Pathway-based analysis of primary biliary cirrhosis genome-wide association studies

Abstract

Genome-wide association studies (GWAS) have successfully identified several loci associated with primary biliary cirrhosis (PBC) risk. Pathway analysis complements conventional GWAS analysis. We applied the recently developed linear combination test for pathways to datasets drawn from independent PBC GWAS in Italian and Canadian subjects. Of the Kyoto Encyclopedia of Genes and Genomes and BioCarta pathways tested, 25 pathways in the Italian dataset (449 cases, 940 controls) and 26 pathways in the Canadian dataset (530 cases, 398 controls) were associated with PBC susceptibility (P < 0.05). After correcting for multiple comparisons, only the eight most significant pathways in the Italian dataset had FDR < 0.25 with tumor necrosis factor/stress-related signaling emerging as the top pathway (P = 7.38 × 10⁻⁴, FDR = 0.18). Two pathways, phosphatidylinositol signaling and hedgehog signaling, were replicated in both datasets (P < 0.05), and subjected to two additional complementary pathway tests. Both pathway signals remained significant in the Italian dataset on modified gene set enrichment analysis (P < 0.05). In both GWAS, variants nominally associated with PBC were significantly overrepresented in the phosphatidylinositol pathway (Fisher exact P < 0.05). These results point to established and novel pathway-level associations with inherited predisposition to PBC that on further independent replication and functional validation, may provide fresh insights into PBC etiology.

INTRODUCTION

Primary biliary cirrhosis (PBC) is the most common autoimmune liver disease and primarily affects women, with a prevalence of 1 in 1000 over the age of 40 years.¹ The serological hallmark of PBC is the formation of anti-mitochondrial antibodies against the pyruvate dehydrogenase complex subunit E2 (PDC-E2).² The antibodies specifically recognize immunoreactive PDC-E2 within apoptotic blebs of biliary epithelial cells.³ Untreated disease involves progressive, non-suppurative granulomatous inflammation and autoreactive T lymphocyte-mediated destruction of the small-to-medium intrahepatic bile ducts leading to chronic cholestasis, portal inflammation, cirrhosis and end-stage liver disease.⁴ The accepted concept of PBC etiology is that it arises on a background of strong genetic susceptibility that is reactive to a variety of potential environmental triggers. The disease has a monozygotic concordance of 63%,⁵ a sibling relative risk of 10.5,⁶ and 1–6% of all patients with PBC have at least one first-degree relative affected.⁷ Other autoimmune disorders also tend to be more common in the families of PBC cases.⁸ To date, there have been three genome-wide association studies (GWAS) for PBC that have reproducibly identified several risk loci that implicate key gene loci involved in adaptive immunity and inflammatory response.^9–12

The genetic associations with PBC risk identified by the genome-wide approach are just those single nucleotide polymorphisms (SNPs) that met the most stringent criterion for statistical significance applied to account for the exceedingly large number of statistical comparisons made in a GWAS. Many more variants are typically associated with disease only at the nominal significance level in a GWAS and are therefore not investigated further.¹³ However, if the excess familial risk for PBC is to be explained, some of these discarded variants must be false negatives and constitute genuine PBC susceptibility loci. Moreover, due to population genetic heterogeneity, different SNPs in or near the same gene or in a functionally related gene may be associated with the disease among individual cases in a GWAS sample. This makes it less likely that a replicable association with the disease would be found when testing SNPs one at a time as is usually done in a GWAS.¹⁴ Single nucleotide polymorphisms and the genes that they belong to are not random entities. The products of specific sets of genes interact as members of discrete molecular and cellular pathways with defined biological function.¹⁵ Collectively, these observations have motivated the development of methods for the secondary or complementary statistical analysis of GWAS data that use biological pathways represented by gene sets, instead of SNPs, as the units of analysis.¹⁶ Pathway-based tests provide a dynamic biologically plausible template to efficiently integrate statistical information from the multitude of SNPs with weaker effects that are otherwise missed by conventional single-SNP GWAS analysis.¹⁷

