Host-microbe interactions have shaped the genetic architecture of inflammatory bowel disease

We conducted an imputation-based association analysis using autosomal genotype level data from 15 GWAS of CD and/or UC (Supplementary Table 1, Supplementary Figure 1). We imputed 1.23 million SNPs from the HapMap3 reference set (Supplementary Methods), resulting in a high quality dataset with reduced genome-wide inflation (Supplementary Figures 2, 3) compared with previous meta-analyses of subsets of these data^2,3. The imputed GWAS data identified 25,075 SNPs that had association p < 0.01 in at least one of the CD, UC or all IBD analyses. A meta-analysis of GWAS data with Immunochip⁶ validation genotypes from an independent, newly-genotyped set of 14,763 CD cases, 10,920 UC cases, and 15,977 controls was performed (Supplementary Table 1, Supplementary Figure 1). Principal components analysis resolved geographic stratification, as well as Jewish and non-Jewish ancestry (Supplementary Figure 4), and significantly reduced inflation to a level consistent with residual polygenic risk, rather than other confounding effects (from λ_GC = 2.00 to λ_GC = 1.23 when analyzing all IBD samples, Supplementary Methods, Supplementary Figure 5).

Our meta-analysis of the GWAS and Immunochip data identified 193 statistically independent signals of association at genome-wide significance (p < 5×10⁻⁸) in at least one of the three analyses (CD, UC, IBD). Since some of these signals (Supplementary Figure 6) probably represent associations to the same underlying functional unit, we merged these signals (Supplementary Methods) into 163 regions, of which 71 are reported here for the first time (Table 1, Supplementary Table 2). Figure 1A shows the relative contributions of each locus to the total variance explained in UC and CD. We have increased the total disease variance explained (variance being subject to fewer assumptions than heritability⁷) from 8.2% to 13.6% in CD and from 4.1% to 7.5% in UC (Supplementary Methods). Consistent with previous studies, our IBD risk loci seem to act independently, with no significant evidence of deviation from an additive combination of log odds ratios.

Our combined genome-wide analysis of CD and UC enables a more comprehensive analysis of disease specificity than was previously possible. A model selection analysis (Supplementary Methods 1d) showed that 110/163 loci are associated with both disease phenotypes; 50 of these have an indistinguishable effect size in UC and CD, while 60 show evidence of heterogeneous effects (Table 1). Of the remaining loci, 30 are classified as CD-specific and 23 as UC-specific. However, 43 of these 53 show the same direction of effect in the non-associated disease (Figure 1B, overall p=2.8×10⁻⁶). Risk alleles at two CD loci, PTPN22 and NOD2, show significant (p < 0.005) protective effects in UC, exceptions that may reflect biological differences between the two diseases. This degree of sharing of genetic risk suggests that nearly all the biological mechanisms involved in one disease play some role in the other.

The large number of IBD associations, far more than reported for any other complex disease, increases the power of network-based analyses to prioritize genes within loci. We investigated the IBD loci using functional annotation and empirical gene network tools (Supplementary Table 2). Compared with previous analyses which identified candidate genes in 35% of loci^2,3 our updated GRAIL⁸ -connectivity network identifies candidates in 53% of loci, including increased statistical significance for 58 of the 73 candidates from previous analyses. The new candidates come not only from genes within newly identified loci, but also integrate additional genes from previously established loci (Figure 1C). Only 29 IBD-associated SNPs are in strong linkage disequilibrium (r² > 0.8) with a missense variant in the 1000 Genomes Project data, which reinforces previous evidence that a large fraction of risk for complex disease is driven by non-coding variation. In contrast, 64 IBD-associated SNPs are in linkage disequilibrium with variants known to regulate gene expression (Supplementary Table 2). Overall, we highlighted a total of 300 candidate genes in 125 loci, of which 39 contained a single gene supported by two or more methods.

