Introduction

Genome-wide association studies (GWAS) have been considerably successful in the past decade. From 2005 to 2022, approximately 400,000 single-nucleotide polymorphisms (SNPs) associated with human traits were included in the NHGRI-EBI GWAS Catalog [1, 2]. However, a limitation of GWAS is that approximately 90% of these crucial signals are located in noncoding regions [3]. Transcriptome-wide association study (TWAS) is a bioinformatic approach that integrates large-scale GWAS, uses expression quantitative trait loci (eQTL) datasets to predict gene expression levels, and attempts to identify disease-related genes and verify associations of interest. This is important for exploring specimens that are not easily collected and phenotypes rarely collected with genetic data. TWAS can assess the association between variation in gene expression levels and phenotypic variation based on different population genotype and tissue-specific gene expression data—an additional analytical approach to GWAS data—and may be used to screen candidate pathogenic genes.

Crohn’s disease (CD), a type of inflammatory bowel diseases (IBD), is a lifelong progressive disease with a tendency of symptoms to flare up or subside as the condition alternates between active and remission periods [4]. All segments of the gastrointestinal tract can be affected. The Montreal classification introduced subgroups of CD by considering the location of the disease, age of onset, and phenotype [5]. In recent decades, many researchers have believed that the pathogenesis of CD is an irrepressible immune response to luminal bacterial antigens. Immune cells participate in this process when infiltrating the gut of CD patients. Therefore, the onset of CD involves the immune system, segments of the digestive tract, intestinal tract contents, and multiple organs, which experience complications.

Over the past decades, GWAS have identified over 240 IBD susceptibility genes or loci outside the human leukocyte antigen region, and 37 of these genes are specific for CD [6, 7]. However, many of the polymorphic sites associated with CD are located in non-transcribed regions and do not cause amino acid substitutions or functional mutations, nor do they exhibit disease susceptibility. For instance, the GWAS results of Japanese patients with CD showed that among the 11 susceptibility gene loci associated with CD, only rs76418789 located in IL23R had an amino acid substitution [8]. The other 10 polymorphic sites may be located in these disease-susceptible regions, thereby affecting the expression of nearby genes and participating in the occurrence and development of CD [9]. However, the function of these SNPs in the occurrence and development of CD remains unclear. Thus, TWAS must be conducted to provide a biological context for interpreting disease risk loci by nominating candidate susceptibility genes not only in GWAS risk regions but also in other regions of potential that cannot be detected by current GWAS.

Therefore, we aim to provide an overview of previous TWAS on CD and summarize the databases and methods used in these studies. This review also discusses the overlapped susceptibility genes in different tissues and the potential pathway involved in CD by GO functional analysis, which may provide clues to explore the pathogenesis, diagnosis, and classification of CD.

Heterogeneity across TWASs of CD

There were seven TWASs for CD overviewed in this review with details in Table 1. Four of these studies conducted TWA in single tissue types of Japanese [9], Korean [10], American [11], and British [12] populations. In particular, the Japanese study also used a cross-tissue eQTL database to explore susceptibility genes based on genotypic data from the Japanese population [9]. The other three studies used cross-tissue and multi-country eQTL designs [13,14,15]. The selection of GWAS datasets, eQTLs, sample sizes, tissue types, and screening criteria varied among previous TWAS (Fig. 1).

Table 1 Summary of the basic information of TWAS studies in CD
Full size table
Fig. 1
figure 1

Workflow of TWAS in CD. The initial stage of TWAS design should consider characteristics of CD such as disease behavior, disease location, and disease status, and select GWAS datasets and eQTL populations with specific CD subtypes to ensure homogeneity within studies. TWAS could choose the public eQTL database or establish an original eQTL. The large-scale genotype dataset is used to predict expression data by eQTL to associated with CD related outcomes. CD Crohn’s disease, TWAS transcriptome-wide association studies, SNP single nucleotide polymorphism, eQTL expression quantitative trait loci