In GWAS pathway analysis, one can map SNPs to genes and test for overrepresentation of statistically significant association signals among genes within a known biological pathway compared with the number of such signals among genes outside the pathway. This is termed the ‘competitive’ approach.¹⁶ Alternatively, one can jointly test all genes within the pathway for an association with the disease. As the latter only considers disease association signals within a pathway and does not compare them to signals outside the pathway, it is termed the ‘self-contained’ approach. So far, these methods have been successfully employed to analyze GWAS of a diverse group of diseases.^18,19 Here, we report the results of pathway analysis of two datasets from previously completed GWAS in independent Italian and Canadian PBC cohorts. We applied a recently developed ‘self-contained’ GWAS pathway analysis method, the linear combination test (LCT) of Luo et at.,²⁰ to identify pathways associated with genetic predisposition to PBC and provide greater insight into the etiology of this complex autoimmune disease. In accordance with recommendations for validating GWAS pathway analysis findings by complementary methods,^14,21 the statistical significance of top pathways that were replicated in both datasets was further confirmed by using two ‘competitive’ pathway-oriented strategies.

RESULTS

Linear combination test for pathway analysis

First, we individually analyzed the Italian and Canadian datasets using the LCT.²⁰ This algorithm uses raw genotype data to first compute genome-wide single-SNP association statistics (see Material and methods for details). After assigning all SNPs between the start site and the 3’ untranslated region to a gene, single-SNP P-values were combined for each gene using the gene-level LCT statistic derived by Luo, et al.²⁰ Genes were classified into pathways using the well-accepted Kyoto Encyclopedia of Genes and Genomes (KEGG) and BioCarta resources. Finally, gene-level statistics for all genes within a pathway were combined using the pathway-level LCT (see Material and methods). For these tests, we set P < 0.05 and false discovery rate (FDR) < 0.25 as a stringent criteria for significance (Material and methods).

In the Italian dataset, 207 695 SNPs out of 468 982 SNPs were located within genes and mapped to 14 527 genes. Of these, 4172 genes were assigned to pathways for LCT analysis. In the Canadian dataset, 143 059 SNPs out of 334 444 SNPs were located within genes and mapped to 14792 genes. Of these, 4226 genes were assigned to pathways for LCT analysis. Pathways with > 10 genes accounted for 175 BioCarta and 172 KEGG pathways in the Italian study and 176 BioCarta and 172 KEGG pathways in the Canadian study. At the gene-level, the LCT identified 253 genes in the Italian sample and 236 genes in the Canadian sample with P-value < 0.05. These genes are listed in Supplementary Table S1. As shown in Supplementary Table S2, there was limited overlap between significant genes in the two datasets.

Pathways suggested from the linear combination test analyses

At the pathway-level, the LCT identified 25 pathways (13 BioCarta, 12 KEGG) in the Italian dataset at the P < 0.05 level (Table 1). Of these, eight pathways achieved the threshold for statistical significance we set for this study (P < 0.05 and FDR < 0.25). Notably, these eight included three pathways that are likely to be important in the pathogenesis of inflammatory and autoimmune processes: tumor necrosis factor (TNF)/stress-related signaling pathway (P = 7.38 × 10⁻⁴, FDR = 0.18), antigen processing and presentation (P = 1.08 × 10⁻³, FDR = 0.18), and chaperones modulate interferon signaling pathway (P = 2.33 × 10⁻³ FDR = 0.192).

Table 1.

