Transcriptome-Wide Association Study for Inflammatory Bowel Disease Reveals Novel Candidate Susceptibility Genes in Specific Colon Subsites and Tissue Categories

Abstract

Background and Aims

Genome-wide association studies [GWAS] for inflammatory bowel disease [IBD] have identified 240 risk variants. However, the benefit of understanding the genetic architecture of IBD remains to be exploited. Transcriptome-wide association studies [TWAS] associate gene expression with genetic susceptibility to disease, providing functional insight into risk loci. In this study, we integrate relevant datasets for IBD and perform a TWAS to nominate novel genes implicated in IBD genetic susceptibility.

Methods

We applied elastic net regression to generate gene expression prediction models for the University of Barcelona and University of Virginia RNA sequencing project [BarcUVa-Seq] and correlated expression and disease association research [CEDAR] datasets. Together with Genotype-Tissue Expression project [GTEx] data, and GWAS results from about 60 000 individuals, we employed Summary-PrediXcan and Summary-MultiXcan for single and joint analyses of TWAS results, respectively.

Results

BarcUVa-Seq TWAS revealed 39 novel genes whose expression in the colon is associated with IBD genetic susceptibility. They included expression markers for specific colon cell types. TWAS meta-analysis including all tissues/cell types provided 186 novel candidate susceptibility genes. Additionally, we identified 78 novel susceptibility genes whose expression is associated with IBD exclusively in immune (N = 19), epithelial (N = 25), mesenchymal (N = 22) and neural (N = 12) tissue categories. Associated genes were involved in relevant molecular pathways, including pathways related to known IBD therapeutics, such as tumour necrosis factor signalling.

Conclusion

These findings provide insight into tissue-specific molecular processes underlying IBD genetic susceptibility. Associated genes could be candidate targets for new therapeutics and should be prioritized in functional studies.

Graphical Abstract

Graphical Abstract.

1. Introduction

Inflammatory bowel disease [IBD] is a chronic inflammatory disorder of the gastrointestinal tract that encompasses two main disease subtypes, namely Crohn’s disease [CD] and ulcerative colitis [UC]. IBD is caused by immune dysregulation and aberrant inflammatory responses to gut microbiota that result in tissue damage. Clinical manifestations of CD are more heterogeneous than those of UC. UC is restricted to the large intestine, whereas CD can affect any part of the gastrointestinal tract and involves the colon in only 25% of cases.^1,2

Germline genetic variants have been associated with IBD susceptibility. The largest genome-wide association study [GWAS] for IBD identified 240 independent risk single nucleotide polymorphisms [SNPs].³ Some of them have been functionally characterized and found to affect established mechanisms of IBD pathogenesis, including impaired autophagy, interleukin [IL]-17/IL-23 axis/type 3 innate lymphoid cells, and failure to suppress aberrant immune responses.⁴ GWAS SNPs have also been associated with genes whose encoded proteins participate in pathways targeted by approved IBD therapies such as infliximab and adalimumab, which are monoclonal antibodies that modulate tumour necrosis factor [TNF] signalling.³ Despite these successes the mechanisms by which GWAS-identified genetic variants, especially non-coding variants, confer susceptibility are not yet fully understood.²

The hypothesis that risk SNPs modify expression of nearby genes and influence development of disease is supported by recent work in multiple tissues.² One recently published study⁵ presented a large gene expression dataset from normal colon tissue and showed strong evidence that genetically regulated gene expression in the colon is involved in IBD genetic susceptibility. Another study⁶ related genetic risk variants to gene expression in circulating immune cells to identify IBD susceptibility genes.

Sequencing RNA from multiple tissues/cell types of thousands of subjects with and without IBD to associate gene expression with disease is costly and not feasible for some tissues. In addition, this approach cannot distinguish whether altered gene expression is a cause rather than a consequence of disease. A solution to these limitations is provided by the transcriptome-wide association study [TWAS] statistical approach, which permits prediction of gene expression from genetic data, and thereby enables imputation of gene expression for subjects included in GWAS. TWAS uses reference imputation panels (i.e. predictive models generated from population-based germline genotype and tissue-specific gene expression data) to associate genetically regulated gene expression with traits and diseases. The TWAS approach provides biological context for interpreting disease risk loci by nominating candidate susceptibility genes not only at GWAS risk regions but also at other potential regions that current GWAS have not been powered to detect.⁷

Previous TWAS for IBD^8–10 reported candidate susceptibility genes based on prior versions of the Genotype-Tissue Expression project [GTEx],¹¹ and the only study that analysed the latest version (v8) of GTEx⁹ did not provide results for CD. The University of Barcelona and University of Virginia RNA sequencing project [BarcUVa-Seq]⁵ recently provided a gene expression dataset of colon biopsies across colon subsites. Another source of relevant data for autoimmune diseases is the array-derived correlated expression and disease association research [CEDAR]⁶ dataset, which includes gene expression data of circulating immune cell types. To the best of our knowledge, no published study has utilized these reference panels to perform a joint TWAS (i.e. a TWAS meta-analysis that combines TWAS results of individual tissues for increasing the statistical power to find associations) across tissues to strengthen the evidence for genes involved in IBD susceptibility.