Full size image

Selection of genotype data

Large-scale genotype data provide fundamental material for TWAS. Many multi-country GWASs reported to be associated with CD provide investigators with a variety of options for data selection [16]. As the genetic structure of disease-causing mutations varies in different populations, the effect sizes and risk prediction scores derived for SNPs in one population may not be directly generalizable to other populations [17, 18]. CD has a great ancestral dependency by comparing the GWAS data from East Asian and European ancestries [19]. As yet, most of the available genetic information is based on data from populations of European ancestry [20,21,22]. Key variants that are low in frequency or absent entirely in European populations are likely to be missed when studied in other populations, especially if the variant is ethnic-specific, leaving additional blind spots for future studies [23, 24]. Therefore, considering the study population during the study design stage, caution should be exercised in genotype dataset selection. Most of the 7 TWAS selected consistent races to establish this relationship. For example, the Japanese and Korean TWAS selected genotype and expression populations from their countries. Dai et al. chose genotyped data of European ancestry as the GWAS population and 24 CD patients and 23 healthy controls from the IBD–BIOM inception cohort from UK [12, 25]. Gettler et al. used both the genotype and expression populations of the Childhood CD Study derived from the RISK cohort [11]. However, the selection of a consistent population race may limit the sample size of the study.

Selection of eQTL data

eQTL datasets are constructed using statistical models based on genotype and tissue expression data from the same population, which is an effective tool for fine-map** GWAS that identifies SNPs associated with complex phenotypic traits and can be used to improve the heritability explained by identifiable genetic factors and to better understand the biological basis of complex traits [26].

Previous TWAS exhibited acute instability when choosing various eQTL datasets based on different tissue types. Since eQTL limited the specific tissue to find differentially expressed genes, overlap** the same GWAS database with eQTLs of different tissues could result in different results of TWAS.

eQTLs are race-specific in several aspects. (1) For a specific phenotype, causal genes might be distinct across ethnicities. (2) Many polymorphisms were rare in some races but common in others, which could ignore some associations between SNP and expression in the application. (3) The degree of association between a specific SNP and the expression level in one race might also differ from that in others. Owing to these diversities, when utilizing the same GWAS dataset, substantial discrepancies exist in the conclusions drawn among different studies conducted in different countries. As can be seen in Additional file 1: data S1, for the same TWAS methods, it tends to screen out more genes with a consistent ethnicity of GWAS population and eQTL population. For example, for whole blood, the Japanese study used GTEx (eQTL from the United States) to screen out only 1 gene, and the Korean study used their own eQTL to screen out a total of 21 genes, which also suggested that eQTLs are race-specific and the consistent ethnicity of GWAS and eQTL populations may increase the accuracy of expression prediction.

Whether eQTLs are disease-specific is uncertain [27]. Because epigenetic modifications differ according to the disease state [28], the relationship between expression map** and genotype could also be affected. In our review, most TWAS combined the expression data of healthy individuals and CD patients into the same eQTL establishment. Uniquely, the Japanese study selected data from CD and UC patients to construct its own eQTL dataset of IBD [9], which may interfere with the homogeneity of eQTLs in expression prediction and further association analysis between CD patients and healthy controls though reduce the test power concomitantly. The relationship between eQTL and GWAS associations at the same locus could be unpredictable for different disease types. A previous study of non-alcoholic fatty liver disease used the disease-specific eQTL to pinpoint individuals that harbor specific genotypes more or less susceptible to the disease [29]. Thus, using disease-specific eQTL to establish the relationship in patients is worth investigating.

The resolution is low and unreliable with a sample size of 100 [27], though the sample size of the original eQTLs in most CD TWASs did not exceed 50 pairs. Various public eQTL databases are available to explore CD, including the Genotype-Tissue Expression (GTEx, https://www.genome.gov/Funded-Programs-Projects/Genotype-Tissue-Expression-Project), eQTLGen (https://www.eqtlgen.org/phase1.html), and the Blood eQTL browsers (http://genenetwork.nl/bloodeqtbrowser/). Although GTEx has been in development for 10 years, it is still worth using because its data is relatively stable and still update yet with a wide range of tissue types and sample size. As the largest eQTL database, GTEx data included genotype data from 714 donors and 11,688 RNA-seq data from 53 tissue sites and two cell lines, with sufficient assay power to establish eQTLs in 48 tissue types/sites. Although the database only included the data of a healthy population, it could be combined with the expression database of CD patients for further analysis. As GTEx included data/patients solely from the United States, its utility is limited for other countries’ populations. eQTLGen included 37 datasets from 31,684 individuals, including cis-eQTL, trans-eQTL, eQTS, and single-cell eQTLGen Consortium; however, only blood samples were used [30].The Blood eQTL Browser, which has 5311 individuals’ data, also included only blood samples [31], which limited the exploration of other tissue types more likely to be causally related to CD. Therefore, a database with more comprehensive and specific classifications across tissue types, diseases, and ethnicities is warranted to facilitate the use of multiple disease-targeting tissue types in future large-scale eQTL studies and to provide a unified platform for mining more robust associations in next-generation studies.