Pathways identified in the Italian dataset using linear combination test analysis^a

Rank	Pathway name	Source	Pathway size (genes)	P-value (LCT)	FDR
1	TNF/stress-related signaling pathway	BioCarta	25	0.0007	0.180
2	Bole of MAL in Rho-mediated activation of SRF pathway	BioCarta	19	0.0009	0.180
3	Antigen processing and presentation	KEGG	88	0.0011	0.180
4	Hypoxia and p53 in the cardiovascular system pathway	BioCarta	23	0.0019	0.192
5	Glycerophospholipid metabolism	KEGG	66	0.0022	0.192
6	Chaperones modulate interferon signaling pathway	BioCarta	19	0.0023	0.192
7	Small cell lung cancer	KEGG	87	0.0035	0.218
8	Acetylation and deacetylation of RelA in the nucleus pathway	BioCarta	16	0.0047	0.245
9	Signal transduction through IL1R pathway	BioCarta	33	0.0059	0.269
10	CBL-mediated ligand-induced downregulation of EGF receptors pathway	BioCarta	13	0.0066	0.276
11	Rho cell motility signaling pathway	BioCarta	32	0.0092	0.333
12	Role of PI3K subunit p85 in the regulation of actin organization and cell migration pathway	BioCarta	16	0.0120	0.377
13	Non-small cell lung cancer	KEGG	54	0.0154	0.436
14	Phosphatidylinositol signaling system	KEGG	77	0.0156	0.436
15	VEGF signaling pathway	KEGG	74	0.0232	0.582
16	Calcium signaling pathway	KEGG	179	0.0282	0.582
17	CD40L signaling pathway	BioCarta	15	0.0300	0.582
18	Endometrial cancer	KEGG	52	0.0311	0582
19	Cyclins and cell Cycle regulation pathway	BioCarta	23	0.0313	0.582
20	MAPK signaling pathway	KEGG	267	0.0335	0.582
21	Complement and coagulation cascades	KEGG	69	0.0383	0.636
22	Hedgehog signaling pathway	KEGG	57	0.0438	0.636
23	Graft-versus-host disease	KEGG	42	0.0460	0.636
24	Ras-independent pathway in NK cell-mediated cytotoxicity pathway	BioCarta	20	0.0469	0.636
25	Classical complement pathway	BioCarta	14	0.0484	0.639

Abbreviations: FDR, false discovery rate; KEGG, Kyoto Encyclopedia of Genes and Genomes; LCT, linear combination test.

^aAll pathways with P < 0.05 are shown.

There were 26 pathways (8 BioCarta, 18 KEGG) in the Canadian dataset at the P < 0.05 level using LCT analysis (Table 2). None of these pathways met the more stringent criterion for statistical significance (P < 0.05 along with FDR < 0.25). However, three of the pathways had an FDR of < 0.5: regulation and function of carbohydrate-responsive element-binding protein (ChREBP) in liver pathway (P = 5.68 × 10⁻⁴, FDR = 0.285), bone remodeling pathway (P = 2.33 × 10⁻³, FDR = 0.493) and apoptosis (P = 3.96 × 10⁻³, FDR = 0.493). For both datasets, there was no significant correlation between the number of genes in a pathway and pathway rank. Complete LCT pathway analysis results for the Italian dataset are presented in Supplementary Table S3 and for the Canadian dataset in Supplementary Table S4. We also present the pathway results from both datasets combined using Fisher’s method for meta-analysis in Supplementary Table S5.

Table 2.

Pathways identified in the Canadian dataset using linear combination test analysis^a