Seventy percent (113/163) of the IBD loci are shared with other complex diseases or traits, including 66 among the 154 loci previously associated with other immune-mediated diseases⁹, which is 8.6 times the number that would be expected by chance (Figure 2A, p < 10⁻¹⁶, Supplementary Figure 7). Such enrichment cannot be attributed to the immune-mediated focus of the Immunochip, (Supplementary Methods 4a(i), Supplementary Figure 8), since the analysis is based on our combined GWAS-Immunochip data. Comparing overlaps with specific diseases is confounded by the variable power in studies of different diseases. For instance, while type 1 diabetes (T1D) shares the largest number of loci (20/39, 10-fold enrichment) with IBD, this is partially driven by the large number of known T1D associations. Indeed, seven other immune-mediated diseases show stronger enrichment of overlap, with the largest being ankylosing spondylitis (8/11, 13-fold) and psoriasis (14/17, 14-fold).

IBD loci are also markedly enriched (4.9-fold, p < 10⁻⁴) in genes involved in primary immunodeficiencies (PIDs, Figure 2A), which are characterized by a dysfunctional immune system resulting in severe infections¹⁰. Genes implicated in this overlap correlate with reduced levels of circulating T-cells (ADA, CD40, TAP1/2, NBS1, BLM, DNMT3B), or of specific subsets such as Th17 (STAT3), memory (SP110), or regulatory T-cells (STAT5B). The subset of PIDs genes leading to Mendelian susceptibility to mycobacterial disease (MSMD)^10–12 is enriched still further; six of the eight known autosomal genes linked to MSMD are located within IBD loci (IL12B, IFNGR2, STAT1, IRF8, TYK2 and STAT3, 46-fold enrichment, p = 1.3 × 10⁻⁶), and a seventh, IFNGR1, narrowly missed genome-wide significance (p = 6 × 10⁻⁸). Overlap with IBD is also seen in complex mycobacterial disease; we find IBD associations in 7/8 loci identified by leprosy GWAS¹³, including 6 cases where the same SNP is implicated. Furthermore, genetic defects in STAT3^14–15 and CARD9¹⁶, also within IBD loci, lead to PIDs involving skin infections with staphylococcus and candidiasis, respectively. The comparative effects of IBD and infectious disease susceptibility risk alleles on gene function and expression is summarized in Supplementary Table 3, and include both opposite (e.g. NOD2 and STAT3, Supplementary Figure 9) and similar (e.g., IFNGR2) directional effects.

To extend our understanding of the fundamental biology of IBD pathogenesis we conducted searches across the IBD locus list: (i) for enrichment of specific GeneOntology (GO) terms and canonical pathways, (ii) for evidence of selective pressure acting on specific variants and pathways, and (iii) for enrichment of differentially expressed genes across immune cell types. We tested the 300 prioritized genes (see above) for enrichment in GO terms (Supplementary Methods) and identified 286 GO terms and 56 pathways demonstrating significant enrichment in genes contained within IBD loci (Supplementary Table 4, Supplementary Figure 10,11). Excluding high-level GO categories such as “immune system processes” (p = 3.5 × 10⁻²⁶), the most significantly enriched term is regulation of cytokine production (p=2.7×10⁻²⁴), specifically IFNG-γ, IL-12, TNF-α, and IL-10 signalling. Lymphocyte activation was the next most significant (p=1.8 × 10⁻²³), with activation of T-, B-, and NK-cells being the strongest contributors to this signal. Strong enrichment was also seen for response to molecules of bacterial origin (p=2.4 × 10⁻²⁰), and for KEGG’s JAK-STAT signalling pathway (p = 4.8 × 10⁻¹⁵). We note that no enriched terms or pathways showed specific evidence of CD- or UC-specificity.

As infectious organisms are known to be among the strongest agents of natural selection, we investigated whether the IBD-associated variants are subject to selective pressures (Supplementary Methods, Supplementary Table 5). Directional selection would imply that the balance between these forces shifted in one direction over the course of human history, whereas balancing selection would suggest an allele frequency dependent-scenario typified by host-microbe co-evolution, as can be observed with parasites. Two SNPs show Bonferroni-significant selection: the most significant signal, in NOD2, is under balancing selection (p = 5.2 × 10⁻⁵), and the second most significant, in the receptor TNFRSF18, showed directional selection (p = 8.9 × 10⁻⁵). The next most significant variants were in the ligand of that receptor, TNFSF18 (directional, p = 5.2 × 10⁻⁴), and IL23R (balancing, p = 1.5 × 10⁻³). As a group, the IBD variants show significant enrichment in selection (Figure 2B) of both types (p = 5.5 × 10⁻⁶). We discovered an enrichment of balancing selection (Figure 2B) in genes annotated with the GO term “regulation of interleukin-17 production” (p = 1.4 × 10⁻⁴). The important role of IL17 in both bacterial defense and autoimmunity suggests a key role for balancing selection in maintaining the genetic relationship between inflammation and infection, and this is reinforced by a nominal enrichment of balancing selection in loci annotated with the broader GO term “defense response to bacterium” (p = 0.007).

