Bivariate genome-wide association analysis strengthens the role of bitter receptor clusters on chromosomes 7 and 12 in human bitter taste

Bitterness is a taste sensation that arises when particular chemicals come into contact with receptors in specialized cells on the human tongue [1–3]. But not everyone perceives the same bitterness for a given stimulus; this individual variation is partially genetically determined and can affect food perception, preferences and intake [4–6]. Genetic effects for bitter taste perception, which are estimated by twin studies, range from 36 to 73% [7–9], with most of the known variation arising from inborn variation in the bitter receptor gene family (T2R) [10–12]. These bitter receptors are found in tissues beyond the tongue and oral cavity, including the airways, gut, thyroid, and brain [13] where they may function as toxin detectors or early-stage sentinel systems. Bitter taste responses may reflect how well the receptors detect ligands in other tissues [14]. Historically, the ability to taste one well-studied bitter compound, phenylthiocarbamide (PTC), has been related to many diseases [15]; more recently and more specifically, variation in the PTC taste receptor is shown to be involved in the immune system [16] and to predict surgical outcome for severe rhinosinusitis [17]. Thus, together with the better-known effects on food intake and nutrition, bitter taste perception is of increasing importance to the medical field.

Given the rising importance of taste genetics, studies have focused on determining the underlying genetic variation that leads to individual differences in bitter perception. Our earlier genome-wide association study [12] (GWAS), which included 1457 adolescents from 626 twin families, replicated the classic association between the bitter taste receptor gene TAS2R38 and the perception of propylthiouracil (PROP; a chemical relative of PTC) and revealed a single nucleotide polymorphism (SNP) within the bitter taste receptor gene TAS2R19, accounting for 5.8% of the variance in the perception of quinine. Whereas the quinine does not activate T2R19 (the protein product of TAS2R19) in vitro [18], the SNP was later shown to form a long-range haplotype with missense variants within a nearby bitter taste receptor gene TAS2R31 [19, 20] whose encoded protein T2R31 can be activated by quinine [18]. The GWAS study, however, could neither detect loci for the other compounds tested, such as caffeine and sucrose octaacetate (SOA), that are likely to be affected by a large number of small-effect alleles nor the previously proposed but yet to be identified second locus for thiourea-containing compounds like PTC and PROP [21], including the suggested loci on chromosomes 5 [22] and 16 [23].

Drawing on studies of complex traits such as body mass index [24] and schizophrenia [25], here we increased the overall sample size by 40% and used multivariate association analysis [26] to identify common genetic variants (minor allele frequency [MAF] ≥ 5%) with small effects. Multivariate GWAS has been used to detect SNP associations that did not reach genome-wide significance in univariate analyses, such as autism spectrum disorders [27] and bone mineral density [28]. This method can detect not only pleiotropic genetic variants but also variants associated with only one of the correlated phenotypes [29]. As shown by Stephens [29], bivariate analysis increases power when there is greater separation of genotype groups (0, 1 or 2 copies of the minor allele) in two- versus one-dimensional space. In Fig. 1a and b, we provide two illustrations of when a joint analysis of two correlated traits can provide greater separation of genotypes associated with the primary trait (trait 1). The first example (a) shows the case where only one trait (trait 1 on the y-axis) is associated with the variant (non-pleiotropic), with bivariate analysis providing better separation of the genotype groups in 2-dimensional space compared with the y-axis alone. A similar boost in signal would be found in a conditional analysis, where the non-associated trait is included as a covariate, as this removes the non-associated part of the variance in the associated trait (i.e. covariance between two traits) and, therefore, enhances the association. The second example (b) shows that maximum separation can be achieved when both traits (trait 1 on the y-axis, trait 2 on the x-axis) are associated with the variant and the effect of the minor allele on the two is in opposite direction. In the case where a variant has the same effect on both correlated traits (Fig. 1c), bivariate analysis provides minimum/no increase in power. The bivariate approach is especially well-justified for bitter taste traits because, with the exception of PROP, perception of these bitter substances are highly correlated (r_p = ~0.6) [30] and their genetic variances largely overlap (r_g = ~0.7) [7, 9].

