Environmental variables and genome-environment interactions predicting IBD diagnosis in large UK cohort

We used data from the UK Biobank, a “prospective cohort study with deep genetic and phenotypic data collected on approximately 500,000 individuals from across the United Kingdom, aged between 40 and 69 at recruitment.”⁹ UK Biobank carried out all data collection and methods in accordance with health research regulations in the United Kingdom. The UK Biobank study was approved by the North West—Haydock Research Ethics Committee in the United Kingdom under NHS Research Ethics Committee number 16/NW/0274. Informed consent was obtained during UK Biobank data collection. Enrollment of participants for long-term follow-up took place between 2006 and 2010; the data used in this study was retrieved in 2019. The phenotypic data was collected from a variety of sources, including in-person surveys and interviews conducted at enrollment on a range of topics including past diagnoses, periodically emailed dietary recall questionnaires, and hospital episode statistics (HES) records detailing past diagnoses and surgical operations.

We performed quality control on the dataset to ensure that our cohort was genetically homogenous and non-related. We excluded individuals who did not have British white ancestry (as in previous studies¹⁰, British white ancestry encompassed those who both identified as “white-unspecified,” “white-British,” or “Irish” and were within 7 standard deviations of the mean for the first 6 genetic principal component measures) because we did not expect homogeneity of effects across ancestral groups and because the number of non-white patients was too small in the UK Biobank to have sufficient power to study. We also removed individuals who had close kin in the cohort based on estimated kinship coefficients. Specifically, participants who were related to multiple other participants to the first, second, or third degree were excluded. From each remaining related pair, a random participant was excluded and a random one was kept. Cryptic relatedness was determined using the package KING, with settings designed to exclude third-degree relatives and closer, as described previously⁹. In addition, we excluded those with sex chromosome aneuploidies from analysis. 364,908 individuals from the original 488,377 passed quality control, including 168,992 males and 195,916 females with a median age of 68.

We identified 5306 individuals who either self-reported a previous diagnosis of IBD or had IBD coded (ICD-10) in their hospital episode statistics (HES). 1946 had CD and 3715 had UC (Fig. 1). There was considerable overlap between those who self-reported IBD and those with IBD in the HES records (46% of IBD cases were identified through both routes), but there were also discrepancies. These were likely due to misremembered diagnoses, patients seeking medical care outside the NHS, or the lack of complete HES coverage prior to 1997. In fact, 18.5% (67,423 individuals) in the cohort did not have any HES data at all.

Number of IBD cases in the UK Biobank based on method of identification. ICD10-coded diagnoses were recorded in hospital episode statistics (HES) while self-reported conditions were gathered by survey at recruitment.

To properly characterize the overlap in self-reported versus hospital record diagnoses, we looked at those individuals with an IBD diagnosis after 1997, which is when HES data became available across the United Kingdom. Of the 1208 individuals who self-reported an IBD diagnosis after 1997, 912 (75.5%) also had an HES-coded IBD diagnosis. This high but incomplete overlap between self-reported and HES-coded diagnoses in the UK Biobank has been noted for other phenotypes, and genetic evidence suggests that these phenotyping methods identify comparable sets of cases¹¹. We also compared timings of diagnoses for those with self-reported IBD after 1997 and found that earliest HES-coded diagnosis was usually either in the same year (27.0%) or later (60.1%) than the recalled date of diagnosis, but that there was a highly variable lag time between recalled date of diagnosis and date of first HES record (Fig. 2). In our analyses below, we therefore rely on HES records and self-reported diagnoses separately.

Lag time between self-reported IBD and HES-coded IBD for patients self-reporting IBD after 1997, measured by number of years from earliest recollection of first diagnosis to first instance of relevant ICD10-code in HES records.

Of note, we do not include indeterminate colitis (ICD10 code: K52.3) in our definition of IBD. We also interpret every HES record as a definitive diagnosis, although we note that coding errors, diagnostic errors, and diagnostic changes can occur in IBD, which may affect the results of this study. We therefore carried out a sensitivity analysis in which we removed the 592 patients (Supplemental Table 2) with either conflicting CD and UC diagnoses or a HES record of indeterminate colitis.