Rank	Pathway name	Source	Pathway size (genes)	P-value (LCT)	FDR
1	Regulation and function of ChREBP in liver pathway	BioCarta	44	0.0006	0.285
2	Bone remodeling pathway	BioCarta	14	0.0023	0.493
3	Apoptosis	KEGG	90	0.0040	0.493
4	Renal cell carcinoma	KEGG	69	0.0101	0.648
5	Phenylalanine metabolism	KEGG	21	0.0102	0.648
6	Keratan sulfate biosynthesis	KEGG	16	0.0114	0.648
7	DNA replication	KEGG	36	0.0128	0.648
8	Tyrosine metabolism	KEGG	46	0.0129	0.648
9	Stress induction of HSP regulation pathway	BioCarta	15	0.0152	0.693
10	The IGF-1 receptor and longevity pathway	BioCarta	15	0.0183	0.693
11	Amyotrophic lateral sclerosis (ALS)	KEGG	19	0.0204	0.693
12	Regulation of BAD phosphorylation pathway	BioCarta	26	0.0247	0.693
13	Basal cell carcinoma	KEGG	55	0.0285	0.693
14	Role of mitochondria in apoptotic signaling pathway	BioCarta	21	0.0329	0.693
15	Colorectal cancer	KEGG	85	0.0332	0.693
16	Phosphatidylinositol signaling system	KEGG	77	0.0335	0.693
17	Cytokine-cytokine receptor interaction	KEGG	263	0.0365	0.693
18	Wnt signaling pathway	KEGG	150	0.0375	0.693
19	Skeletal muscle hypertrophy is regulated via AKT/mTOR pathway	BioCarta	20	0.0385	0.693
20	Hedgehog signaling pathway	KEGG	57	0.0407	0.693
21	Adherens junction	KEGG	78	0.0434	0.693
22	IL-10 anti-inflammatory signaling pathway	BioCarta	17	0.0457	0.693
23	Carbon fixation	KEGG	24	0.0460	0.693
24	Bladder cancer	KEGG	42	0.0460	0.693
25	Ubiquitin-mediated proteolysis	KEGG	135	0.0476	0.693
26	Biosynthesis of phenylpropanoids	KEGG	33	0.0478	0.693

Abbreviations: FDR, false discovery rate; KEGG, Kyoto Encyclopedia of Genes and Genomes; LCT, linear combination test.

^aAll pathways with P < 0.05 are shown.

Two pathways reached the P < 0.05 level in both datasets using the LCT. They were the phosphatidylinositol signaling system (Italian: P = 0.016, FDR = 0.436; Canadian: P = 0.034, FDR = 0.693; meta-analysis P = 4.48 × 10⁻³) and the hedgehog signaling pathway (Italian: P = 0.044, FDR = 0.636; Canadian: P = 0.041, FDR = 0.693; meta-analysis P = 0.013).

Complementary analyses supports specific pathways

The phosphatidylinositol signaling system and the hedgehog signaling pathways were followed up in each dataset by two complementary pathway analysis methods (Table 3). First, we applied i-GSEA4GWAS, a modification of the gene set enrichment analysis (GSEA) approach that uses SNP label permutation (see Material and methods).²² Using the i-GSEA4GWAS algorithm, the association between the phosphatidylinositol signaling system pathway and PBC was found to be statistically significant in the Italian dataset (P = 0.003), but not in the Canadian sample. Similarly, the hedgehog signaling pathway yielded significant results on i-GSEA4GWAS in the Italian dataset only (P = 0.005).

Table 3.

Complementary tests for pathways identified at the P < 0.05 level in both datasets using LCT

Pathway namea	LCT		GSEAb		Fisher’s exact test
	Italian P-value (FDR)	Canadian P-value (FDR)	Italian P-value	Canadian P-value	Italian P-value (Ratioc)	Canadian P-value (Ratioc)
Phosphatidylinositol signaling system	0.0156	0.0335	0.003	NS	1.4E-04	0.0003
	(0.436)	(0.693)			(0.416)	(0338)
Hedgehog signaling pathway	0.0438	0.0407	0.005	NS	0.103	1.000
	(0.636)	(0.693)			(0.281)	(0.158)

Abbreviations: FDR, false discovery rate; GSEA, gene set enrichment analysis; LCT, linear combination test; NS, not significant.

^aThese pathways were defined in the Kyoto Encyclopedia of Genes and Genomes (KEGG).

^bThis analysis was performed using the i-GSEA4GWAS modification of GSEA.

^cRatio of number of genes containing at least one SNP with P < 0.05 to the total number of genes in the pathway.