Selection of software or methods of TWAS

Integrating GWAS and expression data in TWAS could performed by various tools, including PrediXcan for individual-level GWAS data, Fusion and S-PrediXcan for summary-level GWAS data, closely related methods, including SMR and HEIDI, based on Mendelian randomization (MR), and GWAS–eQTL colocalization methods, including Sherlock, coloc, QTLMatch, eCaviar, enloc, and RTC, for detecting genes whose expression regulated by the same GWAS hit [32, 33].

Of the seven CD TWAS, five TWAS used commonly used methods such as Fusion, PrediXcan or MetaXcan to integrate GWAS and expression data. MetaXcan is an expanded method calculating the results PrediXcan without using individual data [33].

Besides, Uellendahl-Werth et.al conducted cross-tissue TWAS associated with gut-brain-axis performed by UTMOST (multivariate-response penalized regression models) to predict cross-tissue gene expression [13]. And they observed that UTMOST could get more moderate associations and effectively select predictive cis eQTL variants compared with S-PrediXcan (logistic regression model) and FUSION (Bayesian linear mixed model) [13]. When comparing across studies, the genes screened by ULMOST and MetaXcan rarely overlap with those screened by traditional methods (Additional file 1: Data S1), which suggests that the methods bias was not the main result responsible for the poor reproducibility. Since Uellendahl-Werth et al. did not present the associated genes screened by S-PrediXcan and FUSION in their results, whether method bias contributed to the low repetition rate is uncertain.

Gettler et al. [11] used the coloc R package, which, as Wainberg et al. previously mentioned, is vulnerable because co-regulation bias makes it difficult to distinguish causality based on GWAS and expression data [32, 34]. Hukku et al. compared PrediXcan and GWAS–eQTL colocalization methods and found that the GWAS–eQTL colocalizations may have a higher specificity and limited sensitivity, and PrediXcan could be possible to report more results with difficult in biological interpreting [35]. Therefore, caution should be used when interpreting TWAS test results derived from controversial methods, as some of them may simply be false hits. [32].

Heterogeneity of various tissue types selected in eQTL

All segments of the digestive tract, including the systemic immune inflammatory response, could be involved in CD. Previous reviews have summarized the differences in the epidemiology, genetics, histology, microbiology, and immunology of ileal and colic celiac disease, suggesting that CD at different lesion sites should be regarded as distinct subtypes [36, 37]. Each segment of the digestive tract, immune cells, and other complex organs can be used as follow-up tissues to explore the unique pathology and etiology of CD. To date, only blood, immune cells, colon tissue, and ileum tissue have been reported in previous TWAS for CD. In this section, several biospecimen-related issues of utmost concern are discussed. Susceptibility genes associated with CD reported in TWASs in different tissue type were took union set and summarized in Additional file 2 Data S2. The overlapped susceptibility genes between TWASs are visualized in Fig. 2.

Fig. 2
figure 2

Summary of susceptibility genes associated with CD in different tissue types founding in TWAS. The bold gene names were the susceptibility genes at least found in two TWASs

Full size image

In this review, GO functional analyses of all susceptibility genes associated with CD in each tissue type (Shown in Additional file 2: Data S2) were conducted by “clusterProfiler” and “pathview” packages of R software [38, 39]. Since the logFC of these associated genes were not available, the GO function enrichment analysis was roughly conducted with a random assignment of foldchange (1 or -1), which can only suggest relevant functional enrichment and cannot indicate up-regulation or down-regulation. The top 10 results for significance (P value < 0.05) of Cellular Component (CC), Molecular Function (MF), and Biological Process (BP) are shown in Fig. 3. And the total significant results were shown in Additional file 3 data S3.

