Non-coding rare variants in PANX3 are associated with chronic back pain

Introduction

Back pain (BP) is a major cause of disability worldwide. According to the Global Burden of Disease Study 2015, low back pain together with neck pain were the leading causes of global years lived with disability in 1990–2015 [10]. The most debilitating form of BP is chronic BP (CBP; back pain lasting more than 3 months). The prevalence of CBP is as high as 30% depending on the population and CBP definition used [22].

Precise biological mechanisms underlying CBP are yet to be elucidated. It is known that CBP is a complex heritable trait with narrow sense heritability in the range of 40%−68% [12]. A variety of genome-wide association studies (GWAS) using data from national-scale biobanks and electronic medical records have suggested a number of genes associated with CBP and similar conditions [3; 11; 31; 34; 35; 37], with at least nine genomic loci reported and replicated in independent samples. Among those loci, one of the most consistent associations across studies is with the SOX5 gene. Previous GWAS of CBP were focused on associations with common polymorphisms, i.e. those with minor allele frequency (MAF) of greater than 1%. Liability-scale SNP-based heritability of CBP explained by common variants has been estimated at 13%, which is substantially less than the 40% CBP heritability expected from classical twin studies [11; 30]. Thus, missing heritability estimated as 27% may be explained by other factors, in particular, by rare (MAF < 1%) variants [38].

Exploring associations with rare variants is a complex task requiring large samples. To solve this task, gene-based association analysis has been proposed. This method simultaneously considers a set of variants from a gene instead of a separate analysis of each variant. Gene-based association analysis allows identification of the genes whose variants are directly responsible for differences in disease risk between individuals. Protein-coding variants of these genes can modify the structure of the corresponding proteins; variants in non-coding intragenic regions can regulate the transcription and translation of these genes, protein complex formation, or posttranslational modifications [19; 24]. Such modifications may tilt the physiological balance from a healthy to a diseased state [6].

Recently, we have developed a comprehensive, flexible framework for gene-based association analysis that can be applied to summary-level GWAS results calculated for each variant on biobank-scale sample sizes [2; 36; 40]. This framework can be incorporated into the analysis of rare and ultra-rare (MAF < 0.01%) variants. Additionally, the framework includes a polygene pruning procedure aiming to reduce the influence of strong association signals located outside the gene. This allows us to identify genes associated with a disease due to their within-gene polymorphism.

In this study, we aimed to analyze the genes whose variants are directly responsible for differences in CBP risk between individuals using data from UK Biobank.

Materials and Methods

GWAS summary statistics

For gene-based analysis we used GWAS summary statistics for chronic back pain obtained from our previous study [37]. In short, we studied UK Biobank participants of European ancestry [33]. Genotyping and imputation data were obtained from the UK Biobank March 2018 data release. Genotyping was conducted using the Affymetrix UK BiLEVE and Affymetrix UK Biobank Axiom arrays. Imputation was performed with the IMPUTE4 program (https://jmarchini.org/impute-4/) [4] using the Haplotype Reference Consortium (HRC) [25] and merged UK10K and 1000 Genomes phase 3 reference panels. Cases and controls were defined based on questionnaire responses. First, participants responded to “Pain type(s) experienced in the last months” followed by questions inquiring if the specific pain had been present for more than 3 months. Those who reported back pain lasting more than 3 months were considered chronic back pain cases. Participants reporting no such pain lasting longer than 3 months were considered controls. Individuals who preferred not to answer or reported more than 3 months of pain all over the body were excluded from the study. A sample of 265,000 participants of European ancestry (defined by SNP-based principal component analysis) was randomly selected as the GWAS discovery sample. The remaining European ancestry participants (N = 174,831) comprised a replication cohort.

GWAS were carried out using BOLT-LMM v.2.3.2 software [21]. Linear mixed-effects models were fitted to test for additive effects of SNP alleles on the chronic back pain phenotype adjusting for age, sex, genotyping platform batch, and the first ten principal components of genetic variation. The following filters were applied: minor allele count ≥ 5 and imputation quality score ≥ 0.3; genotyping and individual call rates > 0.98. Only biallelic autosomal SNPs and indels were analyzed.