Second, we used Fisher’s exact test as a measure of significance for the proportion of the total genes in each pathway that contained at least one SNP with P-value < 0.05 (see Material and methods for details). Applying Fisher’s exact test, pathway enrichment ratios for the phosphatidylinositol signaling system were significant in both cohorts (Italian: P = 1.42 × 10⁻⁵; Canadian: P = 3.45 × 10⁻⁴) with 32 out of 77 genes and 26 out of 77 genes from this pathway containing at least one SNP with P < 0.05 in the Italian and Canadian GWAS, respectively. Although hedgehog signaling also demonstrated enrichment in both datasets, ratios for this pathway were not statistically significant (Table 3).

To further dissect the LCT pathway analysis association signal for these two pathways in each dataset we examined the most significant genes underlying the signals (Figure 1). The genes most strongly driving the association for each replicated pathway differed between the two datasets pointing to the genetic heterogeneity and complexity of the disease under study and power limitations in the dataset sample numbers, especially for ascertaining genes with smaller effect size.

Figure 1.

Genes identified in the phosphatidylinositol and hedgehog signaling pathways. Genes reaching statistical significance for the phosphatidylinositol signaling system (panel a) and the hedgehog pathway (panel b) are shown by color highlights (red, Italian dataset; blue, Canadian dataset)

DISCUSSION

In this study, we used a newly developed pathway-based method, the LCT, to analyze two datasets obtained from previously completed GWAS of Italian and Canadian PBC cohorts. At the conservative cutoff for statistical significance (P < 0.05 and FDR < 0.25) that we adopted, the LCT identified eight pathways associated with the risk of development of PBC in the Italian dataset. We leveraged the availability of data from the two independent, geographically separated PBC case populations to evaluate our findings more broadly. In the interpretation of our results, we focus on those pathways that showed nominal evidence of association in both the Italian and in the Canadian dataset, and emphasize that GWAS pathway analysis is primarily a tool to generate hypothesis for further testing. Two KEGG pathways, phosphatidylinositol signaling and hedgehog signaling systems, attained the P < 0.05 level of significance on LCT pathway analysis in both datasets but did not meet the FDR < 0.25 criterion in either dataset. Both pathways were significant in the larger (Italian) dataset by a modified GSEA, further suggesting their involvement in PBC genetic susceptibility. Simple pathway enrichment ratios for phosphatidylinositol signaling also remained statistically significant in both datasets using Fisher’s exact test-Current experimental evidence offers a variety of putative mechanisms that may underpin the possible role of phosphatidylinositol and hedgehog signaling activity in PBC etiology. Each of these mechanisms serves as a potential avenue that requires follow-up functional investigation.

The phosphatidylinositol signaling system pathway is an integral component of the adaptive immune response and is essential for the maintenance of self-tolerance.²³ Phosphatidylinositol signaling is known to be a key controller of T helper 17 cell differentiation.^24–25 T helper 17 cells, a subset of helper T cells that produce interleukin 17, are major drivers of both inflammation and autoimmunity.²⁶ T helper 17 differentiation may be modulated by dendritic cell interleukin 12 (IL-12) through its effect on interferon-γ.²⁷ It has been demonstrated that dendritic cell IL-12 production is in turn positively regulated by the p110 β catalytic subunit of phosphoinositide 3-kinase (PI3K).²⁸ Interestingly, the PIK3CB gene coding for the PI3K p110β isoform, emerged as the most significant gene of the phosphatidylinositol pathway in the Italian data (P = 9.54 × 10⁻⁴, Figure 1a). The importance of IL-12 to PBC pathogenesis is highlighted by the identification of a strong and reproducible association between the IL12A and ILD2RB2 loci and disease risk in every PBC GWAS conducted thus far.^9–2 Further, aberrant signal transduction via the phosphatidylinositol system in PBC is consistent with the role of this pathway in disorders that share genetic susceptibility factors with PBC, especially rheumatoid arthritis and systemic lupus erythematosus,^29,30

