Genome‐Wide DNA Methylation Identifies Potential Disease‐Specific Biomarkers and Pathophysiologic Mechanisms in Irritable Bowel Syndrome, Inflammatory Bowel Disease, and Celiac Disease

ABSTRACT

Background and Aims

Irritable bowel syndrome (IBS), inflammatory bowel disease (IBD), and celiac disease (CeD) present with similar gastrointestinal (GI) symptoms. DNA methylation‐based biomarkers have not been investigated as diagnostic biomarkers to classify these disorders. We aimed to study DNA methylation profiles of IBS, IBD, CeD, and healthy controls (HC), develop machine learning‐based classifiers, and identify associated gene ontology (GO) terms.

Methods

Genome‐wide DNA methylation of peripheral blood mononuclear cells from 315 patients with IBS, IBD, CeD, and HC was measured using Illumina's 450K or EPIC arrays. A methylation dataset on 304 IBD and HC samples was used for external validation. Differential methylation was measured using general linear models. Classifiers were developed using penalized generalized linear models using double cross‐validation controlling for confounders. Functional enrichment was assessed using GO.

Results

Three hundred and fifteen participants (148 IBS, 47 IBD, 34 CeD, and 86 HC) had DNA methylation data. IBS‐IBD and IBD‐CeD showed the highest number of differentially methylated CpG sites followed by IBD‐HC, CeD‐HC, and IBS‐HC. IBS‐associated genes were enriched in cell adhesion and neuronal pathways, while IBD‐ and CeD‐associated markers were enriched in inflammation and MHC class II pathways, respectively (p < 0.05). Classification performances assessed using area under the receiver operating characteristic curves (AUC) for IBS‐IBD, IBS‐CeD, and IBD‐CeD were 0.80 (95% CI = 0.7–0.87, p = 6.75E‐10), 0.78 (95% CI = 0.68–0.86, p = 4.57E‐10), and 0.73 (95% CI = 0.62–0.83, p = 0.03), respectively. The performance of IBD‐HC was successfully validated using external data (AUC = 0.74 [95% CI = 68–0.80, p < 0.001]).

Conclusions

Blood‐based DNA methylation biomarkers can potentially distinguish chronic GI disorders that present with similar symptoms. GO suggested functional significance of the classifiers in disease‐specific pathology.

Comprehensive genome‐wide DNA methylation profiling of blood cells in irritable bowel syndrome, inflammatory bowel disease, celiac disease, and healthy controls identifies potential biomarkers associated with disease‐specific pathways and shows promise as a biomarker for differential diagnosis of gastrointestinal disorders.

Abbreviations

CeD

celiac disease

CGIs

CpG islands

DMPs

differentially methylated positions

DMRs

differentially methylated regions

HC

healthy controls

IBD

inflammatory bowel disease

IBS

irritable bowel syndrome

PBMCs

peripheral blood mononuclear cells

ROC

receiver operating characteristic curve

UC

ulcerative colitis

Summary.

• Potential of DNA Methylation Biomarkers: Blood‐based DNA methylation biomarkers show promise in distinguishing between gastrointestinal (GI) disorders with overlapping symptoms, such as irritable bowel syndrome (IBS), inflammatory bowel disease (IBD), and celiac disease, and in differentiating these conditions from healthy individuals, thus offering a novel diagnostic approach for these GI disorders.

• Machine Learning Enhances Diagnostic Accuracy: Machine‐learning classifiers demonstrated high diagnostic accuracy (AUC 0.73–0.80) in differentiating these conditions, and external validation confirmed the robustness of IBD‐related biomarkers.

• Disease‐Specific Pathway Insights: Gene ontology analysis identified distinct pathways associated with each disorder, such as cell adhesion and neuronal pathways for IBS, inflammation for IBD, and MHC class II pathways for celiac disease, underscoring the biological relevance of these biomarkers.

1. Introduction

Irritable bowel syndrome (IBS) is a highly prevalent stress‐sensitive, chronic gastrointestinal (GI) disorder characterized by chronic abdominal pain associated with diarrhea and/or constipation. IBS has a worldwide prevalence of 4.5%–11% [1] and is associated with a significant healthcare and economic burden [2]. Currently, no valid test can reliably diagnose IBS, so symptom‐based criteria are used. Most IBS patients have seen at least three physicians and undergone multiple expensive and invasive tests before a diagnosis of IBS, as IBS is often considered erroneously a diagnosis of exclusion [3, 4]. Other GI conditions, such as inflammatory bowel diseases (IBD), celiac disease (CeD), and even GI malignancy, can present with similar GI symptoms but are diagnosed by confirmatory tissue histopathology and treated differently.

IBD is a chronic, immune‐mediated condition manifesting as intestinal mucosal inflammation. IBD affects approximately 3 million Americans, evenly divided between Crohn's disease and ulcerative colitis (UC) [5, 6]. In general, the diagnosis of IBD is made by gastroenterologists and requires endoscopic and histological confirmation. The time between the onset of symptoms and referral to a specialist can vary substantially. One of the major sources of delay in diagnosis is treatment of patients under the assumption that they have IBS rather than IBD [7]. About 10% of IBD patients are misdiagnosed as IBS [7]. Delays in diagnosis are associated with worse outcomes, including increased incidence of complications and the need for early surgery [8, 9, 10].

CeD is a gluten‐induced, immune‐mediated enteropathy, with an estimated worldwide prevalence of approximately 1% [11, 12, 13, 14]. CeD can present with variable manifestations and may also be misdiagnosed as IBS [15]. GI society guidelines recommend serological screening for CeD in patients with IBS, chronic diarrhea, and anemia, for instance. However, up to 5% of patients are seronegative [14]. Additionally, some patients self‐initiate a gluten‐free diet (GFD), thus affecting the yield of CeD serology testing [16, 17]. Substantial delays in diagnosing CeD even up to 13 years can occur in clinical practice [13, 18, 19]. Delays in diagnosis and treatment of CeD can lead to long‐term complications such as osteoporosis, infertility, anemia, and intestinal lymphoma and increased health care utilization and pharmacotherapy [20, 21].