Fig. 3
figure 3

The top 10 results obtained from GO functional analyses of all susceptibility genes associated with CD in each tissue type. BP biological process, CC cellular component, MF molecular function

Full size image

Digestive tissue

Clinical inflammation in CD can infect the entire gastrointestinal tract from the mouth to the anus [40]. The intestinal epithelium is a single layer of columnar epithelium that produces mucus and antimicrobial factors and can establish a buffer zone between the luminal contents and itself [4]. Destruction of the intestinal epithelium may cause bacterial invasion and lead to the occurrence of CD. Thus, in CD studies, epithelial lesion tissue is commonly seen as the top-priority casual biospecimen [41].

A comparative study observed that 19 hub genes were differentially expressed between the colon and ileum [63]. ERAP2 forms a repertoire of ligands for HLA class I, involving in the processing of MHC-I ligands antigen presentation and the antigenic response of infection [64], which associated with various inflammation diseases, such as IBD, ankylosing spondylitis, birdshot chorioretinopathy, Behcet’s disease and psoriasis [65,66,67]. In this review, ERAP2 is also observed an overlap** susceptibility gene in colon and blood, which has potential application in clinical screening and diagnosis of CD. Accumulating evidence suggests that ERAP2 is tractable targets for the regulation of immune responses [68]. In pancreatic cancer cells, gemcitabine could increase the mRNA and protein levels of ERAP2 [69]. However, no drugs targeted ERAP2 in CD yet.

GBAP1 is a pseudogene for the glucocerebrosidase (GBA) gene encodes for the enzyme glucocerebrosidase. Previous studies have demonstrated that GBAP1 can act as a competitive endogenous RNA to competitively bind with microRNAs in gastric cancer [70, 85] and Parkinson's disease [71] through functional prediction, thereby promoting the expression of GBA. However, the role of glucocerebrosidase in CD has not been reported.

GSDMB, a member of the Gasdermins family, was originally known for its role in pyroptosis [72], and most prevalently expressed in gastrointestinal-associated organs, including stomach, small intestine and colon [73]. Studies have found that the expression of GSDMB is increased in the inflammatory mucosa of ileum and colon of CD patients, and the related genes are enriched in cell proliferation, migration, and adhesion other than of pyroptosis [74]. As an inducer of GSDMB, methotrexate could induce upregulation of IEC-derived GSDMB-FL and translocation to the plasma membrane, but not lytic cell death in undifferentiated HT-29 cells. And the development of methotrexate in CD tarfeting GSDMB has entered phase III clinical trials (NCT00132899, Table 2).

Table 2 Targeted drugs had been developed of overlapped genes
Full size table

RNASET2 is the only human member of the Rh/T2/S family of acidic hydrolases [75]. An eQTL analysis observed an association between decreased RNASET2 and TNFSF15-mediated IFN-γ production, a key mediator of mucosal inflammation [76]. The circulating RNASET2 protein levels was decreased in CD patients compared with healthy control [75]. In cancer studies, RNASET2 has been found to be involved in recruitment, activation, and polarization of monocytes and macrophages [77, 78]. However, the role of RNASET2 in CD needs to be further investigated.

SLC22A5 code organic cation transporters (OCTN2), which was widely expressed and also the susceptibility genes observed in colon and blood (can be seen below). OCTN2 is mainly localized at the brush-border of apical membranes of intestinal epithelial cells and has a high transport capacity of L-carnitine in the small intestine, which is vital for β-oxidation of long-chain fatty acids in the mitochondria [79]. Several studies have observed the expression of OCTN2 downregulated in inflamed sites compared with non-inflamed sites both in patient intestinal tissue and mice model [80, 81]. And the PPARα/γ may act as transcription factors in the expression of OCTN2 and further regulate inflammatory response [80]. OCTN2 also could transports drugs, such as TEA, ipratropium, prednisolone, and beta-lactam antibiotics [82,83,84].