Several pathways biologically related to phosphatidylinositol signaling were also uncovered using the LCT analysis at the P < 0.05 level in our data. A possible non-biological explanation for this observation is that genes common to these different pathways underlie the statistical association. However, pathway overlap cannot account entirely for this observation and there are well-established functional connections between the pathways discovered and events upstream and downstream of the phosphatidylinositol signaling system. One particularly note-worthy relationship involves TNF/stress-related signaling, the top pathway in the Italian dataset. This finding corroborates the independent discovery of seven distinct loci harboring genes related to TNF signaling and downstream Nuclear Factor-KappaB (NF-κB) signaling at the genome-wide significance level in the most recent GWAS of PBC.¹² Interactions between specific members of the TNF pathway lead to the induction of apoptosis as well as the activation of NF-κB signaling, which is anti-apoptotic and pro-inflammatory.³¹ Disturbances in this balance between cell death and survival are now recognized as being critical to PBC progression.³² Possible involvement of the phosphatidylinositol pathway in PBC thus appears to fit well with the TNF hypothesis as this signaling system has been shown to mediate the effects of TNF-α on NF-κB activation,^33,34

The hedgehog signaling pathway consists of a family of molecules that control cell-type specification during normal development and are intimately involved in tissue and organ morphogenesis.³⁵ Biliary epithelial cells are the first targets of autoimmune injury in PBC. Increased expression in biliary epithelial cells of hedgehog pathway genes and genes targeted by this pathway has previously been reported in a study of PBC patients.³⁶ Animal models of chronic cholestatic biliary injury also demonstrate activation of hedgehog signaling³⁷ and hedgehog signaling has been linked to the promotion of cholangiocyte chemokine production that may mediate recruitment of inflammatory cells in PBC.³⁸ Multiple lines of evidence suggest that the hepatic fibrosis seen in the natural history of PBC can be partly attributed to epithelial-mesenchymal transition, or to the progressive replacement of biliary epithelial cells by cells of fibroblastic lineage.³⁹ Hedgehog signaling is among the best-known effectors of epithelial-mesenchymal transition.⁴⁰ Another inducer of epithelial-mesenchymal transition, the Wnt signaling pathway (KEGG), ranked 17th among the pathways in the Canadian sample (P = 0.037, FDR = 0.693).⁴¹ Crucially, our analysis of the Italian cohort revealed that the ZIC2 gene, which is pivotal to the cross-talk between hedgehog and Wnt, was the most significant gene not only in the hedgehog pathway, but also for the dataset overall (P = 7.20 × 10⁻⁵, Supplementary Table S1).⁴² The association between Wnt signaling and PBC remains to be elucidated, though the upregulation of genes in this pathway has been reported in an early microarray study of the disorder.⁴³

It is worth noting that the two additional analyses that we used to evaluate the pathways replicated at the P < 0.05 threshold, complement the LCT. Taken together, they test the association between a set of genes and disease predisposition under some of the different underlying genetic architectures that may drive such an association. Although the LCT combines evidence of association from all SNPs that map to a gene, the modified GSEA only accounts for the top SNP signal in each gene and Fisher’s exact test for overrepresentation considers all SNPs in a gene that were nominally significant in the original GWAS.

Our study has several limitations and the results must be interpreted with some caution. The first, and a frequently cited criticism of GWAS pathway analyses in general, was the reliance on canonical pathways that represented < 30% of the total genes mapped in each dataset. However, we sought to reduce the influence of selecting canonical pathways by sourcing our pathways from two standard, manually curated databases containing well-defined pathways. Second, the annotation of protein-coding regions in the human genome is incomplete and moreover, there is substantial non-coding SNP information in intergenic regions that is now known to have both trans-effects as well as long distance cis-effects on the expression of genes in signaling pathways.⁴⁴ Using SNPs within or close to a gene to represent the gene overlooks such distant functional and regulatory relationships. Third, pathways in the Canadian dataset failed to breach the FDR level for statistical significance that we set and the FDRs of the pathways replicated in both datasets was relatively high. Possible explanations include an inadequate sample size, the behavior of the LCT statistic or the genuine absence of a stronger pathway-level association signal for the pathways tested. The sample size for the study was limited, but PBC is an uncommon disease and assembling large cohorts is difficult. Lastly, the current study was limited to common SNPs (> 5%) in the populations studied and many uncommon SNPs as well as structural variants may underlie a considerable portion of the susceptibility to this disease.