In this study, we perform an integrative TWAS analysis to identify novel candidate susceptibility genes whose expression influences IBD pathogenesis. This study includes reference datasets of tissues and blood cell types relevant to IBD and leverages GWAS results for IBD from a dataset including about 25 000 cases and 35 000 controls.³ We nominate genes whose expression in more than 60 tissues and cell types, including specific colon anatomical subsites (ascending, transverse and descending colon), is associated with IBD and its subtypes (CD and UC). Finally, we assess associations specific for epithelial, immune, mesenchymal and neural tissue categories.

2. Materials and Methods

2.1. GWAS summary statistics

We downloaded publicly available IBD, UC and CD GWAS summary statistics from a large study including about 60 000 subjects.³ We performed liftover of SNP coordinates to the GRCh38 genome reference using Crossmap.¹² Reference SNP cluster IDs [rsIDs] were annotated according to dbSNP v151 to match IDs from reference panels.

2.2. BarcUVa-Seq data processing

BarcUVa-Seq data⁵ include genome-wide genotypes and gene expression from ascending (n = 138), transverse (n = 143) and descending (n = 164) colon. Expression data were processed as described elsewhere.⁵ The GENCODE v26 gene model¹³ was used to facilitate integration with GTEx v8 data. Genotypes were imputed with the TOPMed (version r2) reference panel on the Michigan Imputation Server.¹⁴ SNPs were filtered by minor allele frequency [MAF] 0.01 and imputation quality (i.e. R²) 0.8. For each panel, we assessed population heterogeneity using 2318 ancestry-informative marker SNPs with the plink pca method.¹⁵

2.3. CEDAR data processing

CEDAR data⁶ were obtained from the Array Express repository under accession numbers E-MTAB-6666 and E-MTAB-6667 for genotypes and expression data, respectively. The data include gene expression from terminal ileum, transverse colon, rectum, platelets, CD15⁺ granulocytes, CD19⁺ B lymphocytes, CD8⁺ T lymphocytes, CD4⁺ T lymphocytes and CD14⁺ monocytes. Corresponding sample sizes are provided in Supplementary Table 1. Expression arrays were processed with the iluminaio R package.¹⁶ Expression variability between samples was assessed with graphical visualization of expression values in box plots to ensure that no extreme outliers appeared in the dataset. Quantile normalization was performed. Gene annotation was harmonized to GENCODE v26 annotations¹³ to facilitate integration with GTEx v8 data. Genotypes were imputed with the Haplotype Reference Consortium panel on the Michigan Imputation Server,¹⁴ and lifted over to the GRCh38 genome reference with Crossmap.¹² We filtered SNPs by MAF 0.01 and imputation quality (i.e. R²) 0.8. For each panel, we assessed population heterogeneity using 2318 ancestry-informative marker SNPs with the plink pca method.¹⁵

2.4. Gene expression prediction models

We downloaded GTEx v8 elastic net regularized regression-based imputation panels (N = 49 tissues/cell types) from PredictDB.^11,17 We generated gene expression prediction models using gastrointestinal tissue and blood cell gene expression data from BarcUVa-Seq (ascending, transverse and descending and ‘any’ colon, where ‘any’ includes all three subsites) and CEDAR (terminal ileum, transverse colon, rectum, platelets, CD15⁺ granulocytes, CD19⁺ B lymphocytes, CD8⁺ T lymphocytes, CD4⁺ T lymphocytes and CD14⁺ monocytes) datasets, using elastic net regularized regression. CEDAR gene expression was adjusted by sex, age and sequencing batch. BarcUVa-Seq gene expression was adjusted for sex, sequencing batch, probabilistic estimation of expression residuals [PEER] factors¹⁸ and genetic ancestry (two principal components). To be consist with the PredictDB pipeline followed by the GTEx team for generating the GTEx v8 models,^11,17 we considered significant gene models as those with a predictive performance p < 0.05 and R² > 0.1. Summary statistics and SNP weights of BarcUVa-Seq prediction models were loaded into the Colon Transcriptome Explorer [CoTrEx] 2.0 web resource.¹⁹ Altogether, we compiled a total of 62 reference imputation panels of expression prediction models with a median of 4848 significant genes per panel (ranging from 1003 to 10 013 genes) (Supplementary Table 1). As expected, the number of significant prediction models increased with the sample size of the imputation panels.

2.5. Transcriptome-wide association analyses

The TWAS approach, in a first step, predicts gene expression from genotype data of subjects from whom gene expression has not been measured. This is achieved thanks to tissue-specific gene expression prediction models, i.e. reference imputation panels (see previous subsection ‘Gene expression prediction models’). Next, the inferred gene expression is tested for association with a particular phenotype (e.g. IBD). The Summary-PrediXcan (S-PrediXcan) method²⁰ used in this study combines the last two steps into one, and therefore does not need individual-level genotype data; instead, it uses the summary parameters of the statistical association between SNPs and the phenotype of interest, commonly referred to as ‘summary statistics’ (see 2.1. Methods sub-section on the summary statistics we used). Along with GWAS summary statistics, it uses the SNP expression weights to impute the expression of a given gene; and uses the variance and covariances of the included SNPs to correct for linkage disequilibrium (LD) biases.²⁰ Specifically, S-PrediXcan computes a Z-score (Wald statistic) as a measure of the association between predicted gene expression and a phenotype. The main analytical expression used is as follows:

$${\displaystyle \begin{array}{l}{\displaystyle Z_g\approx {\displaystyle \sum _{l\in Mode{l}_g}}{\omega }_{\lg }\frac{{{\displaystyle \widehat{\sigma }}}_l}{{{\displaystyle \widehat{\sigma }}}_g}\frac{{{\displaystyle \widehat{\beta }}}_l}{se\left({{\displaystyle \widehat{\beta }}}_l\right)}}\end{array}}$$

where ω_lg is the weight of SNP l in the prediction of the expression of gene g; ${\widehat{\beta }}_l$ is the GWAS coefficients for SNP l; $se\left({\widehat{\beta }}_l\right)$ is the standard error of $\widehat{\beta },$${\widehat{\sigma }}_l$ is the estimated variance of SNP l, and ${\widehat{\sigma }}_g$ is the estimated variance of the predicted expression of gene g.²⁰ We considered as significant those genes that passed Bonferroni correction (0.05/total number of genes).

On the other hand, we used Summary-MultiXcan (S-MultiXcan)²¹ for the joint analysis of TWAS results across multiple tissues. Briefly, MultiXcan consists of fitting a linear regression of the phenotype on predicted expression from multiple tissue models jointly.²¹ On a similar basis to the S-PrediXcan approach explained above, the MultiXcan framework was extended to be used with GWAS summary statistics. Specifically, S-MultiXcan combines single-tissue S-PrediXcan results, along with LD information from a reference panel for the estimation of their joint effect across tissues on the phenotype.²¹ We considered as significant only those genes that passed Bonferroni correction and that had a p-value ≤ 10^–4 in the panel with lowest p, as advised elsewhere,²¹ to minimize errors due to LD mismatches.

The categorization of expression panels as epithelial, immune/blood, mesenchymal and neural categories was based on their histological origin (for CEDAR and BarcUVa-Seq datasets), and on their classification above the third quartile of the corresponding categories established by Breschi et al.²² (for GTEx v8 datasets).

2.6. Gene annotation

Gene symbols were annotated according to the HUGO Gene Nomenclature Committee.²³ Genes were annotated as novel if they did not appear in the GWAS catalogue genes for IBD,²⁴ were not indicated in large GWAS previously published elsewhere,^3,25 or did not appear in the TWAS-hub resource for IBD.⁸ Genes were annotated at GWAS loci if their transcription start sites were within 1 Mb of any of the top 240 SNPs identified by IBD GWAS.³ In the case of significant genes predicted using BarcUVa-Seq colon panels, we annotated the cells for which genes were expression markers according to a study by Smillie et al. that characterized the colon transcriptome at single-cell resolution.²⁶

2.7. Fine-mapping

In the context of a TWAS, the fine-mapping approach aims to prioritize candidate genes with higher likelihood of being causal for the association. This is especially important for TWAS-associated loci with multiple genes, where the correlation of expression between genes tend to be high and which might bias the results, in a similar manner as LD does with GWAS-identified SNPs. To address this topic, probabilistic fine-mapping was performed using the fine-mapping of causal gene sets [FOCUS] approach.²⁷ FOCUS provides fine-mapping at each of the TWAS-identified loci by integrating GWAS summary statistic data, the SNP expression weights for each tissue and LD-related statistics among all SNPs in each locus. Specifically, it applies a probabilistic framework to assign to every gene in a given TWAS-associated locus a posterior probability [PIP] that indicates the likelihood of a given gene to explain the observed TWAS association signal.²⁷ We used the FOCUS software with default parameters and provided FOCUS with genes passing Bonferroni correction in TWAS analyses, and considered as probably causal those genes included in a credible set with a nominal confidence of 90% and with a PIP > 0.5.

2.8. Pathway enrichment analysis

We included signalling and regulatory pathways from the Pathway Interaction Database²⁸ in pathway enrichment analysis. Enrichment was measured by hypergeometric tests. We only reported pathways that had an enrichment q value <0.05 [false discovery rate computed with the Benjamini–Hochberg method].

2.9. Data availability statement

The data underlying this article were derived from sources in the public domain. IBD, CD and UC summary statistics are available at ftp://ftp.sanger.ac.uk/pub/project/humgen/summary_statistics/human/2016-11-07/; GTEx v8-derived gene expression prediction models are available in Zenodo, at https://dx.doi.org/10.5281/zenodo.3519321; BarcUVa-Seq-derived prediction models are available in the Colon Transcriptome Explorer version 2.0, at https://barcuvaseq.org/cotrex/; and CEDAR data were obtained from the Array Express repository, under accession numbers E-MTAB-6666 and E-MTAB-6667 for genotypes and expression data, respectively.