ZNF300P1 encode a long intergenic noncoding RNA, suggesting its primary function may be to regulate expression of other genes [85]. ZNF300P1 was found upregulated in ileum, rather than in colon or whole blood [86]. Besides, ZNF300P1 may alter tissue-specific expression of TNF and a range of additional genes previously implicated in colitis and/or autophagy. Besides, ZNF300P1 is reported to regulate polarity, proliferation, migration, and adhesion in ovarian epithelial cells [87], suggesting that it may similarly participant in intestinal epithelial functions.

The pathways of susceptibility genes associated with CD in ileum were enriched in lipid-related metabolism (Fig. 3A). Previous observational studies have reported a distinct lipid profile in CD patients compared with healthy population [88, 89]. And growing evidence showed emulsifying omiga-3 fatty acids maybe a potential supplementary in maintaining remission of CD patients [90, 91]. An epidemiological study observed that lower total cholesterol levels, LDL-C, and HDL-C were associated with higher incidence of CD, but not UC [92]. Coincidentally, another study also observed that more lipid components significantly changed in CD patients than in UC patients compared with healthy population [93]. Considering that ileal lesions are present only in CD, the different association of lipid metabolism with these two types of IBD may be due to the location of the lesion in the ileum. However, A shotgun lipidomics study of noninflammatory ileal biopsy tissue identified only phosphatidylinositol 16:0/18:1 was different between healthy controls and CD patients, although the sample size was small [89]. Additional future exploration will be necessary to confirm this observation.

Colon

Unlike the ileum, the colon has a substantial bacterial load, which plays a crucial role in regulating gut health. Changes in the abundance of specific bacteria have been used as biomarkers for screening gastrointestinal disorders, including IBD, irritable bowel syndrome, adenomatous colonic polyps, and colorectal cancer. Changes in the abundance of some bacteria have been used as biomarkers to screen for IBD and other gastrointestinal diseases [94,95,96]. A TWAS conducted in the gut microbiota has detected multiple tissue-specific candidate genes in the sigmoid and transverse colon, respectively, such as TOB2P1 for Enterococcaceae in sigmoid colon, WDR6 for Coprococcus in sigmoid colon, and KCNIP3 for Veillonellaceae in transverse colon [97]. An association study using bioinformatic analysis in colorectal cancer also observed two overlap** pathways, the bile secretion and steroid hormone biosynthesis pathways, enriched by operational taxonomic units (OTUs) and gene expression patterns in colon tissue, respectively [98]. These results indicate a close cross-talk between the intestinal microbiota and the colon transcriptome. Thus, the results of colon TWAS could identify potential genes stimulated by the microbiota and provide hints to explore the colon-specific pathology of CD.

Four TWASs included colon as casual tissue, and each study included at least 2 segments of colon. In several studies, susceptibility genes observed in the sigmoid and transverse colon almost overlap. In Japanese population, among five genes observed in colon, four genes, including ERV3-1, NPIPB9, ZNF713, and WDR31 overlap between sigmoid and transverse colon [9]. However, none of these genes overlap with the results of other reports. For the cross-tissue TWAS, Uellendahl et al. found that 18 genes were differentially expressed in the sigmoid colon, and 31 genes were differentially expressed in the transverse colon [13]. Seven genes were overlapped between two segments of colon. In European ancestries, Cheng et al. found that ZNF300P1 and MICB were significantly differentially expressed in both the sigmoid and transverse colon among the 3 susceptibility genes in colon [14]. In a meta-analysis, Virginia et al. conducted TWAS in three colon segments, including the ascending, transverse, and descending colon. And three colon segments had 11 overlap** genes, including SLC22A5, GSDMB, ENTR1, ERAP2, C4A, FUT2, UBA7, GSDMA, FLRT3, RBM6, and HLA-C [15]. In the comparison between Uellendahl’s and Virginia’s studies, ERAP2 and IL23R in the transverse colon were observed in both. None of these susceptibility genes was replicated across studies in the sigmoid colon, because two of these studies found a relatively small number of genes (Additional file 2: data S2) [9, 14].

Among all susceptibility genes reported in colon regardless of which segments, ERAP2 [13, 15], Interleukin-23 receptor (IL23R) [13, 15], Major histocompatibility complex class I chain- related gene B (MICB) [14, 15], Post-GPI attachment to the proteins 3 (PGAP3) [13, 15], and SLC22A5 [14, 15] were overlapped between TWASs. ERAP2 and SLC22A5 were also overlapped with ileum discussed above.