Here we aimed to identify common genetic variants with small effects (i.e. 1 – 5%) on the perception of bitterness, building on our previous GWAS [12], which was under-powered to detect common genetic variants with small effects. We performed univariate GWAS for the perceived intensity of 5 bitter substances (PROP, quinine, caffeine, SOA, and denatonium benzoate [DB]) using our expanded sample, including 1999 individuals from 929 twin families. As these phenotypes were collected from the same individuals, to boost power we ran a series bivariate GWAS (6 in total) for the correlated phenotypes of quinine, caffeine, SOA and DB [9]. We looked for evidence of pleiotropy for each identified variant. When there was little evidence for pleiotropy, we tested the SNP association to the primary trait conditional on the second. For variants in linkage disequilibrium, we used bidirectional conditional analysis (i.e. including the genotype of one SNP as a covariate at a time to test the association with the other SNP) and plotted the SNP associations for one trait against the other. Finally, to help interpret the genotype-phenotype associations, we examined the potential function of the identified SNPs with bioinformatics tools.

In this study of bivariate GWAS on human taste perception, we identify two putative novel associations, including rs67487380 on chromosome 12 for SOA-elicited bitter taste and rs10261515 on chromosome 7 for DB-elicited bitter taste. In addition, we provide the first independent replication of an association on chromosome 12 for caffeine bitterness (rs2597979) and confirm our previously reported associations for quinine bitterness (rs10772420 on chromosome 12) and PROP bitterness (rs10246939 on chromosome 7). All variants are located within the bitter taste receptor clusters on chromosomes 7 and 12, highlighting the importance of these two regions in the genetics of bitter taste. Further, we show evidence of pleiotropy for those variants on chromosome 12 and the functional importance of the DB-associated SNP.

This is the first GWAS study to identify a SNP (rs67487380 on chromosome 12) association with human perception of SOA. In mice, a major locus for SOA perception (soa) was reported in the early 1990s [33]. Interestingly, the mouse soa locus also affects the perception of other bitter substances, including quinine, DB, PROP, but not caffeine [34, 35]. Here we provide evidence that rs67487380 is also associated with the perception of quinine and DB, but not caffeine or PROP (P > 0.05). SOA activates human T2R46 but no other T2Rs in heterologous expression assays [32]. It is possible that rs67487380 regulates the perception of SOA through its effect on mRNA expression because the G allele for weaker SOA intensity is also associated with a lower expression level of TAS2R46. Nevertheless, rs67487380 could still be a proxy for true causal variants.

The finding of the novel association between DB and the SNP rs10261515 suggests that there may be a second locus on chromosome 7 that affects human bitter taste perception (the first is the locus within TAS2R38 for PROP). Heterologous expression studies using human embryotic kidney (HEK) cells transfected with TAS2Rs have shown that DB activates T2R4 but no other bitter taste receptors in this region (e.g. T2R3, T2R5 and T2R38) [18]. In addition, the human T2R4 is the ortholog of mouse T2R8, which also responds to DB [1]. Our functional annotation results provide further support for T2R4 as a DB bitter taste receptor, since the allele (rs10261515 G allele) for a lower perceived intensity of DB is associated with lower expression level of TAS2R4 mRNA.

The SNP association for caffeine perception replicated a previous GWAS of 608 Brazilian adults [11]. In that study the lead SNP accounted for 8.9% of the variance in caffeine sensitivity, compared with our estimate of 1.9%. Similarly, the Brazilian study accounted for 23.2% of the variance of quinine taste with genetic mutations, which is four times the effect estimated here. This difference in effect sizes is likely due to two main factors. First, the taste scores in the Brazilian sample were corrected for overall-taste-sensitivity (an average score of the perception of sweet, umami, sour, salty and bitter compounds), which removed ~30% of the variance in the perception of caffeine and quinine. Without correction, rs10772420 accounted for 13.2% of the variance in quinine, and the caffeine association was not detected due to low power. Second, the Brazilian study used a detection threshold approach, which measures overall oral sensitivity, compared with our measure of bitter taste intensity. Regardless, both studies identified the same variants for caffeine, quinine as well as PROP, indicating that these are likely to be valid associations among human bitter taste perception and these T2R-rich regions of chromosomes 7 and 12.