3. Results

3.1. Transcriptome-wide associations

We evaluated associations between genetically regulated gene expression and IBD, CD and UC status, separately for each tissue/blood cell type. In the TWAS for IBD we found significant association for a median of 62 genes per tissue/cell type [ranging from ten to 124]. As expected, the number of significant associations increased with the sample size of the imputation panel. Also, we found fewer associated genes in the tissues/blood cell types of the CEDAR dataset, which was based on expression arrays to profile gene expression, than in the BarcUVa-Seq and GTEx datasets, which were based on RNA-seq. We found CD4⁺, CD14⁺ and CD19⁺ cells, rectum tissue, and BarcUVa-Seq transverse colon tissue among the tissues/cell types with the highest percentage of genes significantly associated with IBD. A summary of TWAS results for the three IBD phenotypes is provided in Supplementary Table 2. Complete TWAS results in all tissues and cell types for IBD, CD and UC phenotypes are provided in Supplementary Data 1.

3.2. BarcUVa-Seq colon TWAS

TWAS results generated with BarcUVa-Seq-derived panels [ascending, transverse, descending and any colon] are summarized in Table 1 and shown in Figure 1. We found 124 unique candidate susceptibility genes, including 39 that were novel (i.e. not reported in other large association studies see Methods]). Among the 81 and 57 genes associated with CD and UC, respectively, we found 26 shared genes, and 55 and 31 genes specific for each disease subtype, respectively. CD-specific genes included Liver Enriched Antimicrobial Peptide 2 [LEAP2], and Ubiquitin D [UBD], both novel and specific for descending colon. UC-specific genes included Tripartite Motif Containing 31 [TRIM31], which was specific to ascending colon, and Abhydrolase Domain Containing 11 [ABHD11], which was specific to descending colon. We provide complete annotated results for BarcUVa-Seq candidate susceptibility genes in Supplementary Data 2.

Table 1.

Summary of candidate susceptibility genes whose genetically regulated expression in the colon is associated with IBD

Phenotype [n unique genes]	Colon subsite	Genes [site-specific]	Novel genes	Expression markers of cell types
IBD [86]	Ascending	39 [6]	5	11
	Transverse	43 [6]	10	14
	Descending	37 [8]	6	8
	All colon	62 [16]	11	22
CD [81]	Ascending	35 [10]	7	8
	Transverse	30 [7]	9	5
	Descending	36 [13]	7	6
	All colon	49 [14]	9	12
UC [57]	Ascending	25 [6]	5	10
	Transverse	28 [6]	8	9
	Descending	21 [5]	3	6
	All colon	38 [12]	9	12
Overall unique elements		124	39	33

Figure 1.

Volcano plots showing a summary of TWAS for IBD in [A] ascending, [B] transverse and [C] descending colon. Positive effect size indicates that higher gene expression is associated with higher IBD risk. Coloured points indicate genes passing Bonferroni correction [ascending n = 39, transverse n = 43, descending n = 37]. Darker blue indicates novelty [not previously reported genes]. The top ten genes with strongest evidence [lowest p] are labelled.

To identify cell types within the colon likely to mediate genetic susceptibility to IBD, we intersected lists of candidate susceptibility genes from BarcUVa-Seq TWAS with lists of expression marker genes of specific cell types derived from colon single cell RNA sequencing [scRNA-Seq] profiles.²⁶ We found 33 candidate susceptibility genes were markers for a total of 28 cell types across colon subsites [Figure 2A] and IBD phenotypes [Figure 2B]. Cell types were categorized into epithelial, fibroblast, endothelial, myeloid, T cell and B cell types [see Methods for annotation details]. The candidate susceptibility genes identified in ascending and transverse colon TWAS and in the TWAS for UC were more frequently markers of specific cell types than susceptibility genes identified in descending colon TWAS and in the TWAS for CD, respectively [see Figure 2]. Among these findings, we found ten novel candidate susceptibility cell marker genes, which are described in Table 2. These included two fibroblast markers, three markers of myeloid cell types [e.g. inflammatory monocyte], four markers of epithelial cell types [such as M cell, goblet cell and enterocyte] and two markers of T cells. All participate in immune-related pathways except for CLDN4, which participates in epithelial tight junction maintenance. Complete cell type annotation of BarcUVa-Seq-derived candidate susceptibility genes is provided in Supplementary Table 3.

Figure 2.

Distribution of candidate susceptibility genes that are markers of expression of specific cell types in the colon. [A] Distribution by colon anatomical subsite; [B] distribution by IBD subtype [CD, UC].

Table 2.

Summary of genes whose genetically regulated expression in the colon is associated with IBD. Subset of ten genes reported as expression markers of cell types in the colon and that have not previously been reported as IBD susceptibility genes