Thus, there is a significant unmet need for an objective biomarker‐based test that can distinguish IBS, IBD, and CeD, allow an earlier diagnosis and appropriate treatment plan, and reduce unnecessary medical costs and tests (e.g., computed tomography [CT] scans, abdominal ultrasounds, stool studies, breath tests, repeat endoscopies). The importance of a diagnostic test for patients was supported by a recent large‐scale survey of individuals with IBS who reported that a fast and accurate diagnostic test for IBS is one of the top 10 research priorities [22]. Such a discovery has the transformative potential of shifting the paradigm of diagnosing IBS, IBD, and CeD.

Epigenetic modifications including DNA methylation are mechanisms that regulate gene expression and higher‐order DNA structure. DNA methylation has emerged as a leading mechanism linking gene–environment interactions to long‐term behavioral development, particularly in complex disorders [23, 24]. In normal mammalian somatic genomes, DNA methylation mainly occurs at cytosines in a CpG dinucleotide context. CpG methylation is generally absent from short stretches of CpG‐rich sequences known as CpG islands (CGIs) which typically occur at or near the transcription start site of genes [25]. Hypermethylation of CGI promoters is tightly linked with transcriptional repression of the affected gene and therefore have been viewed as an epimutation causing the silencing of a gene. In contrast, recent studies show that gene body methylation is positively correlated with gene expression and can be potential therapeutic targets [26].

Aberrant DNA methylation has been associated with a variety of cancers and non‐cancerous disorders, including psychiatric and neurodegenerative disorders [27]. Our previous study on DNA methylation and targeted bisulfite sequencing in peripheral blood mononuclear cells (PBMCs) identified changes in DNA methylation in IBS patients compared to healthy controls (HC) [28]. This study highlighted the role of neuronal and oxidative stress pathways in the pathophysiology of IBS. DNA methylation changes in blood have been associated with inflammation in patients with IBD [29]; however, there is a lack of blood‐based DNA methylation studies in CeD. Additionally, there are no studies investigating DNA methylation as a biomarker for the differential diagnosis of IBS from other diseases that mimic IBS.

Therefore, we hypothesized that DNA methylation marks in PBMCs can serve as a biomarker for diagnosing IBS and for distinguishing IBS from IBD and CeD. Similarly, DNA methylation‐based biomarkers may be used for noninvasive diagnoses of IBD and CeD. The aims of this study were to (1) study the differences in genome‐wide methylation profiles of IBS, IBD, CeD, and HC; (2) develop DNA methylation‐based classifiers to discriminate IBS, IBD, CeD, and HC; and (3) investigate gene ontology (GO) terms and pathways associated with the classifiers.

2. Materials and Methods

2.1. Study Population

Male and female participants with IBS, IBD, CeD, and HC ages 18–55 were recruited by community advertisement or at GI clinics at the University of California, Los Angeles (UCLA). IBS and HC samples were collected between 2009 and 2020. IBD and CeD samples were collected between 2012 and 2020. In addition to the UCLA cohort, we included banked PBMC samples from patients with IBD obtained from the Crohn's and Colitis Foundation (CCF) Study of a Prospective Adult Research Cohort with IBD (SPARC IBD) with identical inclusion/exclusion criteria. The SPARC IBD cohort has been previously described [30]. The diagnosis of IBS and bowel habit subtypes were based on Rome III [31] or IV [32] criteria depending on the time of recruitment and confirmed by a clinician with expertise in IBS. HC had no personal or family history of IBS, IBD, CeD or other chronic pain conditions. Additional exclusion criteria for IBS and HC included infectious or inflammatory disorders, active psychiatric illness over the past 6 months assessed by structured clinical interview for the DSM‐IV (MINI) [30], use of corticosteroids in the past 6 months, use of narcotics in the past 2 months, and alcohol abuse. Questionnaires administered to the participants are described in detail in Appendix S1. The UCLA Bowel Symptom Questionnaire (BSQ), Hospital Anxiety and Depression Scale (HAD) [33], and Adverse Childhood Experiences (ACE) [34] questionnaires were administered to all UCLA patients in this study including IBS, IBD, and CeD. Additional questionnaires administered to IBS patients included Irritable Bowel Syndrome Severity Scoring System (IBS‐SSS) [35].

The diagnosis of IBD was confirmed by endoscopy with pathologic tissue confirmation. Since treatments such as inflammation‐reducing drugs used in IBD have been associated with changes in DNA methylation [36], we recruited patients who were treatment naïve or currently on no IBD treatment including biologic agents (including ustekinumab, risankizumab, or vedolizumab) or other agents (tofacitinib, upadicitinib, ozanimod, 6‐mercaptopurine, azathioprine, methotrexate, or 5‐aminosalicylic acid [5‐ASA]). Additional exclusion criteria included UC or Crohn's disease treated surgically without evidence of subsequent disease and history of coexistent IBS or CeD. Patients with IBD reported active GI symptoms (e.g., abdominal pain, bloating, diarrhea, and/or blood in stool). Disease activity was assessed with the following instruments at the time of sample collection: Simple Endoscopic Score for Crohn's Disease [37] (SES‐CD, UCLA and SPARC IBD cohorts), Crohn's Disease Activity Index [38] (CDAI, UCLA cohort), Short‐CDAI [39] (SCDAI, SPARC IBD cohort), Simple Clinical Colitis Activity Index [40] (SCCAI, UCLA cohort), and Ulcerative Colitis Disease Activity Index [41] (UCDAI, SPARC IBD cohort) as described in Appendix S1.