Among IBD cases there were 2602 males and 2704 females with median age 69, which is nearly identical to the median age of the entire cohort, 68 (Supplemental Fig. 1). There was a sex difference in IBD prevalence (1.54% in males vs. 1.38% in females, p = 1.63e−04). Previous studies have indicated that this sex difference can be attributed to the increased prevalence of UC in older males, and this is indeed the case in the UK Biobank (1.12% in males vs. 0.927% in females, p = 2.85e−08)¹².

Genetic risk

Using 232 previously identified biallelic SNPs³ associated with IBD, CD, and UC, we calculated polygenic risk scores (PRSs) for each of the three disease phenotypes according to the following standard formula

where \(\beta\) is effect size and g is genotype represented by 0, 1, or 2. We then tested whether our PRSs associated with their respective disease phenotypes in a logistic regression, with age, sex, genetic ancestry, and location of the UK Biobank assessment center as covariates (Eq. 2).

Environmental associations and GxE interactions

We used Cox proportional hazards regressions to model the risk imparted by various environmental variables in right-censored survival analyses where the event of interest was IBD diagnosis as noted in self-report surveys (using the participant’s recalled year of diagnosis) or HES records (taking the earliest hospital episode as the date of diagnosis). To minimize confounding and to control for demographic factors and known genetic risk, we typically included the following covariates in our models: PRS, age, sex, 10 genetic principal components (ancestry), and UK Biobank assessment center location (Eq. 3). There were exceptions, however (also refer to Table 1): When testing the effects of 24-h dietary recall variables, we added daily caloric intake as an extra covariate to control for total consumption. When testing for geographic variables—namely, socioeconomic status, latitude at birth, latitude at recruitment, and sun exposure during the summer and winter—we removed the location of the assessment center from the usual list of covariates to avoid collinearity. This information is summarized in Table 1.

We corrected for multiple hypothesis testing using Benjamini–Hochberg adjustments to control false discovery rate at FDR < 0.05 within each disease subtype. All significant results were checked graphically; poorly fit models and data which did not visually conform to the proportional hazards assumption were discarded (only models for hormone replacement therapy were discarded).

Based on previous findings and biological plausibility^5,6,13,14, we chose 38 environmental variables to test for association with IBD (Table 1). For each environmental variable, any individual with incomplete or illogical data (e.g., started smoking before quitting smoking) was excluded from analysis. All but three variables were assessed through a single recall event which was collected via touchscreen survey and verbal interview at recruitment; for the vast majority of participants (~ 95%), there was no further follow-up for these variables. For the small minority of participants with follow-up responses to survey questions, the averaged value of their responses and the date of the initial survey were used in the analysis. The three exceptions to this method of data collection were appendectomy, socioeconomic status, and 24-h diet recall. Appendectomy was both assessed by recall at enrollment and gathered from HES records, whichever occurred earlier. Socioeconomic deprivation was determined by matching participant zip codes at recruitment against the UK government’s Index of Multiple Deprivation (IMD) from 2010. And in contrast to all other variables, 24-h diet recall was assessed through multiple recall events. Participants were sent questionnaires five times between 2009 and 2011 at approximately 4–6 month intervals asking them to recall what they consumed in the past 24 h. For accuracy, we excluded those who did not respond to at least three of the five questionnaires.

We also asked whether any of the environmental variables interacted with the PRS by adding an interaction term to the Cox model (Eq. 4).

We chose to separate our analysis of self-reported IBD and HES-coded IBD because we found a variable lag time between self-reported IBD after 1997 (when HES data became available) and HES-recorded diagnoses in the UK Biobank (Fig. 2). That is, in each survival analysis and regression model we used either self-reported IBD or HES-recorded IBD depending on whether the analysis was prospective (in which case we relied on HES data) or retrospective (in which case we used self-reported data). Our approach is displayed in Fig. 3.