Gene-based association analysis

The gene-based association analysis using the summary statistics was performed by three methods: SKAT-O[18], PCA[39] and ACAT-V[20] implemented in the sumFREGAT R-package[36]. The results from these methods were combined using the aggregated Cauchy omnibus test, ACAT-O [20]. Only protein-coding genes with at least two SNPs having the GWAS summary statistics were analyzed. SNPs were annotated to genes based on dbSNP version 135 SNP locations and NCBI 37.3 gene definitions. SNPs were annotated to a gene if they were located between the gene’s transcription start and stop sites. We used 1000 Genomes functional annotations for genetic variants to determine nonsynonymous, coding and non-coding variants. Correlations between genotypes of variants within each gene were estimated using the genotypes of the 265,000 UK Biobank participants of European Ancestry from the discovery sample [1].

We considered four scenarios:

1. all SNPs located between the transcription start and stop positions of a gene;

2. non-coding SNPs (introns, 5′UTR and 3′UTR, non-coding exon variants);

3. protein coding SNPs (exons);

4. nonsynonymous substitutions.

To reduce the influence of association signals located outside the region of interest, we applied a polygene pruning procedure[2] to exclude each SNP that is in high LD (r² > 0.5) with genome-wide significant SNPs located outside the region and more strongly associated with CBP.

Genes having a combined ACAT-O p-value ≤ 2.5 × 10^–6 (standard gene-based significance threshold) after the pruning procedure were considered as genome-wide significant genes. Genes with p-value ≤ 2.5 × 10^–5 were considered as suggestively significant. For genes and scenarios with genome-wide significant p-values, replication in the separate UK Biobank cohort (N = 174,831) was performed. Only SNPs that were present in the discovery sample and passed the polygene pruning were used in replication in UK Biobank. The threshold for replication was set as p-value ≤ 0.05/(number of genes selected for replication).

Replication of PANX3 in FinnGen

For the one novel gene attaining genome-wide significance, which was subsequently replicated in UK Biobank (PANX3), an additional validation was performed using FinnGen Biobank GWAS results version 6 (Finngen-r6, https://r6.finngen.fi/).

FinnGen individuals were genotyped with Illumina and Affymetrix chip arrays (Illumina Inc., San Diego, and Thermo Fisher Scientific, Santa Clara, CA, USA). Genotype imputation was done with the population-specific SISu v3 imputation reference panel [17] with Beagle 4.1 (version 08Jun17.d8b) as described in the following protocol: dx.doi.org/10.17504/protocols.io.nmndc5e. Variant call set was produced with GATK HaplotypeCaller algorithm by following GATK best-practices for variant calling. Genotype-, sample- and variant-wise QC was applied in an iterative manner by using the Hail framework v0.1 and the resulting high-quality WGS data for 3,775 individuals were phased with Eagle 2.3.5. Post-imputation quality-control involved checking expected conformity of the imputation INFO-value distribution, MAF differences between the target dataset and the imputation reference panel and checking chromosomal continuity of the imputed genotype calls.

We examined a biologically similar phenotype “Dorsalgia” (N cases = 35,486, N controls = 194,289), an electronic health record (EHR)-based back pain phenotype defined by ICD-10 codes. EHR-based back pain phenotypes have been used extensively in research over the past 3 decades [7–9; 23], and we have recently used ICD-10-defined phenotyping for GWAS of back pain-related phenotypes [16]. “Dorsalgia” reflects back pain associated with health care utilization, and therefore captures both acute and chronic back pain but predominantly the latter, as chronic/recurrent back pain is more associated with health care seeking [5; 15; 28]. Using LD score regression method, we estimated the genetic correlation between dorsalgia phenotype and CBP as 0.87 (p-value < 8 × 10^–79), which is an indicator of major shared genetic influences between self-reported chronic back pain and the EHR-based “dorsalgia” back pain phenotype.

For the gene-based association analysis, we used LD correlations between genotypes of variants provided by FinnGen. The polygene pruning procedure was not necessary as there were no genome-wide significant GWAS p-values within 5Mb distance from the start/end positions of the PANX3 gene. The threshold for replication was set at p-value ≤ 0.05.

We also tested the gene-based association between PANX3 and phenotype “other intervertebral disc disorders” (https://risteys.finngen.fi/phenocode/M13_INTERVERTEB) using FinnGen Biobank GWAS results.