In conclusion, the linear combination method may be useful as a secondary step to single-marker analysis for mining a combination of known and novel biologically plausible disease-related pathways from GWAS data. Pathways such as TNF signaling, antigen processing and presentation, and apoptosis, each of which is an established contributor to genetic predisposition to PBC, were among the top pathways identified.¹² Two pathways, phosphatidylinositol signaling system and hedgehog signaling, were replicated at the nominal level of significance in the available datasets and these findings were backed by a complementary pathway analysis approach in at least one of the datasets. Genetic variation in these two pathways has not been frequently associated with PBC in prior work. The findings need to be validated in other independent PBC GWAS cohorts. If explored in greater depth and confirmed by future experimental studies, these results have the potential to yield new targets that may be of value for preventive intervention and therapeutic development against PBC.

MATERIALS AND METHODS

Study populations

This study included both an Italian and a Canadian cohort. All PBC cases in both GWAS met the American Association for the Study of Liver Diseases diagnostic criteria for PBC.

The Italian dataset consisted of 449 PBC cases and 940 controls of homogenous Italian descent with genotypes for 468 982 SNPs from the GWAS described in detail by Liu et al.¹¹ All retained subjects had homogeneous Italian descent genetically inferred by principal components analysis that applied specific criteria to eliminate outliers and individuals of Sardinian origin from the dataset. The cases had a mean age of 552 years, 90.3% were female, 85.4% were anti-mitochondrial antibodies- positive and 31.7% had liver cirrhosis. Stringent quality control standards were implemented as previously described and all SNPs retained had sample call rates > 95%, minor allele frequency > 0.05 and were in Hardy-Weinberg disequilibrium test P > 10⁻⁵. Pairs of subjects with cryptic relatedness as defined by an identity-by-state score > 0.1 were removed.

The Canadian sample was from the GWAS described in Hirschfield et al.⁹ and consisted of 530 PBC cases, 398 controls and 334 444 SNPs. The cases had a mean age of 60.7 years, 93% were female, 95.5% were anti-mitochondrial antibodies- positive and 5.2% had received a liver transplant Study genotyping was done at the University of Toronto using the Illumina HumanHap370 BeadChip. Single nucleotide polymorphisms with minor allele frequency < 0.01 were excluded and cryptically related individuals, who had an identity-by-state score > 0.25, were removed. Other data filtering standards were identical to the Italian GWAS.

Ethics statement

All participants in both primary studies provided written informed consent and were enrolled on protocols approved by a local Institutional Review Board or ethics committee at each center.

Linear combination test for pathway analysis

The two datasets were analyzed individually using the LCT described by Luo et al.²⁰ The test was made publicly available as part of a free software package that was used for the present pathway analysis (https://sph.uth.tmc.edu/hgc/faculty/xiong/software-A.html). The LCT provided adequate type I error rates in simulation studies that we conducted. The algorithm used raw genotype data to first compute genome-wide single-SNP association statistics. All SNPs between the start site and the 3′-untranslated region were then assigned to the gene using NCBI dbSNP Build 129 and human Genome Build 36.3. As SNPs within genes are correlated due to linkage disequilibrium, traditional methods for combining independent P-values cannot be used to bring together Single-SNP P-values for all SNPs in the gene. Therefore, to test the association of each gene with the disease, we combined P-values for all SNPs within the gene using the gene-level LCT statistic derived by Luo, et al.:²⁰

$$T_{\mathrm{L}}=\frac{e\prime Z}{\sqrt{e\prime R_{\mathrm{g}}e}}$$

where e = [1,1,…,1]^T. Z = [Z₁,…, Z_k]^T for a gene with k SNPs (given that Z₁ = Φ⁻¹ (1 − P₁) where, P₁ is the P-value of a statistic with a normal or asymptotic normal distribution), R_g is the correlation matrix of Z and T_L follows a standard normal distribution under the null hypothesis.