Patients with CeD reported active GI symptoms, for example, abdominal pain, bloating, diarrhea, and had their diagnosis confirmed by the presence of Marsh II–III lesions on duodenal biopsies and positive serology (anti‐tissue transglutaminase [tTG] and/or anti‐endomysial antibodies). Patients were currently adherent to a GFD > 2 weeks, history of a coexistent diagnosis of IBD, IBS, or other causes of malabsorption, or any other medical condition that could explain their GI symptoms were excluded. Questionnaires administered to patients with CeD included the validated Celiac Symptom Index (CSI) [42] and Celiac Disease Adherent Test [43].

All study participants who had a current or past history of > 1/2 pack per day of cigarettes were excluded. A small percentage of participants who were former smokers (one Crohn's disease [4%] and three CeD patients [9%]) were included in the study. We recorded the use of medications including selective serotonin reuptake inhibitors (SSRIs), serotonin‐norepinephrine reuptake inhibitors (SNRIs), tricyclic antidepressants (TCAs), benzodiazepines, statins, and beta blockers (Table S1), which may affect DNA methylation [44]. None of the HC used these medications. Participants recruited at UCLA were compensated. The study was approved by the UCLA Institutional Review Board, and all subjects signed a written informed consent prior to the study.

2.2. Statistical and Bioinformatic Analyses

2.2.1. Clinical and Demographic Data

Group differences in demographic characteristics including age, sex, body mass index (BMI), race, ethnicity, and smoking status were assessed using t‐tests, analysis of variance (ANOVA), or Fisher's tests. Summary statistics were created for disease activity scores for IBS, IBD, and CeD.

2.2.2. DNA Methylation Data Processing

The methods used to process DNA methylation data and implementation of machine learning algorithms are presented in detail in Appendix S1. In short, raw Illumina DNA methylation array data (IDAT) files generated by the Illumina iScan scanner were processed using Enmix DNA methylation analysis pipeline [45].

2.2.2.1. DNA Methylation Data Preprocessing

Quality Control (QC) metrics were generated, and samples that did not pass the QC threshold were eliminated. The sample numbers reported throughout the study are the ones that passed the QC threshold. Background correction and dye bias correction were applied to the data, and the resulting signal intensities were normalized using the “quantile normalization” method. Cell types were estimated as described previously [46]. QC information‐based filtering was implemented to filter unwanted probes out of the 855,790 CpGs on the array.

Additionally, CpGs on X and Y chromosomes, single‐nucleotide polymorphisms (SNPs) and repeats, non‐specific or cross‐reactive probes, and probes showing low variability and extreme methylation values were removed. The remaining sites, ~200,000 CpGs, were used as input for differential methylation analysis and classifier development. All the analyses were performed using R programming language (https://cran.r‐project.org/). Additional details of the analyses are presented in Appendix S1.

2.2.2.2. Differential Methylation

Group differences in methylation at CpG sites also known as differentially methylated positions (DMPs) between IBS, IBD, CeD, and HC were analyzed using general linear models using “limma” package in R [47] using age as a covariate. In a separate analysis that included additional covariates, we tested methylation differences between disease and HC groups in models controlling for age, array batch, sex, and proportions of various cell types including CD4 T‐cells, CD8 T‐cells, neutrophils, monocytes, and natural killer cells. p‐values were adjusted for multiple tests using false discovery rate (FDR). DMPs were visualized using Manhattan plots, generated using “gap” package in R [48]. Differentially methylated regions (DMRs) were identified using the “DMRCate” R package [49]. An FDR < 0.05 was considered significant for differences in DMP between diagnostic groups. However, in cases where we did not find any significant CpGs, trends for association (p < 0.05) were considered. The DMRs are defined as regions having at least three DMPs within a 500 bp window [50]. Significance was based on harmonic mean of the individual component FDRs (HMFDR) < 0.05, and trends (p < 0.05) were reported for comparisons with HMFDR > 0.05.

2.2.2.3. Correlation Between Methylation Levels at DMPs and Disease Symptom Severity Scores

We tested potential relationships between methylation levels of disease‐associated CpGs and severity scores in IBS, IBD, and CeD patients using general linear models. Within UC patients, we tested group differences in methylation levels of IBD versus HC‐associated DMPs between high and lowUCDAI scores. Within CD patients, we tested an association between the methylation levels of DMPs and SCDAI scores. Within CeD patients, we tested associations between CeD vs HC‐associated sites and CSI scores. FDR< 0.05 was considered significant.

2.2.2.4. Gene Ontology (GO) Analysis

Enrichment of GO terms and/or pathways associated with DMPs and machine learning (ML)‐based classifiers for each comparison was assessed using “missMethyl” package in R [51]. Using missMethyl package, we mapped the significant/differentially methylated CpG sites to Entrez Gene IDs and tested for GO term (including, “BP”—biological process, “CC”—cellular component, “MF”—molecular function) or KEGG pathway enrichment using a non‐central hypergeometric test [44]. This method takes into account the number of CpG sites per gene on the 450K/EPIC array and multi‐gene annotated CpGs [51]. An FDR < 0.05 was considered significant or trends (p < 0.05) were reported. For analyses of classifiers, we constructed a model with all samples and analyzed the associated GO terms.