Results

For the discovery sample, gene-based analysis was performed for 18,502 genes. Two of them, SOX5 and PANX3, reached the genome-wide significance threshold (p-value < 2.5 × 10^–6) (Table 1). We detected suggestive (p-value < 2.5 × 10^–5) signals for 39 genes (see Supplementary Table 1).

Table 1.

Significant results of the discovery and replication gene-based analysis for four SNP sets.

Gene	NSNP	All			Intronic			Exonic	Non-synony-mous
		PD	PR	PV	PD	PR	PV	PD	PD
PANX3	176	6.23 × 10−10	0.043	0.018	4.07 × 10−10	0.020	0.033	0.549	0.330
SOX5	21520	1.57 × 10−7	0.003	-	1.28 × 10−7	0.003	-	0.878	0.997

^*N_SNP is the number of all informative SNPs in the gene. P_D: discovery p-values. P_R: replication p-values in UK Biobank. P_V: validation p-values in FinnGen. All p-values reflect those after polygenic pruning. Bold font marks the most significant SNP set on discovery.

For both genes, the strongest associations (smallest p-values) were achieved for intronic SNPs. The signals were replicated using the replication cohort from UK Biobank and the same SNP set (Table 1). For exonic and non-synonymous sets, the results were not significant.

The full GWAS summary statistics for all SNPs in the SOX5 and PANX3 genes are presented in Supplementary Tables 2 and 3, respectively. In the SOX5 gene, there were 13 common intronic SNPs (MAF > 10%) that reached the SNP-wise genome-wide significance threshold (p-value < 5 × 10^–8). PANX3 contained only one genome-wide significant rare intronic SNP (rs532090383) with MAF = 8 × 10^–5 (MAC = 42) and moderate imputation quality (r² = 0.37). It also had two SNPs with p-value < 10⁻³ with low MAF < 5.9 × 10^–5. All three SNPs are not in LD with each other (r² < 2 × 10^–7).

Since SOX5 is a well-known chronic back pain gene [11; 34; 37], we focused only on characterization of the novel PANX3 gene. We performed sensitivity analyses and repeated the gene-based tests in the discovery sample without rs532090383: the combined p-value became non-significant for all SNPs and intronic SNPs (p-value = 0.205 and 0.139, respectively).

As PANX3 was discovered due to a single SNP with very low MAF and moderate imputation quality, an elevated false discovery rate probability associated with the relatively low statistical power could be of concern. For this reason, we performed an additional replication using a similar phenotype, ‘Dorsalgia’, in the FinnGen Biobank GWAS results, version 6. The resulting combined gene-based p-value was significant for intronic and all SNPs scenarios (p = 0.033 and 0.018, respectively). GWAS summary statistics for all PANX3 SNPs in FinnGen are presented in Supplementary Table 4.

Discussion

In this work, we conducted a gene-based analysis of chronic back pain. The liberal threshold for imputation quality (≥ 0.3) and MAC ≥ 5 that we used allowed us to include rare (MAF < 1%) and ultra-rare (MAF < 0.01%) variants in the analysis. We discovered and replicated two genes. One of them, SOX5, has previously been found to be associated with CBP through common intronic variants. The other gene, PANX3, has not previously been reported in studies of CBP. We discovered that the association of PANX3 with chronic back pain involves rare and ultra-rare intronic variants.

PANX3 (pannexin 3) is one of three channel-forming proteins of the gap junction pannexin family. It has been detected in a variety of tissues including cartilage, bone and skeletal muscle, where it is implicated in processes of tissue development, regeneration and disease [29]. In many cell types of the musculoskeletal system, such as osteoblasts and chondrocytes, PANX3 has been shown to regulate the balance of proliferation and differentiation [13; 14; 29]. In mice, Panx3 was shown to alter the fate of chondrocyte cells from proliferation to differentiation by regulating intracellular ATP/cAMP levels [14]. Panx3-mediated ATP release accelerated hypertrophic differentiation, a process critical for skeletal development [14].

