01
Homo sapiens	CHB + JPT − LD block	2.37*	1.34	2.05**
Homo sapiens	CEU − non LD block	2.88**	1.58*	1.58*	0.080***
Homo sapiens	YRI − non LD block	2.30*	1.41	2.06*
Homo sapiens	CHB + JPT − non LD block	3.70***	1.45	2.76**
b.
Pan troglodytes	5′UTR + Exon 1	-2.381**	-3.60**	-3.77**	---
Pan troglodytes	Intron 4	-1.947*	-2.19	-2.45	---
Pan troglodytes	Exon 5	-1.012	1.60	1.32	---
Pan troglodytes	Exon 6 + 3′UTR	-1.821*	-2.58*	-2.74*	---

In contrast to the pattern of selection found among humans, several OLR1 gene regions in P. troglodytes (chimpanzee) showed evidence for intra-specific positive selection. Specifically, negative Tajima's D values were found for the 5′UTR region, intron 4 and 3′UTR (D = - 2.381, -1.947, and -1.821 respectively, Figure 2d), indicating intra-specific positive selection in these regions. In contrast, the other regions of the gene did not show evidence of positive selection with D values ranging from -1.800 to 0.000 (Table 2b). Taken together, the data from humans and chimpanzees indicate different types of selective pressure (balancing vs. positive) acting on a segment that includes intron 4 in both humans and chimpanzees.

The evolutionary history of the OLR1 locus is consistent with the species phylogeny

The action of selection on a locus can cause its phylogeny to deviate significantly from that of the species^{10, 11}. To test whether this was the case for the OLR1 locus, we reconstructed the evolutionary history of the locus using both coding and non-coding data (Supplemental Figure 1). Although the locus history is largely in accord with the species phylogeny, sequences from species with multiple individuals often fail to form monophyletic groupings, with some of the alleles representing trans-species polymorphisms (e.g., alleles in humans that are more similar to alleles in chimpanzees than they are to alleles derived from other humans¹²). Although only balancing selection can maintain trans-species polymorphisms long-term, transient trans-species polymorphisms could also be consistent with directional selection or neutrality¹⁰. To test this possibility, we compared whether the maximum likelihood estimated topologies, i.e. the branching patterns of phylogenetic trees, from the entire set of coding and non-coding sequences were significantly different from topologies constrained to obey the primate phylogeny and monophyly of alleles from each species, using the Shimodaira-Hasegawa test ¹³. The maximum likelihood topologies were not significantly different from the species phylogeny (both tests were not significant at p = 0.05), indicating that trans-specific OLR1 polymorphisms do not exist in primates.

Discussion

Analyses of primate sequences of OLR1 indicate complex patterns of selection based on species and the part of the gene being assessed. Our results support the hypothesis that intron 4 and its surrounding region have been shaped by differential selection in human vs. non-human primate lineages. To summarize, while analyses of the coding region indicate selective neutrality, we found strong evidence of ancient positive selection in the intron 4 region in primates (Table 1). Comparable analyses of 18 chimpanzee sequences also indicated intra-specific positive selection (Figure 2d, Table 2b), whereas in humans (especially Europeans) there was strong evidence for balancing selection in intron 4 (Figure 2a,b,c Table 2a). Hence, the mode of recent selection varies between these very closely related species.

Differential patterns of selection in OLR1 can be interpreted in light of how the LOX-1 and LOXIN proteins function and their impact on phenotypic variation. For LOX-1 to bind its ligands, it must contain the sequence encoded in exon 5. The ability to bind ligands, however, has been significantly associated with several predominantly human diseases (cardiovascular disease, hypertension, Alzheimer's; see Supplemental Table 1). In contrast to its presumed negative effect on cardiovascular disease, LOX-1 function is necessary for binding to bacteria and to trigger the subsequent response to infection³.

Antagonistic pleiotropy acting on OLR1 is consistent with the pattern of balancing selection in certain human populations. Balancing selection, which maintains alleles in a polymorphic state, may in the case of OLR1 be due to both the direction and strength of selection varying with immediate environments. For example, for a late-onset disease such as CAD, the selection pressure may be weaker than that from infection in certain populations (e.g., humans in tropical regions), whereas it may be stronger in humans in non-tropical climates such as Europe. In contrast, selection pressure for late-onset non-communicable disease among non-human primates is likely low, and therefore locus variation in these has more likely been shaped by positive selection in response to early onset diseases such as those due to infectious agents. A possible explanation for the observation of balancing selection in OLR1 is heterozygote advantage. However, our data are insufficient to directly address this possibility.

This and previous physiological and immunological studies support a likely role for balancing selection in humans, but the case in non-human primates is less obvious. Clearly, our interspecific analyses provide support for ancestral positive selection, but functional data in non-human primates is lacking. Therefore, it is impossible to infer exactly what the biological role(s) of these variants is in non-human primates. Nonetheless, the analyses clearly provide support to the idea that the polymorphisms in intron 4 confer differences, based on variable selective effects between species. However, the exact variants cannot be determined as they are not directly functionally comparable between humans and primates.