Schematic of the primary analyses we carried out for each variable. All variables were measured at a single time point except 24-h diet recall, which represented the average value from 3 to 5 surveys taken at multiple time points. IBD = inflammatory bowel disease, IMD = Index of Multiple Deprivation (a measure of socioeconomic status), OCT = oral contraceptive therapy.

Prospective analyses were conducted for variables that dealt with environmental exposures around the time of enrollment. Data on diet patterns, socioeconomic status, summer and winter sun exposure, latitude at recruitment, and regular non-steroidal anti-inflammatory drug (NSAID) use were all collected via a touchscreen survey at the time of enrollment. Their association with IBD was tested in prospective analyses that relied on HES data for diagnoses of IBD after enrollment. These analyses began at the time of enrollment and proceeded until either a diagnosis of IBD was made or the date of the patient’s most recent HES record.

Prospective analyses were also carried out for the 24-h dietary recall variables, which are distinct from the dietary patterns data collected via the touchscreen survey. These 24-h dietary recall variables were collected through a series of 5 questionnaires sent to participants over the period 2009 to 2011. These prospective analyses began at the time of the first questionnaire completed and proceeded until either a diagnosis of IBD was made through HES or the date of the patient’s most recent HES record.

Retrospective analyses were carried out for perinatal and childhood variables–namely, birth by cesarean section, being breastfed, maternal smoking around birth, and prolonged childhood exposure to antibiotics. Analyses for perinatal variables began at birth and proceeded until a self-reported diagnosis of IBD (since HES records were not available during these years) or the date of enrollment into the UK Biobank (since this is when self-reported diagnoses were collected via the touchscreen survey). The analysis of childhood exposure to antibiotics begans instead at age 19–the definition of “childhood” for this variable in the UK Biobank–but otherwise was the same.

Finally, time-varying retrospective analyses were carried out for lifespan variables whose statuses could change across a participant’s life. These included appendectomy, smoking status, hormone replacement therapy use, and oral contraceptive use. These were the only variables which underwent a time-varying analysis. The analyses began at birth and proceeded with time-varying changes to a participant’s exposure status until either self-reported IBD or the date of enrollment.

Because of the uncertainty around the dating of IBD diagnosis in the UK Biobank, and because recorded IBD diagnosis typically lags behind real diagnosis of disease, we performed robustness analyses around the point of truncation whereby individuals suspected of having IBD before the truncation point were removed even if their recorded date of diagnosis fell after that time. We did this through one of three ways: (1) additionally removing all IBD cases diagnosed within 2 years after the survey, (2) additionally removing all those who had surgeries (excisions into small intestine, colon, and rectum) commonly performed in IBD patients before the survey, and (3) additionally removing all those who had either IBD-related surgeries or endoscopies before the survey (relevant OPSC-4 codes in Supp. Table 1). Participants ruled out from analysis based on their histories of surgeries and endoscopies were enriched for future IBD (for surgery only: 4 out of 11,918 vs. 68 out of 329,783, OR = 1.63, p = 0.32 by Fisher’s exact test; for both surgery and scope: 4 out of 6393 vs. 68 out of 355,312, OR = 3.27, p = 0.039), indicating that our methods did indeed target potential IBD cases who simply had not been identified as such before truncation. These removals may introduce new biases into the analysis, so we only used these analyses to test the robustness of the statistically significant results in the original data and not to draw new conclusions.

For analyses without left truncation, we performed a different robustness analysis whereby we removed individuals whose date of IBD diagnosis fell within 2 years after a change in environmental status—for instance, getting an appendectomy or quitting smoking.

Lastly, to assess the possible role of recall errors, we conducted sensitivity analyses for the lifespan variables—i.e., appendectomy, smoking, hormone replacement therapy, and oral contraceptive therapy (OCT)—using an alternative prospective analysis. Follow-up began at enrollment and proceeded until either a diagnosis of IBD on HES records or the date of the last HES record. The statuses of these variables were fixed at baseline for the vast majority of individuals since post-enrollment data for these variables were gathered by the UK Biobank only in a minority (~ 5%) of participants. The sensitivity results for current OCT use were discarded because only 2 active OCT users were diagnosed with IBD after enrollment.