IL23R is one of popular genes affects disease susceptibility and highly expressed on cell membrane of memory T cells and other immune cells, such as natural killer cells, monocytes, and dendritic cells [99]. IL23R interacts with IL-23, regulating the of immune activity and against infection by bacteria and viruses [99]. And the functional IL23R pathway polymorphisms play a role in modulating neonatal development of intestinal tolerance and bacterial colonization [100]. There were several of humanized monoclonal IgG, including Brazikumab, Risankizumab and Mirikizumab, could binds p19 of IL23 has entered clinical trial, and most of them has enter phase 3 clinical trials (Table 2). Due to the presence of protective or disease-associated variants in IL23R and related genes, only one locally acting oral peptide (JNJ-67864238) directly antagonizing IL-23R was found but was recently terminated after meeting criteria for futility [NCT04102111] [101].

MICB almost exclusively expressed in the intestinal epithelium [102]. MICB was reported in many human cancers via immune evasion [103,104,105]. And the immune cells, including natural killer (NK) cells and T cells, involved in MICB were also connected with CD. However, the functions of MICB in CD were still lack of evidence. In addition, MICB has only been reported as a CD susceptibility gene in whole blood, but not in blood immune cells, which needs further study.

PGAP3 is ubiquitously expressed and code a Glycosylphosphatidylinositol (GPI)-specific phospholipase involving in lipid remodeling of GPI-anchored proteins[106]. The function of PGAP3 was most reported in brain morphogenesis and mental development [107, 108]. However, the mechanism of PGAP3 in CD was still under studied.

To be noted, the susceptibility genes associated with CD in colon enriched in the cell component of vesicle membrane (Fig. 3B), including exosomes, microvesicles and apoptotic bodies from endosomes, plasma membrane, plasma membrane/endoplasmic reticulum, respectively [109]. And the vesicle may be related with bacteria–host communication, which may involve in internalization of bacterial extracellular vesicles of epithelial cells [110]. Endocytic routes of intestinal epithelial cells, including macropinocytosis, clathrin-mediated endocytosis and lipid raft-mediated processes, may involve in CD pathogenesis [109]. Furthermore, extracellular vesicle (EV), mainly secreted by immune cells and intestinal epithelial cells, could package double-strand DNA (dsDNA), activating the STING pathway to provoke inflammatory responses [123]. Whole blood seems to be a tissue type with less disease-specificity. However, it is also the most accessible biospecimen in clinical practice and could thus obtain sufficient test power with a large sample size. The differentially expressed gene is most likely to have the potential to be a biomarker and could be extensively used in clinical practice to help earlier diagnosis and disease classification.

Among seven TWASs associated with CD, five TWASs selected whole blood as targeted tissue. And among total 144 susceptibility genes in whole blood related with CD, ATG16L1 [12, 13], Caspase recruitment domain 9 (CARD9) [12, 14], ERAP2 [12, 13], MICB [12, 14], NOD2 [10, 12, 13], SLC22A5 [12, 14], and Tumor necrosis factor superfamily 15 (TNFSF15) [9, 10, 12] were overlapped between TWASs. ATG16L1, ERAP2, MICB and SLC22A5 were overlapped with intestinal and discussed in the above.

Among the over 40 risk loci associated with CD identified to date, polymorphisms in NOD2 account for the largest proportion of the genetic risk for this disease [124]. Experiments have demonstrated that NOD2 recognizes bacterial muramyl dipeptides and recruits ATG16L1 to bacterial entry sites on the plasma membrane, further regulating the intestinal barrier function and limiting transcellular permeability and bacterial translocation [125, 126]. However, the differential expression of NOD2 in circulatory—but not intestinal—tissues is puzzling. NOD2 is widely expressed in macrophages and dendritic cells but to a lesser extent in intestinal epithelial cells [127] and T cells [128], which may explain this phenomenon. In addition, NOD2 can act as a viral sensor protein and is activated by the orally bioavailable dinucleotide SB 9200 [129], but whether it is effective against CD is unknown.