Phenotype	Colon subsite	Locus	Gene symbol	Gene name	GWAS SNP [p value]	Cell type	Cell category	Pathway	Z score	TWAS p
IBD	Transverse	1q21.3	CTSS	Cathepsin S	rs17800987 [1.07E-16]	Inflammatory monocytes	Myeloid	Trafficking and processing of endosomal Toll-like receptors	−4.73	2.27E-06
IBD	Colon	1p36.12	WNT4	Wnt Family Member 4	rs12568930 [1.00E-17]	WNT2B+ Fos-lo 2 cells	Fibroblasts	WNT ligand biogenesis and trafficking	−6.71	2.01E-11
UC	Colon				rs34920465 [9.01E-16]				−5.99	2.08E-09
CD	Colon	5p13.1	C7	Complement C7	rs6451494 [8.258E-56]	RSPO3+ cells, WNT2B+ Fos-lo 2	Fibroblasts	Terminal pathway of complement	4.75	2.04E-06
CD	Descending	6p22.1	UBD	Ubiquitin D	rs11859512 [4.76E-13]	Microfold cells	Epithelial	—	4.33	1.47E-05
UC	Ascending	6p22.1	TRIM31	Tripartite Motif Containing 31	rs2270191 [1.139E-22]	Goblet cells, enterocytes, immature enterocytes	Epithelial	Interferon gamma signalling	-4.70	2.58E-06
IBD	Colon	6p21.32	HLA-DOB	Major Histocompatibility Complex, Class II, DO Beta	rs6927022 [5.00E-133]	Cycling B cells, germinal centre [GC] cells, follicular cells; dendritic cells	B cells; myeloid	MHC class II antigen presentation	−4.96	7.19E-07
UC	Colon				rs9271176 [4.20E-91]				−5.52	3.35E-08
IBD	Transverse	6p21.32	PSMB9	Proteasome 20S Subunit Beta 9	rs6927022 [5.00E-133]	CD8+ lamina propria [LP] cells	T cells	Proteasome	−4.50	6.90E-06
	Descending								−5.64	1.74E-08
	Colon								−5.23	1.71E-07
CD	Descending				rs185605448 [4.84E-04]				−4.82	1.43E-06
	Colon								-4.46	8.27E-06
UC	Colon	7q11.23	CLDN4	Claudin 4	rs11981405 [1.77E-07]	Transit-amplifying [TA] cells, immature goblet, immature enterocytes, stem cells, secretory TA, enterocytes, goblet, Best4+ enterocytes, enterocyte progenitors	Epithelial	Tight junction	4.96	6.98E-07
IBD	Colon	9q34.3	C8G	Complement C8 Gamma Chain	rs10781499 [4.00E-56]	Enterocytes	Epithelial	Terminal pathway of complement	–5.11	3.27E-07
IBD	Colon	14q13.2	NFKBIA	NFKB Inhibitor Alpha	rs2384352 [3.12E-13]	Tregs, innate lymphoid cells [ILCs]; CD69+ Mast	T cells; myeloid	TNF signalling	4.85	1.22E-06

GWAS SNP refers to the SNP identified by GWAS³ with lowest p value among those located up to 1 Mb from the TSS of the associated gene. The pathway with the lowest enrichment q value [computed by a hypergeometric test] is indicated. If the enrichment q value is ≥ 0.05 the pathway is not provided. REACTOME and KEGG pathway sources were used. Gene symbols in bold refer to genes that were reported as differentially expressed between patient-derived and healthy colon biopsies.³⁰ Positive Z scores indicate that higher gene expression is associated with higher IBD risk.

To further support our findings, we sought overlap between candidate susceptibility genes described in this section and IBD-associated genes reported in other studies from IBD patient-derived colon biopsies. We found that seven of the 38 significantly associated genes identified in BarcUVa-Seq colon TWAS for UC were reported as differentially expressed between biopsies from treatment-naïve UC patients [n = 14] and healthy biopsy samples [n = 16].²⁹ These included genes at 1q23.3 [FCGR3B], 6p21.32 [HLA-DRB1, HLA-DQB1, TAP2 and HLA-DOB], 6p21.33 [MICB] and 6p22.1 [TRIM40] loci. Importantly, directions of effect were concordant across studies [i.e. TWAS and differential expression] for five of seven genes [all except HLA-DOB and MICB].

Finally, to test for consistency of results across datasets from lower intestinal tissues, we correlated the predicted effect of the TWAS associations between tissues [see Figure 3]. As expected, we found the lowest correlations for the CEDAR dataset, which might be due to the technology used for assessing gene expression [arrays] in contrast to RNA-Seq, used by GTEx and BarcUVa-Seq. We found high correlations between BarcUVa-Seq-derived effects [all sites] and GTEx transverse colon-derived effects [r ≥ 0.75 in all three IBD phenotypes]. A lower correlation of results with GTEx sigmoid colon might be due to the higher component of muscularis tissue than of epithelial tissue present in the samples of this dataset.²²

Figure 3.

Replication of TWAS results in lower intestinal tissues. Correlation of the predicted effect of gene expression across tissues identified in TWAS for [A] IBD, [B] CD and [C] UC. Hierarchical clustering of tissues is shown. Correlation values are indicated by the colour scale.