Our data provides a unique perspective on the role of selection in a gene that has multiple and putatively antagonistic roles, indicating that future studies examining the phenotypic role of OLR1 might best be focused on the region where we found evidence of selection. In addition, we argue that association studies with OLR1 will need to adjust for ethnicity, since the patterns of selection vary by population. Pleiotropy leading to a complex signature of selection may play out differently depending on the environmental context, and may provide a paradigm for genes, such as OLR1, that confer risk for multiple diseases.

In summary, the results suggest that antagonistic pleiotropy may obscure genetic associations between CAD and OLR1, a gene whose multiple functions influence diverse, antagonistic, or competing biological processes. For example, atherogenesis is affected by both immunity and ox-LDL internalization. In addition, the two variants, LOX-1 and LOXIN, affect at least one of these processes antagonistically. Nonetheless, our data is consistent with the concept that evolution shapes genetic structure differently in relatively short-lived human populations with a high pathogen load (e.g. developing world populations), where there is likely to be strong selection for the variant that increases the proportion of LOX1 product, as compared to populations in the developed world with access to quality care and nutrition. In the latter environment infection may not provide as large a selective pressure, allowing LOXIN to increase in frequency. In both cases, changing patterns of allelic variation may over the long term reduce power to detect phenotype association in some, but not all populations. Antagonistic roles may also drive balancing selection in some populations, further masking effects on disease, especially when separate phenotypes are being assessed, such as myocardial infarction, stroke, or carotid intima-media thickness (Supplemental Table 1). These complex biological roles may explain why functional studies on LOX-1 have clearly explained its role in the pathogenesis of CAD, whereas epidemiological studies have produced inconsistent and at times opposite outcomes.

Our data showing complex patterns of selection in OLR1 supports the argument that environmentally dependent antagonistic pleiotropy is likely to affect the pattern of genetic variation in this gene and hence association results as such studies are impacted by allelic distributions and patterns of LD. Therefore, assessing patterns of selection can identify genes, such as OLR1, for which knowledge of function is key to defining disease risk, as they may be more sensitive to gene-environment interactions than those with simpler patterns of selection.

Experimental procedures

Samples

A total of 48 samples were collected and sequenced from human and non-human primates, covering a total of 12 species. The three human samples used were de-identified. An additional six reference sequences, one each from H. sapiens, Pan troglodytes, Pongo pygmaeus, Nomascus leucogenys, Macaca mulatta and Callithrix jacchus were downloaded from the Ensembl database. Details regarding the samples are preetned in Supplemental Table 4.

DNA was extracted from blood samples using the phenol-chloroform method¹⁴.

Amplification and sequencing

Primers were designed using conserved regions of the Ensembl 6 primate alignment (http://www.ensembl.org/Homo_sapiens/Gene/Compara_Alignments?align=511&db=core&g=ENSG00000173391&r=12%3A10310902-10324737). At least 50 bp of flanking intronic sequence were included in all exon amplimers. The entirety of intron 4 was amplified because previous data suggested that it included a variant that affected alternative splicing. Primer sequences and reaction conditions for all the primers used are shown in Supplemental Table 5, the amplified regions are shown in Supplemental Figure 2. Amplification reactions were performed with the Taq Platinum according to the manufacturer's protocols (Invitrogen). Each fragment was sequenced in both the forward and reverse direction. Sanger sequencing was done in the Vanderbilt University Genome Resources Core.

Tests for ancient selection on coding and non-coding regions

We tested for ancient selection in both coding and non-coding regions of two different taxon sets. The coding region contained exons 1, 2, 3, 4, 5 and 6. The non-coding region sequence contained intron 4. The first taxon set (primate data set) included a single sequence from each of 12 different primate species, whereas the second taxon set (human-chimp data set) included four human sequences and 17 de novo chimpanzee sequences.

Sequence alignment

All coding sequence alignments were performed in protein space using the MAFFT software, version 6.811 ^{15, 16} and back-translated to codons using the PAL2NAL software ¹⁷, whereas all non-coding sequence alignments were performed in nucleotide space using the MAFFT software, version 6.811 ^{15, 16}.

Test for ancient selection in coding regions

We tested the coding region of the primate data set for evidence of positive selection by estimating the ω ratio of the non-synonymous substitution rate (d_N) to the synonymous substitution rate (d_S) using the CODEML module from the PAML software package, version 4.4 ¹⁸. To do so, we first evaluated the log likelihood of the null M1a model, which allows codon sites to exhibit variable selective pressure but no positive selection. Under M1a, ω values at different codon sites can belong to one of two categories; sites in the first category have ω values that range above 0 but < 1 (and are assumed to be under purifying selection), whereas sites in the second category have ω values equal to 1 (and are assumed to be neutral). We then compared the log likelihood of the M1a model to that estimated by the alternative model M1b, which allows codon sites to exhibit variable selective pressure with positive selection. Under M1b, ω values at different codon sites can belong to one of three categories; sites in the first category have ω values that range between 0 and 1 (and are assumed to be under purifying selection), sites in the second category have ω values equal to 1 (and are assumed to be neutral), whereas sites in the third category have ω values greater than 1 (and are assumed to be products of positive selection.