2.2.2.5. Classifier Development and Evaluation

Using the genome‐wide DNA methylation data, we developed machine ML‐based classifiers to test their performance as diagnostic biomarkers in classifying the GI diseases and HC. Generalized linear models (GLM) were fit via penalized maximum likelihood using “glmnet” package, considering diagnosis as outcome, normalized and filtered DNA methylation probes (CpG sites) as predictors, age, cell‐type proportions, and technical batch as fixed covariates. These covariates were selected on the basis of their correlation with the principal components derived from the methylation data as is standard in analyses of methylation [44]. Of the variables, tested age, cell‐type proportions, and the methylation batch variables were found to be associated with the methylation‐based principal components (Figure S1). Race/ethnicity was not used as a covariate due to missing race/ethnicity data on some participants specifically in the IBD and CeD groups (Table 1). We used a double cross‐validation method [52] to evaluate and test machine learning models, which is a preferred method since the models are trained and tested on independent datasets. Double cross‐validation process includes two nested cross‐validation loops referred to as outer (10% held‐out set) and inner loops (90% training data). The training dataset was further subdivided into tuning/calibration (90% data) and validation sets (10%) which form the inner loop. The calibration set was used for model building (hyperparameter tuning), and the validation set was used to estimate the errors. The model with the lowest prediction error within the inner loop was selected as the best model. This model was then applied to the held‐out test dataset, and the class labels were predicted. Multiple splits (k = 10) of inner and outer datasets were run in order to avoid the bias with respect to variable selection resulting from use of a single training set (Figure S2). Lasso, ridge, or elastic net regression models (penalty term representing shrinkage α = 0–1) were fitted, and the model associated with an α resulting in the best cross‐validated performance was chosen as the final model. The prediction score that is the weighted sum of CpGs was calculated for each classifier. These scores are termed methylation risk scores (MRS) and have been used as an extension of polygenic risk scores (PRS), which capture multi‐factorial information leveraging high‐dimensional data to aid in the prediction of clinical phenotype [44]. Optimal cutoffs were chosen so as to have a minimum difference between the sensitivity and specificity and the values close to the area under the receiver operating characteristic (ROC) curve (AUC) [53]. Performance metrics including accuracy (mean and 95% confidence interval [CI]), sensitivity, specificity, F1 score, and p‐values were generated based on the 2 × 2 confusion matrix constructed using the predicted and true test sample labels.

TABLE 1.

Demographic characteristics of the study population.

	IBS (N = 148)	UC (N = 22)	Crohn's disease (N = 25)	CeD (N = 35)	HC (N = 86)	p
Age; mean (SD)	30.9 (10.9)	45.8 (14.5)	36.6 (14.7)	36.1 (13.4)	30.5 (10.2)	1.22e‐07 a
Sex (% female)	96 (65%)	12 (55%)	12 (48%)	23 (66%)	47 (55%)	0.31 b
BMI; mean (SD)	25.8 (5.6)	25.0 (5.9)	24.0 (5.2)	24.9 (4.5)	26.6 (4.2)	0.31 a
Race
Caucasian	77 (52%)	14 (64%)	20 (80%)	25 (71%)	35 (41%)	0.0005 b
Black	14 (9%)	1 (5%)	0	0	10 (12%)
Asian	27 (18%)	1 (5%)	1 (4%)	2 (6%)	22 (26%)
Multiracial	22 (15%)	1 (5%)	1 (4%)	0	10 (12%)
American Indian	3 (2%)	1 (5%)	0	1 (3%)	6 (7%)
Unknown	5 (3%)	3 (14%)	3 (12%)	5 (14%)	3 (3%)
Ethnicity
Hispanic	34 (23%)	3 (14%)	1 (4%)	5 (14%)	23 (27%)	0.001 b
Non‐Hispanic	109 (74%)	15 (68%)	21 (84%)	25 (74%)	63 (73%)
Unknown	5 (3%)	4 (18%)	3 (12%)	4 (11%)	0 (0%)

Abbreviations: BMI, body mass index; CeD, celiac disease; IBS, irritable bowel syndrome; SD, standard deviation; UC, ulcerative colitis.

^a ANOVA p‐value.

^b Fisher test p‐value.

2.2.2.6. Sensitivity Analyses

We performed a sensitivity analysis to assess the performance of the classifiers excluding patients who consumed medications that can potentially alter DNA methylation [44] including SSRIs, benzodiazepines, statins, or beta blockers (Table S1). Since IBD PBMC samples were derived from two sites (UCLA and CCF), we repeated IBS versus IBD comparison and included the “site” variable as an additional batch covariate.

2.2.3. Assessment of Classifier Performance

2.2.3.1. External Validation Set

We downloaded whole blood HM450K DNA methylation data on IBD versus HC from the Gene Expression Omnibus (GEO) database (accession #GSE87648) for validating the performance of our IBD versus HC model. The processed dataset includes 460,398 probes on 304 samples (204 IBD and 100 HC). This dataset consists of whole blood DNA methylation data on newly diagnosed Crohn's disease patients (n = 103; 50 women and 53 men, mean [SD] age = 38.7 [16.3]) and UC patients (n = 101; 45 women and 56 men, mean [SD] age = 37.1 [14.0]), and 100 HC (49 women and 51 men, mean [SD] age = 34.3 [10.3]). We filtered the external data to include probes that overlapped with the probes on the Epic array to match the internal IBD and HC datasets. We trained our algorithm on our internal IBD and HC samples using “glmnet” with 10‐fold cross‐validation as described in the previous section, and used this trained model to predict the disease status of external samples. The results were evaluated using AUC and 95% CI.

2.2.3.2. Permutation Testing

Permutation testing is a statistical method to help understand if the results of an experiment are meaningful or just happened by chance. Permutation‐based testing to assess classifier performance has been used extensively in classification problems in computational biology to test if there is a real class structure in the data [54]. To create null distributions, we generated 100 random permutations to generate AUCs from classifiers generated from shuffled labels. We compared the IBS versus IBD, IBS versus CeD, IBD versus CeD, and IBS, IBD, and CeD versus HC models generated on actual class labels, to the corresponding null distributions and calculated the p‐values for differences in the distributions.

3. Results

We analyzed DNA methylation profiles of 315 participants including IBS (N = 148; 65% women, 45 constipation‐predominant IBS [IBS‐C], 54 diarrhea‐predominant IBS [IBS‐D], 49 mixed or unsubtyped IBS [IBS‐M or IBS‐U]), CeD (N = 34; 68% women), IBD (N = 47; 22 UC and 25 Crohn's disease, 49% women), and HC (N = 86; 55% women). Table 1 shows the demographic characteristics of study participants. There were significant overall group differences in mean age, race, and ethnicity (p = 5.97e‐06, 5.0e‐4, 1.0e‐3, respectively). There were no significant group differences in the percentage of women or BMI (Fisher p = 0.31 and ANOVA p = 0.26, respectively).