The role of PANX3 in osteoarthritis (OA) cartilage degeneration has been demonstrated previously. Its expression is upregulated in knee osteoarthritic (OA) tissues in mice and humans [26]. Interestingly, Panx3 knockout mice are markedly resistant to the development of OA following surgical destabilization of the medial meniscus [26] but exhibit accelerated OA progression in ageing [27]. Similarly, with regard to intervertebral disc (IVD) degeneration, Panx3 was found to be involved in mediating the response to altered mechanical stress following IVD injury but not in the progression of age-related histopathological IVD degeneration [32]. In other words, PANX3 seems to act in a context-dependent manner in both diseases: it may play either protective or detrimental roles depending on whether disease is injury-induced or age-related [29]. We can speculate that this may explain why PANX3 has not been described with musculoskeletal disorders to date: the power to detect an association signal could be reduced dramatically due to the context-dependent action of PANX3 coupled with a heterogeneity in phenotypes, where injury-induced and age-related pathologies can mix together.

The genetic correlation of 94% between chronic back pain and intervertebral disc disorder (IDD) has been estimated recently [3]. Indirect evidence in favor of a cartilage-mediated mechanism of PANX3 involvement in CBP is a relatively low gene-based p-value (3.52 × 10⁻⁵) we observed for the “other intervertebral disc disorders” phenotype in the FinnGen sample, a result far exceeding that for “Dorsalgia” (p-value = 0.018). We can speculate that a possible mechanism of action of the PANX3 on the back pain is due to its effect on the intervertebral discs.

We have replicated PANX3 gene using two samples from different biobanks – UK biobank and FinnGen. Given that the “Dorsalgia” phenotype was defined using ICD-10 codes reflecting health care use for back pain, rather than the self-reported symptom of CBP as in the UK biobank discovery sample, this shows external validity in distinct albeit related back pain phenotypes, and other populations.

We should note that the significant result of the PANX3 gene-based analysis in the discovery sample was driven mostly by one ultra-rare SNP, rs532090383, with moderate imputation quality (p-value < 1.18 × 10^–12, MAF = 8 × 10^–5, r² = 0.37). This could cast a doubt on the accuracy of the discovery results. This SNP wasn’t significantly associated in the UK Biobank replication sample (p-value < 0.21) and had an association p-value = 1.14 × 10^–5 in the full sample. The absence of the SNP-level replication can be explained by an instability of the distribution of a small number of rare allele carriers between discovery and replication samples that were sampled in random (see Supplementary Data). We evaluated the odds of such rare allele association to be replicated. We randomly divided the full sample on the discovery (N = 265,000) and replication (N = 174,831) samples 10,000 times and tested the association between CBP and dosages of rs532090383 in all samples. We obtained the discovery p-value < 5 × 10^–8 in 200 (2%) experiments. None of them were replicated (p-value < 0.05). This illustrates that ultra-rare SNP associations are not statistically stable and not expected to be replicated given the sample sizes. The gene-based analysis considering all SNPs in the region simultaneously avoids these potential problems and leads to more robust results.

On the gene-based level, we replicated PANX3 association in both replication samples (even though rs532090383 was not associated in the UK Biobank replication sample and was absent in FinnGen). We performed sensitivity analysis using the whole UK Biobank sample (N = 439,831). We removed all SNPs with imputation quality less than 0.7 and performed gene-based analysis for introns in PANX3 gene. The results still remained nominally significant (p-value = 0.036). We want to emphasize that the gene-based analysis solves the issue of random distribution of ultra-rare alleles because it takes into account the association of all SNPs in the region. Thus, the results of the gene-based association analysis shouldn’t be interpreted in the light of single SNP associations only. In any case, replication of the gene-based results in an independent sample is needed.

For the most strongly associated SNP in the UK Biobank sample, rs532090383, the rare allele A has an OR = 11.3 (95% confidence interval from 3.8 to 33.4). It is possible that including all rare variants from this gene in a polygenic risk model could increase model prediction accuracy. One limitation of the current study is the use of imputed genotypes rather than directly sequenced ones. Rare and ultra-rare variants are not well captured in imputed data and, therefore, the power of the analysis is lower than expected compared to sequencing data of a similar sample size. The analysis of whole-exome or whole-genome sequencing data of comparable sample sizes in future studies may reveal other new associations with CBP, including those of rare and ultra-rare variants. However, we do not expect that PANX3 would be discovered using whole-exome sequencing data given its association with CBP due to intronic variants.

To conclude, we found that SOX5 and PANX3 genes influenced CBP. The PANX3 gene has not previously been described as associated with CBP. Its association was shown to be due to non-coding rare intron polymorphisms.