The functional annotation of the caffeine-associated SNP showed that the highly correlated SNP (rs10743938) is a missense mutation that could affect the function of T2R31. This is the first evidence linking this bitter taste receptor to the perception of caffeine, while genetic variants in TAS2R31 have been shown to affect the perception of quinine [19] (confirmed in the present study), acesulfame potassium and saccharin [36] (the latter two are non-nutritive sweeteners with bitter aftertaste). Prior cell-based functional studies [18] reported that caffeine does not activate T2R31 in heterologous expression assays; rather, it activates T2R7, -10, -14, -43, and -46, and that the summed expression level of these activated T2Rs increases with the perceived intensity of caffeine [37]. However, comparing results from bioinformatics and cell-based analyses can be limited by two major factors. Here, we report associations for the index (lead) SNP with the lowest P-value, but since this SNP is in a linkage disequilibrium block, the association could be driven by any variant within the block. Second, these cell-based functional assays [18] were conducted in heterologous systems (i.e. HEK cells transfected with TAS2Rs), which may not always recapitulate human sensory experience faithfully [38]. We observed a similar difference for quinine, with the lead quinine-associated SNP rs10772420 constituting a missense mutation in TAS2R19. Yet T2R19 does not respond to quinine in functional expression assays [18]. Instead, rs10772420 is more likely to be a proxy of missense variants within TAS2R31 [19, 20], whose encoded protein T2R31 responds to quinine [18]. Therefore, a better method to identify causal SNPs for the foreseeable future is to tightly integrate genetic-perceptual association results with those of taste receptor cell-based assays using human taste tissues, such as taste buds or cultured human taste cells [39].

This study provides the first evidence for antagonistic genetic pleiotropy in bitter taste. The two SNPs rs10772420 and rs2597979 have opposite effects on the perceived intensity of quinine and caffeine (Additional file 9: Figure S1, Additional file 10: Table S9) and this largely enhances the strengths of their associations (P-value) in the bivariate analysis (Fig. 1b). As bitter-tasting substances (e.g. caffeine) can have both beneficial and detrimental effects, the antagonistic pleiotropy may be an evolutionary consequence that avoids over and under consumption.

The top SNPs for quinine, caffeine, and SOA were correlated (r² = 0.08 – 0.43) and each could have various effects on one another. These correlations are due to the linkage disequilibrium between polymorphisms within bitter taste receptor genes on chromosome 12, which results in common haplotypes for nearby genes and long-range haplotypes for more distant ones [20, 36]. Previous studies have revealed a complex bitter substance – receptor relationship, with one bitter compound activating multiple T2Rs and one T2R responding to multiple bitter substances [18, 32, 40]. Taken together, it is likely that the perception of a bitter taste can be mediated by multiple T2Rs, and SNPs identified in the present study could represent haplotypes that regulate several T2Rs together. We have attempted to illustrate this in Fig. 7 by taking the perception of quinine and caffeine as an example. The lead SNP for quinine (rs10772420) is correlated with several SNPs (the regional association plot in Fig. 3) that regulate the T2Rs for quinine (cell-based functional analysis results in Table 3). Also, the lead SNP for caffeine (rs2597979) is correlated with SNPs that regulate T2Rs for caffeine (Table 3). In addition, the common T2Rs for the two tastes are regulated by SNPs that are in linkage disequilibrium with the two lead SNPs. We note that the real regulatory network can be more complex than this, such that one T2R can be regulated by multiple SNPs. Whereas we used conditional analysis (Table 2) and plotted the SNP associations against the three tastes (Fig. 4) to show that each of the lead SNPs represents the main signal in the linkage disequilibrium block, the clusters of nearby bitter receptors and many variants in high linkage disequilibrium create challenges in separating causal from non-causal variants.