CARD9 was a member of CARD family and an adaptor molecule predominantly expressed in lymphoid tissues and immune cells. The expression of CARD9 was observed significant reduced in CD patients compared with healthy controls [130]. CARD9 is a central signaling molecule in the innate immune via mediating NF-κB signaling and against fungi, bacteria, virus and mycobacteria [131,132,133,134], which is closely related to the pathophysiological of CD development.

TNFSF15 is a Th-1-polarized cytokine that participates in systemic inflammatory responses and functions in regulating immune cells, inducing apoptosis, inducing inflammation, and inhibiting tumorigenesis, which suggests that TNFSF15 is possibly a susceptibility gene in the blood. However, two studies of IBD observed that TNFSF15 was overexpressed in colonic tissues [135, 136]. The protective effect of TNFSF15 polymorphisms on CD has been reviewed elsewhere [99, 137]. Notably, TNFSF15 was only screened in two Asian populations, Japanese and Korean, and was not found in other populations, and it can be tentatively speculated that the association of TNFSF15 with CD is stronger in Asian than in Western populations. Many previous studies have supported this hypothesis. A Japanese study reported a trend for a positive association between TNFSF15 SNPs and the risk of anal lesions in CD [138]. Similar results were obtained in Chinese [139] and Korean population s[140]. In European population, a protective effect of TNFSF15 was observed in CD but fail to define a clinical subgroup of CD patients specifically associated with TNFSF15 [19]. A most recent study compared the susceptibility genes associated with IBD between two population from East Asian and European ancestries, respectively. In this study, researchers found that the genetic basis of CD appears to be more ancestral than that of UC due to the allele frequency of NOD2 and the influence of TNFSF15 [19]. And a meta-analysis also observed that East Asians gene have unique SNPs of TNFSF15 associated with IBD [142]. This phenomenon suggests that CD is more likely to result from the interaction of dietary behaviors and environmental factors with host immune mechanisms. The onset of CD can even be traced back to infancy; breast milk, containing oligosaccharides, contributes to the establishment of intestinal flora in infants and has a longer duration of benefit in inhibiting the adhesion of enteropathogenic bacteria and protecting against the development of the disease [142]. A recent study reported a transcriptome-wide association with gut microbiota [97]. Since numerous studies have demonstrated that the gut microbiota has a close relationship with CD pathology, the TWAS in the gut microbiota may provide new insights into investigating novel pathological mechanisms. This new gut microbiota-based tool may be influenced by diet and medication use, and its applicability to CD research remains to be demonstrated. Thus, broader exploration is needed, and the results should be interpreted with caution. Although over 200 genetic loci associated with IBD have been identified by GWAS, these variants can only explain a small proportion of the heritability of IBD (approximately 26% for CD and 19% for UC) [143]. No genetic markers have been reported to be predictive of complications [37, 144]. Considering that total expression is affected by genetic and environmental factors and that predicted expression in TWAS is only a part of the total expression, gene expression data assessed by genotype data and eQTLs have strong limitations and biases. The predicted expression in TWAS was generally slightly higher than the total expression correlations. The analysis of the correlation between predicted expression and CD may result in the significance of non-causal genes over causal genes due to linkage disequilibrium [32]. When a large proportion of genes are in linkage disequilibrium, the linkage disequilibrium region may also contain causal associations not related to the gene set. Even if there is no causal relationship between the gene set and the phenotype, it can still exhibit a significantly high rate [145]. Therefore, more work is needed to understand the genetic structure of CD.

Before the emergence of GWAS, most genotypic-phenotypic associations failed to replicate owing to small sample sizes, improper reliance on standard significance thresholds, failure to account for associations with low prior probabilities, and failure to assess the same SNPs across studies [1, 146]. We must acknowledge that the replication of TWAS results is not easy. The following reasons make the replication for CD TWAS even harder. (1) As a complex disease, the subtypes of CD patients, such as age of onset, lesion site, disease behaviors, disease process (active/remission), and medication used (hormone/biologicals), have different expression profile [151, 152]. One study suggested a more significant overlap between eQTL and methylation QTL, and both studies suggested that related effects may be specific to cell type or disease status [151, 152]. During the past 100 years, the incidence of inflammatory bowel disease has sharply risen, then plateaued in the western world, whereas countries outside the western world seem to be in the first stage of this sequence [142]. The fact that CD patients are often treated for life and tend to concentrated in a few well-known treatment centers, it is feasible to obtain sufficient numbers of patients and biological samples through recruitment in well-known treatment centers. And a certain tissue type within the same study design is recommended to be consistent, allowing for increased sample size on a limited budget. In the future, larger-scale, cell/tissue-specific, and status-specific studies will be vital to resolve this problem. With the advancement of technology, single-cell RNA sequencing and single-cell TWAS have already emerged, substantially improving the homogeneity of samples and further facilitating targeted interpretations of TWAS outcomes and disease mechanisms for individual cell types or specific disease states [153].