About 35% IBS, 28% IBD, and 31% CeD patients used medications such as antidepressants, statins, or NSAIDs (Table [Link], [Link]). Severity of IBS symptoms was moderate with mean (standard deviation [SD]) overall severity score of 9.45 (4.18; range 0–20) and IBS‐SSS of 237 (89.70; range 0–500) which represents moderately severe symptoms (Table S2A). UC disease activity assessed by UCDAI and SCCAI (mean [SD] = 4.29 [1.25] and 6.17 [4.84], respectively) indicated mild to moderate disease. Crohn's disease activity assessed by CDAI, SCDAI, and SES‐CD indicated mild to moderate disease (mean [SD] = 146.36 [78.10], 243.82 [73.77] and 4.40 [5.30], respectively) (Table S2B and S2C). CeD activity assessed by CSI scores indicated moderate disease (mean [SD] = 38 [10.98]) and CDAT scores (mean [SD] = 14.94 [3.88]) indicated a poor adherence to a GFD by the CeD patients in this study (> 13 indicates non‐adherence, Table S2).

Based on the clinical data, a history of anxiety and/or depression was the most common co‐morbidity (28%) associated with IBD, followed by thyroid disease (8%) and gastroesophageal reflux disease (GERD, 4%) in the UCLA cohort, and IBD‐associated arthropathy among SPARC IBD patients (5%). Patients with CeD also reported the diagnosis of thyroid disease (14%), anemia and GERD (9%), and anxiety or depression (6%).

3.1. DMPs and DMRs Associated With Disease Groups

When adjusted for age, there were significant differences (FDR < 0.05) in CpG methylation between IBS versus IBD (N of DMPs = 248), IBS versus CeD (N of DMPs = 6), IBD versus CeD (N of DMPs = 655), IBD versus HC (N of DMPs = 4130), and CeD versus HC (N of DMPs = 311). Between IBS and HC, 98 CpGs showed a trend for differential methylation at p < 0.001, but none at FDR < 0.05 (Figure S3A–F, Table S3).

Analysis of locations of DMPs in the regulatory regions of the genes within each comparison suggested that a majority of DMPs in IBS versus IBD were in the promoter region. Of these, a majority of CpGs showed decreased methylation in IBS compared to IBD, suggesting potential epigenetic silencing in IBD patients. Only a minority of probes were promoter associated in IBD versus CeD, IBD versus HC, and CeD versus HC. A large number of promoter‐associated probes were hypermethylated in IBD compared to HC (Table S4).

We then tested potential relationships between methylation levels of disease‐associated CpGs and severity scores in IBS, IBD, and CeD patients. No significant association was seen between IBS‐SSS and methylation of CpG sites (p < 0.05) within IBS patients. Within UC patients, we tested 4130 CpG sites (differentially methylated between IBD vs. HC at FDR < 0.05), for the difference between high and low UCDAI disease severity scores. Of these, 44 CpG sites were associated with UCDAI scores (p < 0.05, FDR > 0.05, Table S5). The top gene ontology GO terms associated with the 44 genes included “leukocyte migration involved in inflammatory response.” Within patients with Crohn's disease, 90 of the 4130 CpGs tested were associated with Crohn's disease severity as measured by the SCDAI score (p < 0.05, FDR > 0.05, Table S5). GO terms associated with Crohn's disease severity included inflammation‐related terms such as “positive regulation of T cell activation.” A lack of association at FDR < 0.05 may be due to multiple factors, including a smaller sample size of UC and Crohn's disease compared to a combined IBD cohort (disease severity measures were different for UC and Crohn's disease and therefore were analyzed separately). However, since the CpG sites were preselected based on an FDR threshold of adjusted p < 0.05 and tested based on specific a priori hypotheses, a trend for association may be important.

Similarly, for CeD, we tested 312 CpG sites (FDR < 0.05 between CeD and HC) and identified 66 CpG sites associated with CeD severity index scores (CSI, p < 0.05). DNA methylation levels of an intergenic CpG site cg04132186 and of CpG sites in genes ADORA2A and BAHCC1, correlated with CSI scores (FDR = 0.08, Figure S4, Table S5). The GO terms associated with the 66 CpG sites included “MHC class II protein complex assembly.”

We also compared genome‐wide methylation profiles of IBS bowel habit subtypes. Between diarrhea‐predominant IBS (IBS‐D, N = 54) and constipation‐predominant IBS (IBS‐C, N = 45), there was one CpG site in FAM71E2 gene that was hypomethylated in IBS‐D (FDR adjusted p = 0.03) and one CpG site in BCAR1 gene which showed a trend for hypomethylation in IBS‐D compared to IBS‐C (FDR adjusted p = 0.08). Not much is known about the function of FAM71E2 gene. Studies have reported that BCAR1 protein contributes to the regulation to a variety of signaling pathways including cell adhesion, migration, invasion, apoptosis, hypoxia, and mechanical forces [55]. Due to BCAR1's role in regulating integrin‐dependent and cell–cell adhesion, it can influence the integrity of the epithelial barrier in the gut, potentially contributing to the pathophysiology of IBS‐D [55].

Analysis of consecutive differentially methylated CpGs sites (FDR < 0.05) within a region of the gene suggested that there were 49 and 1726 DMRs associated with IBS versus IBD and IBD versus CeD, comparisons, respectively (FDR < 0.05). IBS versus CeD resulted in one DMR with a p < 0.05, but none at FDR < 0.05. CeD versus HC and IBD versus HC were associated with 58 and 2667 DMRs, respectively (FDR < 0.05). Although not significant at HMFDR < 0.05, we found trends for association with 20 DMRs between IBS and HC comparison (Table S6).