Perceptual studies of bitter taste also have reported that individual differences in perceived bitterness from multiple compounds show positive correlations. Most relevant to the present work, past studies demonstrated a strong correlation of perceived bitter taste intensities among DB, SOA, and quinine [30]. This observation harkens to that reported in the present study for rs67487380 on chromosome 12. Furthermore, individual differences in bitterness from SOA, caffeine and quinine were also observed, suggesting a linkage between SOA receptor variants and caffeine receptor variants [30]. This too reflects associations observed in the present data set. Perhaps, a linkage disequilibrium block accounts both for the genetic architecture as well as the bitterness perception associations.

Prior work using pedigree segregation analysis has proposed that the perception of PTC (a structurally related chemical to PROP) is modulated by a 2-locus model [21], but the location of a second locus has been unclear for nearly 30 years. Here we found neither support for an association with TAS2R1 on chromosome 5 nor an association on chromosome 16, both of which was suggested by prior family-based linkage studies [22, 23], but identified a putative secondary locus within the DIRC3 gene on chromosome 2, which accounted for an additional 1.83% of the variance (4.58% of the genetic variance) in the perception of PROP paper. While we found no evidence for replication using three independent datasets – from one published study (i.e. the Brazilian sample) and two unpublished (the Silk Road and the Italian samples), we note that there are considerable differences across studies (e.g. sample age [all other studies used adults], ethnicity, and delivery method [the Brazilian study used PROP solution]), which may have influenced our ability to replicate their findings. Ideally, we need to test for this association using the same methods and materials (i.e. the perceived intensity of saturated PROP paper measured from adolescents with European ancestry), but at this stage the signal does not appear to be sufficiently robust to be detected with alternative methods.

The strengths of the present study include the use of the largest sample to date, and the collection of multiple taste phenotypes from the same individuals, which increases the statistical power via bivariate association analysis. We show that the association signals (P-value) for quinine and caffeine (rs10772420 and rs2597979 respectively) were stronger in the bivariate compared with the univariate analysis, but the estimated effect size remains the same. The signal boosts in these already established associations serve as a proof of principle for using bivariate GWAS. We also show that, through the discovery of the association of DB, a signal can be enhanced when only one of two correlated traits is associated. This is useful for identifying non-pleiotropic SNPs for correlated phenotypes. We used multiple levels of analysis (conditional on genotype and phenotype) as well as cluster plots to disentangle the pleiotropic nature of these SNPs with bitter tastes and provide additional support for the signals identified in the bivariate analyses. We attempted to obtain data to replicate every novel association. However, we were unable to test the association for SOA and DB because we are currently the only group in the world that has collected these two bitter taste phenotypes along with genomic data. Given the enhancement in the known signals for both quinine and caffeine in the bivariate analyses, together with the post-hoc bioinformatics analyses, as well as prior functional analyses, we believe the SOA and DB hits are unlikely to be false positives. Further, findings from multivariate GWAS of other phenotypes, e.g. levels of plasma lipids [41], have been replicated in independent studies. The variants for SOA and DB account for less than 10% of the genetic variance (< 2% of trait variance) of their associated traits, suggesting that there are more variants with smaller effects. The remaining genetic variance could be partly due to rare variants because SNPs with an MAF smaller than 5% were excluded here and rare variants can have a large effect on complex traits [42].

This study reveals the influence of multiple variants on bitter taste and demonstrates the benefits of multivariate analysis over the conventional univariate GWAS. Recent advancement in the methodology of multivariate GWAS (i.e. MTAG [43]) could make multivariate analysis easier to apply because it uses individual summary level results from different studies and does not require correlated phenotypes to be collected from the same sample. Whereas our previous twin analysis provided strong evidence of pleiotropy for the perception of several bitter compounds (except for PROP), there are numerous causal models that could underlie this shared genetic etiology. Identification of specific SNPs/genes involved offers a useful starting point for determining the biological pathways linking perception of bitter substances and for delineating of the mechanisms involved. Future studies integrating bioinformatics and functional analyses using human taste tissues will provide stronger evidence in identifying true causal variants, which could assist personalized nutrition and precision medicine.