3.3. Joint analyses of TWAS results

To gain more power for discovery we performed a meta-analysis of all TWAS results obtained separately for IBD, CD and UC [summarized in Table 3]. In these joint analyses, we combined the TWAS results of all tissues/cell types [see Methods] and found 466, 395 and 290 significant genes for IBD, CD and UC risk, respectively, comprising 596 unique candidate susceptibility genes. These findings included 186 novel genes (i.e. not reported in other large association studies [see Methods]). Overall, we found candidate susceptibility genes nearby (i.e. with the gene Transcription Start Site [TSS] within 1 Mb of) 106 of the 240 top SNPs reported in IBD GWAS.³ The Manhattan plot for IBD TWAS is shown in Figure 4. The most significant association is with Endosome Associated Trafficking Regulator 1 [ENTR1] [p = 8.27 × 10^–61]. We found 85 unique signalling and regulatory pathways significantly enriched in significantly associated genes [Supplementary Table 4]. Among these pathways we found IL-12, IL-23, integrins and TNF-related pathways, which have high therapeutic relevance for IBD. Novel genes in these therapeutic pathways are summarized in Table 4.

Table 3.

Summary of significant candidate susceptibility genes identified in the joint analyses of TWAS results across all tissues/cell types

Disease subtype	Genes	Genes at GWAS loci	Novel genes	Fine-mapped genes
IBD	466	388	136	50
CD	395	32	116	44
UC	290	27	88	31
Unique elements	596	440	186	47

Genes: significant genes passing Bonferroni correction and with lowest individual p ≤ 1E-4; GWAS loci: within 1 Mb of any top SNP found at corresponding GWAS; Novel: not reported in other large genome-wide association studies [see Methods]; Fine-mapped: genes included in fine-mapping credible sets and with >50% probability of being causal in their given signal.

Figure 4.

Manhattan plot of TWAS joint analysis for IBD. Each point represents a gene. Genes significantly associated are coloured. Novel genes with p < 1E-16 are labelled.

Table 4.

Summary of novel genes involved in signalling pathways of high therapeutic relevance for IBD

Gene symbol	Gene name	Locus	GWAS SNP	TWAS P	Mean Z	Pathway	Drug name[s]
HLA-A	Major Histocompatibility Complex, Class I, A	6p22.1	rs10826797 [3.99E-13]	1.05E-06	1.72	IL12-mediated signalling events	Ustekinumab
MAP4K4	Mitogen-Activated Protein Kinase Kinase Kinase Kinase 4	2q11.2	rs13001325 [2.51E-23]	8.82E-07	−1.09	TNF receptor signalling pathway	Infliximab, adalimumab
TRAF2	TNF Receptor Associated Factor 2	9q34.3	rs10781499 [4.00E-56]	1.78E-19	−2.44	TNF receptor signalling pathway	Infliximab, adalimumab
COL11A2	Collagen Type XI Alpha 2 Chain	6p21.32	rs6927022 [5.00E-133]	4.94E-06	1.05	Beta1 integrin cell surface interactions	Vedolizumab

GWAS SNP refers to the SNP identified by GWAS³ with lowest p value among those located up to 1 Mb from the TSS of the associated gene. Pathway refers to a signalling pathway in which the gene is significantly enriched [q value < 0.05] and which is related to the drug indicated.

We next performed fine-mapping of significantly associated genes to prioritize those with high probability of explaining the association signal in loci where multiple genes were identified. We found 50, 44 and 31 fine-mapped genes for IBD, CD and UC, respectively, comprising a total of 47 unique genes. Of these, we identified six novel genes [Supplementary Table 5], including five protein coding genes and one long non-coding RNA gene. These include genes that participate in the complement immune response.

3.4. Category-specific joint analyses

Next, to identify genes that participate in IBD susceptibility-related molecular mechanisms exclusively in specific tissue categories, we performed joint analyses combining different sets of TWAS results [see Methods] into epithelial [n = 18], immune/blood [n = 14], mesenchymal [n = 11] and neural [n = 15] categories based on histological and transcriptional characteristics [details in Methods and Supplementary Table 1]. Most significant genes found in these analyses had been previously identified in the joint analyses with all TWAS results [described in the previous section], but these tissue-category stratified joint analyses reported 93, 101 and 66 additional significant genes for IBD, CD and UC, respectively. Some of these genes were specific to IBD subtype and tissue category [see Table 5]. For example, we identified 26 genes specific to CD and the immune/blood category, which represented 11.9% of the total significant genes found in that analysis. In contrast, we found eight genes [5.2%] specific to UC and the immune/blood category [Table 5; Supplementary Figure 1]. A total of 78 category-specific genes (immune [N = 19], epithelial [N = 25], mesenchymal [N = 22] and neural [N = 12]) were not previously described by other studies [description given in Supplementary Table 6]. For example, we found that Aph-1 Homolog A, Gamma-Secretase Subunit [APH1A] underexpression in neural tissues was associated with IBD [p = 2.41E-06]. This gene participates in presenilin action in Notch and Wnt signalling, and in syndecan-3-mediated signalling events, among others.²⁸

Table 5.