Test for ancient selection in non-coding regions

We tested the non-coding primate data set for evidence of positive selection by estimating the ζ ratio using the batch file developed by Olivier Fedrigo ¹⁹ for use with the HyPhy software package, version 2.0 ²⁰. The ζ ratio is similar in concept to the ω ratio. To assess whether a non-coding region is evolving under purifying selection, neutrally, or under positive selection, it compares the ratio of the substitution rate of a non-coding region (d_NONCODING) relative to the substitution rate of the synonymous sites from the adjacent coding region (d_S), which are assumed to be evolving neutrally ^{21, 22}. Similar to the coding region tests, we first evaluated the log likelihood of the null model (non-coding M1a or ncM1a), which allows non-coding sites to exhibit variable selective pressure but no positive selection. Under ncM1a, ζ values at different sites can belong to one of two categories; sites in the first category have ζ values that are greater than 0 but range between 0but <1 (and are assumed to be under purifying selection), whereas sites in the second category have ζ values equal to 1 (and are assumed to be neutral). We then compared the log likelihood of the ncM1a model relative to that estimated by the alternative model (non-coding M1b or ncM1b), which allows sites to exhibit variable selective pressure with positive selection. Under ncM1b, ζ values at different sites can belong to one of three categories. Sites in the first category have ζ values that range from 0 to 1 (and are assumed to be under purifying selection), sites in the second category have ζ values equal to 1 (and are assumed to be neutral), whereas sites in the third category have ζ values greater than 1 (and are assumed to be undergoing positive selection).

Tests for intra-specific selection

Genotype data for SNPs spanning the whole gene (chr12:10,202,166 to 10,216,057) were downloaded from the HapMap database available at http://hapmap.ncbi.nlm.nih.gov/cgi-perl/gbrowse/hapmap24_B36/#search server. The considered populations were the Chinese (CHB, 45 unrelated Han Chinese from Beijing, China), Japanese (JPT, 45 unrelated Japanese from Tokyo, Japan), Nigerian (YRI, 30 Yoruba mother–father–child trios from Ibadan, Nigeria) and European (CEU, 30 mother– father–child trios from the CEPH collection Utah residents with ancestry from northern and western Europe) from the second release of the project. Offspring genotypes were removed to analyze only unrelated subjects. Formatted files were submitted to the Phase v2.1.1 software for haplotype reconstruction^23-25. Analysis of Molecular Variance (AMOVA) and Tajima's D was then computed using Arlequin 2.0 version^{26, 27}. Through AMOVA, we calculated F_st, which is a measure how genetic variation is partitioned within vs. among populations, and therefore serves as a means to discern genetic differences among populations. SNPs used in these analyses are listed in Supplemental Table 6.

In a second phase of the analysis, the haplotypes were split into two regions: the LD block that includes intron 4 and the rest of the gene. Analyses were then repeated in these regions separately and compared (Figure 1 and Supplemental Table 6).

Analyses on humans were also replicated in the samples from the 1000 genomes project (Supplemental Table 3). We used data from 55 Tuscans from Italy (TSI); 86 samples from Great Britain from both England and Scotland (GBR); 92 Finnish from Finland (FIN); 86 Yoruba from Ibadan, Nigeria (YRI); 58 African American form the Southwest (ASW); 99 Southern Han Chinese (CHS); 97 Han Chinese from Bejing, China (CHB); 88 Japanese from Tokyo, Japan (JPT). Unfortunately, only 4 sequences from the CEU populations were available for this gene region and were excluded from our analyses.

Tajima's D was also calculated on 18 P. troglodytes sequences using the DNAsp software, version 5.10.01 ²⁸. Tajima's D was calculated separately for different regions (overall coding region including all exons, 5′UTR and exon 1, exon 2 and surrounding intronic region, exon 3 and surrounding intronic region, exon 4 and upstream region, exon 4, exon 6, surrounding region and 3′UTR). In all cases significance of selection was determined using a p value cutoff of 0.05.

Phylogenetic Analysis

We estimated the evolutionary history of both coding and non-coding regions of all sequences from all taxa using the maximum likelihood (ML) optimality criterion, as implemented in the RAxML software, version 7.2.6 ²⁹. The ML estimates of the phylogenies of both sets of sequences were obtained assuming a general time reversible model of nucleotide evolution, empirically measured nucleotide frequencies, and allowing for rate heterogeneity across sites. The appropriate number of bootstrap replicates was assessed through the frequency-based stopping criterion implemented in RAxML ³⁰. Because the ML estimates of the evolutionary history of both coding and non-coding regions may differ from the standard species phylogeny ^{31, 32}, we evaluated whether each of the ML trees was significantly different from the species phylogeny using the Shimodaira-Hasegawa test ¹³, as implemented in RAxML.