Our additional analyses including a comprehensive list of covariates, such as age, array type, sex, race, and proportions of various cell types within the disease groups including CD4 T‐cells, CD8 T‐cells, neutrophils, monocytes, and natural killer cells suggested a smaller list of CpG sites associated with various disease pairs (Table S7). IBS versus IBD and IBS versus CeD were associated with one and three DMPs, respectively. There were no associated DMPs with IBD versus CeD. With respect to comparisons between GI diseases and HC, IBD and CeD versus HC were associated with nine and three DMPs, respectively, whereas there were no DMPs associated with IBS versus HC. This may be expected given the collinearity/confounding effect of immune cells with the presence of IBD and CeD since inflammation/immune cell types are closely related to the pathophysiology of these two diseases [56, 57].

To predict function of differentially methylated sites and regions, we performed GO analysis. Genes associated with DMPs in IBS versus IBD comparisons showed enrichment of GO terms such as “immunoglobin mediated immune response” and “negative regulation of immune response” suggesting a differential regulation of inflammation‐related genes between IBS and IBD. IBS versus CeD was associated with “MHC class II protein complex binding” and “peptide antigen assembly with MHC protein complex” suggesting a differential activation of MHC class of protein and immune response genes between IBS and CeD. IBS versus HC comparison DMPs associated IBS versus HC showed an enrichment in “ion transport activity” and “neuron fate determination” terms. DMPs associated with IBD versus CeD were associated with GO terms associated with immune system‐related pathways such as “adaptive immune response” and “cell–cell adhesion” and IBD versus HC comparisons showed enrichment of “immune response” and “adaptive immune response” pathway genes. CeD versus HC was associated with enrichment of terms such as “immune system process” and “regulation of immune system process” and “leukocyte cell–cell adhesion.” All the abovementioned GO terms were significant at FDR < 0.05 except IBS versus IBD and IBS versus HC comparisons, in which trends for GO terms (p < 0.05) have been reported. The representative GO terms enriched in IBS, IBD, and CeD‐associated DMPs are shown in Figure S5A–C. The comprehensive list of terms is listed in Table S8.

3.2. DNA Methylation‐Based Classifiers to Discriminate IBS, IBD, CeD, and HC

To identify disease‐associated classifiers, penalized regression models were trained independently on pairs of diagnoses, including IBS, IBD, CeD, and HC (Appendix S1: Table S1), and tested on a holdout test dataset resulting in disease‐specific classifiers.

3.2.1. DNA Methylation‐based Classifiers for IBS, IBD, and CeD

The selected classifiers showed high accuracy for discrimination between various patient groups. Table 2 shows the performance metrics for IBS compared to HC, IBD, and CeD. The AUC and accuracy for IBS versus IBD classifier were 0.85 and 0.80, respectively (p = 6.75E‐10) and those for IBS versus CeD were 0.82 and 0.78, respectively (p = 4.57E‐10). Figure 1 shows the ROC curves for these comparisons. For IBD versus CeD, AUC and accuracy were 0.78 and 0.73 (p = 0.002), respectively. When comparing IBS versus Crohn's disease and UC separately, the classifiers for IBS performed slightly better in discriminating against Crohn's disease compared to against UC (Crohn's: accuracy [95% CI] = 0.80 [0.66–0.9]; UC: accuracy [95% CI] = 0.75 [0.60–0.87]).

TABLE 2.

Performance metrics for IBS, IBD, and CeD‐associated classifiers.

	IBS vs. IBD b	IBS vs. CeD	IBD vs. CeD	IBS vs. HC a	IBD vs. HC b	CeD vs. HC
Number of markers	136	36	181	866	550	202
AUC	0.85 (0.77–0.93)	0.82 (0.72–0.91)	0.78 (0.67–0.88)	0.69 (0.61–0.77)	0.92 (0.88–0.97)	0.85 (0.77–0.93)
Accuracy (95% CI)	0.80 (0.7–0.87)	0.78 (0.68–0.86)	0.73 (0.62–0.83)	0.69 (0.61–0.75)	0.82 (0.73–0.89)	0.80 (0.70–0.88)
Sensitivity	0.79	0.81	0.73	0.76	0.89	0.76
Specificity	0.80	0.74	0.74	0.62	0.77	0.82
F1	0.80	0.81	0.76	0.71	0.82	0.74
Accuracy p‐value	4.57E‐10	2.21E‐07	1.84E‐073	3.06E‐05	3.41E‐16	7.11E‐09

Note: F1 denotes the weighted average of precision and recall.

Abbreviations: AUC, area under the receiver operating characteristic (ROC) curve; CeD, celiac disease; IBD, inflammatory bowel disease; IBS, irritable bowel syndrome; HC, healthy controls.

^a For IBS versus HC comparison, DNA methylation data from 450K and EPIC arrays were used, and for all other comparisons only subjects with EPIC array data were used.

^b Covariates included age, cell‐type proportions, technical batch effects for all columns, however, IBD versus HC comparison included site of IBD sample collection as an additional covariate.

FIGURE 1.

The receiver operating characteristic (ROC) curves and the area under the ROC curves (AUC) for IBS versus IBD, IBS versus celiac disease, and IBD versus celiac disease comparisons. The x‐axis represents the sensitivity, and y‐axis represents the specificity, and each point on the ROC curve represents a sensitivity/specificity pair corresponding to a particular decision threshold. The AUC values for IBS versus IBD, IBS versus celiac disease, and IBD versus celiac disease classifiers were 0.85, 0.82, and 0.78.

3.3. DNA Methylation‐based Classifiers for IBS, IBD and CeD Compared to HC

When comparing GI diseases to HC, the AUC and accuracy for IBS versus HC were 0.69 and 0.69, respectively (p = 3.06–05). The classification accuracies between the different bowel habit subtypes within IBS including IBS‐D, IBS‐C, and IBS‐M versus HC were similar to the overall IBS group (Table S9). AUC and accuracy of IBD versus HC were 0.92 and 0.82 (p = 1.33E‐08), and those of CeD versus HC were 0.85 and 0.80 (p = 7.11E‐09). The markers associated with each classifier are included in Table S10.