Summary of TWAS joint analyses combining results of specific tissues/cell types

Phenotype	Analysis	Tissues/cell types	Significant genes [not found in joint analysis of all TWAS]	Category-specific gene [% of significant genes]	Novel category- specific genes
IBD	All	61	466	—	—
	Immune/blood	14	239 [32]	19 [8.0%]	6
	Epithelial	18	271 [33]	25 [9.2%]	9
	Mesenchymal	11	252 [33]	23 [9.1%]	13
	Neural	15	230 [15]	9 [3.9%]	3
CD	All	61	395	—	—
	Immune/blood	14	218 [38]	26 [11.9%]	11
	Epithelial	18	231 [31]	23 [10.0%]	11
	Mesenchymal	11	220 [31]	16 [6.8%]	10
	Neural	15	200 [25]	17 [8.5%]	8
UC	All	61	290	-	-
	Immune/blood	14	154 [17]	8 [5.2%]	2
	Epithelial	18	175 [25]	17 [9.7%]	8
	Mesenchymal	11	154 [19]	12 [7.8%]	5
	Neural	15	151 [21]	16 [10.6%]	2
Unique elements	—	—	—	—	78

In addition, we investigated gene pathway enrichment among category-specific IBD-associated genes. We identified 31 additional significantly enriched pathways not found in the pathway analysis of genes from the main TWAS meta-analysis described in the previous section [results in Supplementary Table 4]. Full results for all TWAS joint analyses [main and category-specific meta-analyses of TWAS results] are provided in Supplementary Data 3.

4. Discussion

In this study, we identified candidate genes that may modulate the inherited risk of IBD and could eventually be exploited as novel therapeutic targets. We integrated transcriptomic and genetic information to predict gene expression in 59 957 genotyped subjects, including 25 042 with IBD, and discovered new associations between gene expression and IBD status. Additional insight into colon subsite-specific mechanisms was provided by site-specific expression prediction models trained on the recently published BarcUVa-Seq expression quantitative trait locus dataset. To gain new insights into the mechanisms underlying IBD, we performed a large, multi-dataset TWAS,²¹ including predictive models for colon epithelium-enriched tissues and blood cell types of high relevance for IBD.

There are notable advantages of TWAS over other traditionally used approaches [such as GWAS and differential gene expression analysis] for nominating candidate genes that participate in disease pathogenesis. On the one hand, GWAS just identify risk SNPs and, except some obvious cases where an SNP lies in coding regions, this approach does not provide the candidate downstream functional effects of the SNP on the phenotype/disease. On the other hand, differential gene expression analysis using observational rather than predicted gene expression measures does not provide causal inference. In this sense, the genetic variants that regulate gene expression are not affected by the disease, and therefore the direction of the effect, from gene expression to the disease, and not the opposite, can be made for the TWAS-identified genes.

TWAS based on expression models trained on BarcUVa-Seq ascending, transverse and descending colon allowed comparison of IBD-associated genetically regulated gene expression across different colon subsites. We identified susceptibility genes specific for colon subsites and IBD subtype. The strongest association signal [p = 1.33 × 10^–104] involved Phosphodiesterase 4B [PDE4B], whose expression was associated with IBD only in ascending colon. PDE4B is a candidate therapeutic target for paediatric-onset IBD,³⁰ and expression of PDE4B was associated with UC in patient-derived colon biopsies.²⁹ The TSS of PDE4B is located over 1 Mb from any top GWAS SNP, and the gene has not previously been associated with IBD susceptibility. Another novel gene involved in genetic susceptibility to IBD is UBD, whose expression in descending colon was significantly associated with CD status. UBD is an expression marker for the microfold [M] cell, a type of colon epithelial cell associated with colon inflammation in UC-derived colon biopsies.²⁶UBD has also been reported to be upregulated in patient-derived colon biopsies²⁹ and may be a target for anti-TNF-α treatment.³¹ As these examples demonstrate, the candidate susceptibility genes identified in this study could be promising therapeutic targets for IBD treatment. Expression levels of two other genes, TRIM31 and Claudin 4 [CLDN4], were associated with IBD status for the first time. TRIM31 is an expression marker of colon goblet and enterocyte cells and was significant only in the ascending colon TWAS for UC and in the joint analysis of TWAS results of epithelial tissues, suggesting tissue type specificity. TRIM31 downregulation has been linked to bacterial invasion.³²CLDN4 is involved in the control of colon epithelial barrier function, including the maintenance of tight junction integrity. These examples highlight new directions for emerging treatment approaches.

BarcUVa-Seq TWAS revealed 39 novel IBD candidate susceptibility genes, including expression markers of 28 cell types found in the colon.²⁶ This finding allowed us to link IBD risk SNPs to colon-specific cell types that may affect genetic susceptibility. The risk SNP rs12568930 at 1p36.12 was associated with WNT2B+ Fos-lo 2 cells [a subtype of colon inflammatory fibroblasts] through the expression of Wnt Family Member 4 [WNT4]. The mechanisms of intestinal fibrosis in IBD are poorly understood, which impedes the development of anti-fibrotic therapies.³³

Next, we meta-analysed TWAS results in joint analyses that combine single-tissue results to increase the statistical power to identify associations [see Methods], given the shared patterns of genetically regulated expression across human tissues. The advantages of this integrative approach have been described elsewhere.²¹ Joint analysis of all TWAS results showed 596 genes whose genetically regulated expression might be involved in IBD genetic susceptibility, including 186 genes that were not previously reported in other large association studies [Table 3]. Our meta-analysis highlighted ENTR1 as an important susceptibility gene. This gene encodes a protein involved in presentation of TNF receptors on the cell surface, and the modulation of TNF-induced apoptosis.³⁴ We also reported other novel genes encoding proteins that play important roles in TNF signalling. For example, TNF Receptor Associated Factor 2 [TRAF2],³⁵ involved in TNF signalling, may be targeted by anti-TNF IBD therapeutics.