We tested for enrichment of cell-type expression specificity of genes in IBD loci in 223 distinct sets of sorted, mouse-derived immune cells from the Immunological Genome Consortium¹⁷. Dendritic cells showed the strongest enrichment, followed by weaker signals that support the GO analysis, including CD4+ T, NK and NKT cells (Figure 2C). Notably, several of these cell types express genes near our IBD associations much more specifically when stimulated; our strongest signal, a lung-derived dendritic cell, had p _stimulated < 1×10⁻⁶ compared with p _unstimulated = 0.0015, consistent with an important role for cell activation.

To further our goal of identifying likely causal genes within our susceptibility loci and to elucidate networks underlying IBD pathogenesis, we screened the associated genes against 211 co-expression modules identified from weighted gene co-expression network analyses¹⁸, conducted with large gene expression datasets from multiple tissues^19–21. The most significantly enriched module comprised 523 genes from omental adipose tissue collected from morbidly obese patients¹⁹, which was found to be 2.9-fold enriched for genes in the IBD-associated loci (p = 1.1 × 10⁻¹³, Supplementary Table 6, Supplementary Figure 12). We constructed a probabilistic causal gene network using an integrative Bayesian network reconstruction algorithm^22–24 which combines expression and genotype data to infer the direction of causality between genes with correlated expression. The intersection of this network and the genes in the IBD-enriched module defined a sub-network of genes enriched in bone marrow-derived macrophages (p < 10⁻¹⁶) and is suggestive of dynamic interactions relevant to IBD pathogenesis. In particular, this sub-network featured close proximity amongst genes connected to host interaction with bacteria, notably NOD2, IL10, and CARD9.

A NOD2-focused inspection of the sub-network prioritizes multiple additional candidate genes within IBD-associated regions. For example, a cluster near NOD2 (Figure 2D) contains multiple IBD genes implicated in M.tb response, including SLC11A1, VDR and LGALS9. Furthermore, both SLC11A1 (also known as NRAMP1) and VDR have been associated with M.tb infection by candidate gene studies^25–26, and LGALS9 modulates mycobacteriosis²⁷. Of interest, HCK (located in our new locus on chromosome 20 at 30.75Mb) is predicted to upregulate expression of both NOD2 and IL10, an anti-inflammatory cytokine associated with Mendelian²⁸ and non-Mendelian IBD²⁹. HCK has been linked to alternative, anti-inflammatory activation of monocytes (M2 macrophages)³⁰; while not identified in our aforementioned analyses, these data implicate HCK as the causal gene in this new IBD locus.

We report one of the largest genetic experiments involving a complex disease undertaken to date. This has increased the number of confirmed IBD susceptibility loci to 163, most of which are associated with both CD and UC, and is substantially more than reported for any other complex disease. Even this large number of loci explains only a minority of the variance in disease risk, which suggests that other factors such as rarer genetic variation not captured by GWAS or environmental exposures make substantial contributions to pathogenesis. Most of the evidence relating to possible causal genes points to an essential role for host defence against infection in IBD. In this regard the current results focus ever closer attention on the interaction between the host mucosal immune system and microbes both at the epithelial cell surface and within the gut lumen. In particular, they raise the question, in the context of this burden of IBD susceptibility genes, as to what triggers components of the commensal microbiota to switch from a symbiotic to a pathogenic relationship with the host. Collectively, our findings have begun to shed light on these questions and provide a rich source of clues to the pathogenic mechanisms underlying this archetypal complex disease.