Among the most frequently mentioned TWAS genes, such as ATGL16L1, NOD2 and IL23R, were most reported by coding risk variants in GWAS studies instead of replicating the results of RNA-seq. Among the seven studies, only one TWAS provided the gene list associated with CD obtained by RNA-seq data [12]. Of the 95 associated genes screened by TWAS and the 35 associated genes screened by RNA-seq, only two genes (RPL9 and STMN3) overlapped. Since the most of RNA-seq data involved in the other six TWAS were either not associated with CD or did not include appropriate cases and controls (HC only, CD only, CD and UC), we found another two well-designed studies for comparison. Two Asina studies identified differentially expressed genes by RNA-seq in CD patients [154, 155]. Unfortunately, there was no susceptibility gene overlap between RNA-seq results and TWAS results neither in the ileum or colon. And the susceptibility genes screened by TWAS was less overlapped with the results of RNA-seq. There are several reasons for this phenomenon: (1) According to the distance of gene effect, eQTL includes cis-eQTLs (local) and trans-eQTLs (distal) [156]. A previous study observed that pervasive cis-eQTLs affect the majority of human genes (~ 75%) [157, 158], but a large twin study claimed that only 10% of the variation in gene expression was explained by cis-eQTL [123]. However, cis-eQTLs remain the only reliable tool in the TWAS method [32], which was limited in assessing the allele-specific expression [159]. (2) Stretch enhancers are large chromatin-defined regulatory elements that regulate the expression of cell type-specific genes and are enriched in disease-associated genetic variants in disease-associated cell types. However, eQTL effect sizes for stretch enhancers may be smaller than for ubiquitous promoter regions, which may lead to prediction bias [34]. (3) The pleiotropy, including horizontal pleiotropy and vertical pleiotropy, is widely existed in genome but the exact extent is still unknown [160]. Most of the genes may be indirect causative genes for complex traits, and some of the GWAS gene expression predicted by eqtl may be amplified due to horizontal pleiotropy [35]. (4) Gene expression may be affected by heritable epigenetic variation, small signaling molecules or other environment factors [161]. For example, NOD2 expression could be induced by bacterial lipopolysaccharide, short-chain fatty acids, hormonal vitamin D, and TNF-α [162], which make it harder to predict the real expression levels. Although rarely reported in RNA-seq studies, these gene expressions are involved in mucosal immunity as previous reported [99, 162, 163]. This suggests that RNA-seq and TWAS may have complementary roles in explaining genetic associations of complex traits.

Counterintuitively, a susceptibility gene, such as NOD2, identified in a tissue type is not always consistent with its function. This observation raises the question of whether differential expression results obtained by eQTLs can explain causal associations, and a growing body of data has raised this question. As Wainberg et al. pointed out, the TWAS method is merely a statistical test to predict expression and disease risk from genetic evidence, which can be used to screen candidate disease-causing genes but does not guarantee causality [32].

In conclusion, the following three considerations might benefit future TWAS for CD, facilitating a more rational study design. (1) Despite its generic nature, we require GWAS data from different countries and disease states with large sample sizes. (2) The demand for a comprehensive classification, including race, tissue, lesion site, status, and progressive time points, is increasing with the accumulation of eQTL data. (3) Transcriptome-wide data combined with new technologies, such as single-cell approaches, will provide novel insights into the pathological mechanisms of CD and progress in TWAS. (4) In future TWAS, the results and data of intermediate processes should also be provided to facilitate the integration of data from multiple studies, dig deeper into genetic information, and provide more predictions for drug discovery.