Finally, we performed joint analyses of single-tissue TWAS results by histological category to identify associations specific to particular tissue types, which may point to specific molecular mechanisms underlying IBD genetic risk and may give insight into potential targeted therapies. These category-specific analyses identified additional susceptibility genes, allowed us to link risk SNPs to specific tissue types, and provided insight into tissue-type specific mechanisms, as revealed by pathway enrichment analysis.

An important limitation of the TWAS approach is the possibility of spurious correlation between IBD causal SNPs and SNPs regulating gene expression of nearby genes, which could drive non-causal associations, as reported elsewhere.³⁶ This affects especially the human major histocompatibility complex [MHC] region, which features high LD between SNPs and includes several immune-related genes, such as human leukocyte antigen [HLA] genes. Indeed, the mean of significantly associated genes per locus in IBD joint analysis was three genes, whereas 6p21.33 [MHC-related] and 3p21.31 were associated with 52 and 56 genes, respectively. The high number of associations motivated fine-mapping of these loci. Our fine-mapping approach modelled correlation among significant signals and assigned a probability to explain the observed association signal for every gene in a given locus at a nominal confidence of 90%.²⁷ The number of significant signals per locus was reduced after considering only fine-mapped genes. In particular, the 6p21.33 locus retained two probable causal genes out of 52 significantly associated genes.

Among significantly associated genes with strong evidence for causality after the fine mapping of other loci, we found six genes not previously reported. These included the Programmed Cell Death 1 Ligand 2 [PDCD1LG2] gene at 9p24.1, which has been linked to immunosuppression by inhibition of T-cell proliferation³⁷ as well as the Complement C6 and C7 genes [C6, C7] at 5p13.1, which are also involved in immunoregulatory processes. In addition, we found a long non-coding RNA [Lnc-ATXN2L-1], a type of molecule that remains understudied and is considered a promising topic of research.³⁸

In comparison with other published TWAS studies for IBD,^8–10 this study provided more robust statistical associations. This is due to the use of a large number of gene expression datasets, including some of high relevance to IBD, not previously included in other TWAS [i.e. BarcUVa-Seq, CEDAR], meta-analyses including TWAS of many tissues/cell types, and fine-mapping of significant association signals. Many of the significant associations we observed have been identified by other large association studies, including GWAS^3,24 and TWAS for IBD,⁸ and other studies based on patient biopsy samples²⁶ [see Table 3], but our analysis still discovered novel associations.

Our results may guide other investigators to prioritize potential genes of interest for further functional studies. Indeed, the candidate genes we proposed would require extensive validation in an experimental setting, through, for example, the use of engineered organoid models, or CRISPR screens, which was beyond the scope of this study. We supported the robustness of our results by strong statistical significance and by showing overlap with genes described by other high-impact studies. Also, associated genes were enriched in relevant pathways for IBD, mostly immune-related, which might be potential therapeutic targets.

Funding

This work was supported by the National Institutes of Health [R01 CA204279, R01 CA143237 and R01 CA201407 to G.C.]; the Agency for Management of University and Research Grants [AGAUR] of the Catalan Government [2017SGR723 to V.M.]; the Instituto de Salud Carlos III, co-funded by Instituto de Salud Carlos III funds – a way to build Europe [PI14-00613, and PI17-00092 to V.M.]; the Spanish Association Against Cancer [AECC] Scientific Foundation [GCTRA18022MORE to V.M.]; and the Centro de investigación biomédica en red. Epidemiología y salud pública [CIBERESP] [CB07/02/2005 to V.M.]. R.C.T. received funding through the Marie Skłodowska-Curie actions - Horizon 2020 - European Union grant no. 796216; M.O.S. received a postdoctoral fellowship through the ‘Fundación Científica de la Asociación Española Contra el Cáncer [AECC]’; C.H.D. received funding through the National Institutes of Health training grant T32 5T32CA163177-07 [PI: Craig Slingluff, MD]; V.D.O. received a predoctoral fellowship through the Ministerio de Educación, Cultura y Deporte - Gobierno de España FPU16/00599.

Conflict of Interest

The authors disclose no relevant conflicts of interest.

Author Contributions

V.D.O.: Data collection and curation, statistical analysis, analysis and interpretation of data, draft writing. F.M.N.: Statistical analysis, analysis and interpretation of data, draft review and editing. G.I.S.: Analysis and interpretation of data, draft review and editing. J.G., F.R.M., M.O.S., C.H.D., M.D.: Analysis and interpretation of data, draft review and editing. A.D.V.: Statistical analysis, analysis and interpretation of data, draft review and editing. R.C.T.: Study conception and design, supervision, data collection and curation, analysis and interpretation of data, draft review and editing. G.C.: Funding acquisition, analysis and interpretation of data, draft review and editing. V.M.: Funding acquisition, study conception and design, analysis and interpretation of data, draft review and editing, supervision.