Genes were mapped to pathways from the BioCarta database (http://www.biocarta.com/genes/index.asp) and from the Kyoto Encyclopedia of Genes and Genomes (KEGG; http://www.genome.jp/kegg/pathway.html). The default pathway list included with the LCT software contained 299 BioCarta and 202 KEGG pathways. However, we decided to confine the final analysis to pathways containing > 10 genes to avoid testing pathways that were too small. At this stage, P-values for all genes within a pathway were combined using the pathway-level LCT statistic to test association of each pathway with the disease:

$$T_{\mathrm{P}}=\frac{e\prime T_{\mathrm{L}}}{\sqrt{e\prime R_{\mathrm{P}}e}}$$

where T_L = [T_L1,…,T_Lm]^T for a pathway with m genes, R_p is the matrix of correlations between the test statistics for all genes in the pathway and T_P is asymptotically distributed as the standard normal distribution under the null hypothesis.

Results of the LCT analysis were adjusted for multiple comparisons using FDR control by the Benjamini-Hochberg procedure.⁴⁵ Statistical significance was set at P < 0.05 and q-value or FDR < 0.25, a frequently adopted criterion in GWAS pathway studies.¹⁶ A meta-analysis of LCT pathway resulting from both datasets was also conducted using Fisher’s combined probability test.

Additional analyses for replicated pathways

To further validate pathways that were replicated at the nominal significance level (LCT pathway P < 0.05) in both datasets, we conducted additional GWAS pathway analyses on each dataset focused only on the replicated pathways using two complementary strategies.

First, we used the i-GSEA4GWAS adaptation of the classical GSEA genome-wide pathway association method.²² Classical GSEA, as in Wang et al.,⁴⁵ uses the single-SNP association test statistic for the most significant SNP in each gene to represent the gene. All genes are ranked in descending order of their test statistic value. A weighted Kolmogorov-Smirnov-like running sum statistic is calculated to determine, within a particular pathway, overrepresentation of highly ranked genes from the ranked list of all genes. The Kolmogorov-Smirnov-like statistic is normalized to account for differences in the number of genes across pathways. After this point, i-GSEA4GWAS differs from classical GSEA in that it multiplies the normalized statistic by a correction factor. This factor depends on the proportion of significant genes in a pathway and attenuates the possibility of pathways being unduly influenced by a few genes that are very highly ranked. Finally, statistical significance for pathways is calculated after phenotype label permutation in classical GSEA and SNP label permutation in i-GSEA4GWAS. Single nucleotide polymorphisms are permuted across pathways and the method provides a computationally efficient approach to follow-up results from a more comprehensive primary pathway analysis. For i-GSEA4GWAS, we (a) used single-SNP χ² GWAS analysis results from PLINK (version 1.05) for each dataset,⁴⁷ (b) tested only those pathways that replicated on LCT in both datasets, (c) used pathway definitions identical to LCT, (d) used the same rules for mapping SNPs to genes as in LCT and (e) performed 1000 SNP label permutations.

The second complementary pathway analysis strategy for validation involved determining the statistical significance of pathway enrichment ratios using Fisher’s exact test. For each dataset, all genes containing at least one SNP with P-value < 0.05 in single-SNP χ² GWAS analysis were listed. Enrichment ratio for a pathway was calculated as the number of genes in this list that map to the pathway divided by the number of genes in the pathway. As before, only pathways that replicated on LCT in both datasets were tested using pathway definitions identical to LCT. Fisher’s exact test was used to determine the probability that the association between genes in the list and genes in the pathway was explained by chance alone. Data were analyzed through the use of IPA (Ingenuity Systems, www.ingenuity.com).