3.4. Assessment of Classifier Performance

3.4.1. External Validation Set

Classifiers developed from our internal IBD versus HC data were used to predict the classes of samples from an independent IBD versus HC DNA methylation dataset as described in Section 2. The classifier successfully classified a majority of participants from the independent external dataset into correct disease categories (Figure S6, AUC = 0.74 (0.68–0.80), p < 0.0001).

3.4.2. Permutation Testing

Permutation testing showed that the AUC values for the true labels for IBD versus HC, IBS versus HC, IBS versus IBD, IBS versus CeD, and IBS versus HC were significantly better than the permutated labels (p < 0.05, Figure S7). In other words, permutation testing revealed that the performance of our model was significantly better than what would be observed just by chance.

3.5. GO Terms and Pathways Associated With Classifiers

The GO terms associated with the classifiers were similar to the GO terms associated with DMPs and suggested that IBS was associated with epigenetic changes in neuronal and immune system pathways. IBD‐associated classifier was associated with epigenetic changes in inflammation and immune response pathways, and CeD classifiers were enriched in major histocompatibility complex (MHC) pathways‐related terms (p < 0.05, Table S11).

3.6. Sensitivity Analyses

We repeated IBS versus IBD and IBS versus CeD comparisons, excluding patients who were taking medications, including antidepressants, NSAIDs, and statins, which are reported to affect DNA methylation [44], and recalculated the performance metrics. The AUC for the new model was comparable to that of the original model in both comparisons (Table S12), suggesting that the models chosen were robust to the effects of these confounders.

4. Discussion

We report a comprehensive analysis of genome‐wide DNA methylation data in the most common GI disorders including IBS, IBD, and CeD, and HC. Here, we identify the differentially methylated CpG sites between IBS, IBD, and CeD which can serve as potential blood‐based diagnostic biomarkers. The genes associated with the differentially methylated CpG sites corroborated with the pathophysiologic mechanisms of the disease, and the biomarkers we report here showed a significantly predictive AUC with potential for differentiating IBS, IBD, and CeD. Following replication and validation in a larger cohort, these biomarkers can potentially allow an earlier diagnosis and appropriate treatment plan and reduce unnecessary medical costs.

Our prediction model was trained and validated on EPIC DNA methylation data from samples recruited at UCLA and the CCF, and independently evaluated on a 450K external dataset on whole blood samples. Our results and replication across cohorts and platforms suggest that our model may be robust across populations and settings, but this requires further study. Additional strengths of our study include accounting for confounders by including covariates such as batch effects, cell counts, and age in our models, applying strict exclusion criteria such as smoking and medication intake, and performing additional sensitivity analyses such as excluding patients taking antidepressants and accounting for differences by collection sites.

DNA methylation has been studied in the intestinal mucosa and whole blood samples of IBD [58, 59] and intestinal epithelium and saliva of CeD [60, 61] patients. Our study replicated some of the findings reported by these studies. For example, the differential methylation of human leukocyte antigen (HLA) region has been reported in CeD patients [61], which we also report here. In the study reporting methylation differences between IBD and HC [59], a gene RPS6KA2 associated with inflammation was reported as hypermethylated in IBD patients and was also observed as hypermethylated in IBD versus HC in our cohort (FDR = 0.002). However, our study included methylation profiling of blood samples in IBS and these GI disorders that mimic IBS, which makes it easier to compare across disease profiles. In addition to univariate tests which can have limited power to detect smaller linear and nonlinear changes, we utilized a machine learning framework which enabled us to not only develop predictive models but also select multiple relevant CpG sites without having to conduct repetitive association tests and consequential stringent multiple hypothesis adjustment. Moreover, our thorough preprocessing steps (including filtering specific sites and individuals, accounting for confounding effects) as well as the use of a double cross‐validation methodology lay the framework for reduced risk of overfitting and therefore improved reproducibility (and out‐of‐sample prediction). Finally, our stringent inclusion criteria of symptomatic untreated IBD and CeD patients diagnosed with standard diagnostic criteria but not on anti‐IBD medications or a GFD, respectively, likely helped make our model more amenable to replicability and applicability to the target population of undiagnosed patients with chronic bowel symptoms.

The goal of this study was primarily to identify diagnostic biomarkers for the selected GI disorders, and blood samples provide insights into systemic changes but may not provide much information on pathophysiological changes. However, the methylation changes we observed in the PBMC samples were also reported to be associated with changes in colonic mucosal gene expression. For example, differential expression of inflammation‐related genes in the colon of IBD patients has been previously reported by several studies [62, 63, 64, 65]. In CeD patients, studies have reported a key role for immune system genes involved in innate and adaptive immune response, in particular, the Th1 pathway [66, 67]. The role of HLA‐DQ genes in celiac disease is well known, and HLA‐DQ2 and HLA‐DQ8 are reported to be the most important genes for the predisposition to this disorder [68].

An association between IBS and the expression of barrier function and neuronal pathway genes has been reported in multiple studies from our group as well as by others. For example, in a study on colonic mucosal gene expression in IBS patients compared to HC including cohorts from UCLA, and publicly available data from the Mayo Clinic and the University of Nottingham, we reported differential expression of neuronal genes in IBS‐C patients compared to HC [69]. Additionally, another study on the expression of microRNA and mRNA in colonic mucosal biopsies and epithelial cells suggested a role for the cell adhesion and barrier function‐related genes in IBS [70]. Our group also found that gene expression in colonic mucosal biopsies of 105 patients and 60 HC (a subset of DNA methylation patients) on QuantSeq 3′ RNA sequencing platform suggested an association of IBS with GO terms such as “leukocyte cell–cell adhesion” and “immune response” (unpublished results). These studies further strengthen and support the disease relevance of biomarkers identified here. However, as the results derived from gene‐set enrichment analyses were not significant after the adjustment for multiple comparisons, these findings warrant replication in an independent cohort.

Although we found a set of biomarkers that are potentially capable of capturing disease status, the causal role and a contribution to the disease pathogenesis are unclear. The systemic changes measured using blood may be a result of several factors including the disease itself, medication, and other factors. The GO analyses which were aimed at understanding the functional relevance of biomarkers suggested an association of disease‐specific classifiers with pathways that are relevant to the pathophysiology of the corresponding diseases. For example, IBS‐associated DMPs were associated with cell adhesion, neuronal signaling, and pain pathways, which are the most widely studied pathways in IBS [71, 72, 73, 74]. Similarly, inflammatory and immune pathways and related terms were associated with the IBD classifiers further supporting the functional significance of associated classifiers [75]. Additionally, the CeD classifier was associated with MHC class II receptor activity. MHC class II encoded by HLA is a chief genetic determinant of CeD, and certain HLA‐DQ allotypes (DQ2.2, DQ2.5, and/or DQ8), which are allelic variants within the constant region of HLA genes are known to predispose to the disease by presenting posttranslationally modified gluten peptides to CD4+ T‐cells [76]. However, the systemic epigenetic changes identified in this study are predictive biomarkers and hypothesis generating, and their contribution to the disease pathogenesis is not clear.

There is currently no reliable diagnostic test for either IBS bowel habit subtypes with acceptable test characteristics that are needed for widespread clinical use. Available commercial blood tests (for IBS‐D) lack the diagnostic accuracy needed to discriminate IBS from organic GI disorders, including IBD [77]. A noninvasive blood test to diagnose IBS, IBD, and CeD would be of significant value in clinical practice. When validated and refined using additional studies, our blood‐based DNA methylation test's diagnostic potential could be improved by developing algorithms utilizing other noninvasive tests that aid in distinguishing gut inflammation and CeD (e.g., fecal calprotectin, C‐reactive protein [CRP], and celiac serologies) [16, 78].

This study has limitations. Although machine learning is a powerful tool that learns latent patterns in the data to make predictions, it is susceptible to overfitting by learning patterns that arise from known or unknown confounding variables [79, 80]. To avoid overfitting, we nested the hyperparameter optimization using double cross‐validation or nested cross‐validation (Figure S2). Since this method involves testing on a dataset that was independent of the training dataset, it provides almost unbiased estimates of the true errors [52]. Only one suitable external genome‐wide methylation dataset comparing blood samples in patients with IBD and HC was found. Although the performance accuracy of our model was slightly better on the internal dataset compared to the external data, it is known that external validations generally do not perform as well as the development model [81]. In addition, there were several methodological differences between the external IBD and our IBD datasets which may have affected the performance. For example, we used EPIC array versus the 450K used in the external data, which contains only about half the number of probes compared to the EPIC array. While the external data were on whole blood samples which include all white blood cells, and platelets, we analyzed PBMCs that include only a subset of white blood cells, that is, lymphocytes and monocytes.

We generally observed a relatively lower performance of IBS versus HC models; however, that is expected given that IBS is not considered a structural disease. However, a diagnostic test in discriminating against HC is not needed in clinical practice since HC lack GI symptoms. Nevertheless, a positive test for IBS versus HC is likely to diagnose IBS with 70% sensitivity, suggesting that it could have promise as a rule‐in test, particularly in the setting of negative noninvasive tests (e.g., negative CeD serologies, normal fecal calprotectin). Additionally, our GO analysis suggested that the biomarkers associated with IBS versus HC comparison were associated with pathways relevant to IBS pathophysiology. We could not locate similar publicly available DNA methylation blood samples from IBS and CeD patients to further validate our findings. Although we tried to account for some confounders, there might be limitations to this study due to heterogeneity across IBS patients, unavailability of race and ethnicity information on some patients, and other confounders. Therefore, although we found promising biomarkers which can be potentially used for the differential diagnosis of gastrointestinal diseases, these results warrant replication in a larger cohorts and longitudinal datasets to understand the functional role of epigenetic changes in the disease pathogenesis.

In conclusion, using comprehensive data on DNA methylation in various GI conditions and HC, our study shows that blood‐based DNA methylation changes show promise as a noninvasive biomarker to distinguish IBS, IBD, and CeD, leading to an earlier accurate diagnosis and a rule‐in test for IBS. GO analysis supports the functional significance of the classifiers in disease‐specific pathology. Future studies should test these markers further and assess their utility in predicting treatment response and identifying novel therapeutic targets.

Author Contributions

S.M.J.: study design, data analysis, writing, and editing; M.T.: data analysis and editing; E.V.: data analysis; J.D.L.: patient recruitment and editing; L.D.L.: patient recruitment and editing; M.F.: patient recruitment and editing; H.N.H.: patient recruitment and editing; T.S.: patient recruitment and editing; G.A.W.: patient recruitment and editing; B.N.L.: patient recruitment and editing; J.S.S.: patient recruitment and editing; E.A.M.: funding and editing; J.P.J.: data acquisition and editing. L.C.: study design, patient recruitment, data acquisition, writing, editing, and funding.

Conflicts of Interest

Dr. Sauk has consulted for CorEvitas, Prometheus, and Abbvie. Dr. Weiss has consulted for Guidepoint, Regeneron, and EverlyHealth. Dr. Chang has served as a member of the scientific advisory board or consultant for Alfasigma, Ardelyx, Arena, Atmo, Bausch Health, Food Marble, GlaxoSmithKline, Ironwood, and Trellus Health. She has received research support from the National Institute of Health, Arena, AnX Robotica, and Ironwood Pharmaceuticals. She has stock options with Food Marble, ModifyHealth, and Trellus Health. Drs. Chang and Mahurkar‐Joshi have two related patents.

Supporting information

Appendix S1

Appendix S2

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.