Background

Here, we demonstrate an integrative PheWAS approach to establish high-confidence, suggestive metabolite-disease associations for three fatty acid metabolites, omega-3 fatty acids, omega-6 fatty acids, and docosahexaenoic acid for 1,326 disease endpoints. Suggestive metabolite-disease associations were established if there was concordant direction of effect and significance for metabolite level and genetic risk score for the metabolite.

Methods

[1] Cohort design

Participants were ascertained from the UK Biobank (UKB), a cohort study of approximately 500,000 individuals from across the United Kingdom and including multiple ancestries, although predominantly European ancestry. For this analysis, previously imputed genotype and phenotype data were utilized. 409,148 individuals after sample QC were considered. There are up to 121,643 individuals in the UKB with reported metabolite levels. Of this we extracted only individuals passing the sample QC, which resulted in a total sample size of 101,793. Metabolite levels were extracted for ω3FA (field ID: 23444), ω6FA (field ID: 23445) and DHA (field ID: 23450). For individuals with more than one reporting of the metabolite level, the median of the levels was taken. ICD-10 summary diagnosis codes were extracted from UKBB to create case and control cohorts using the case inclusion and control exclusion criteria for each disease-specific phenotype code (phecode) using the phecode mappings (https://phewascatalog.org). There was a total of 1,755 disease end points to assess; however, only diseases with at least 50 cases were analyzed, which resulted in evaluation of 1,326 diseases.

[2] Disease-metabolite associations

A metabolite Z-score for the metabolite level was evaluated using the following formula: z=(x-μ)/σ, where x is the median metabolite level, μ is the mean of the median metabolite levels and σ is the standard deviation of the median metabolite level. Logistic regression models were implemented to assess the association between each Z-scoreMetabolite and disease phecode (case vs. control). Age, sex, age2 and the first ten global principal components were included in the model as covariates. Model structure: Disease status ~ Zscore-Metabolite + age + sex + age^2 + PC1 + ⋯ + PC10

[3] Disease-metabolite PRS associations

The Bayesian approach to PRS calculation, PRScs, was used, estimating posterior SNP effect sizes under continuous shrinkage (CS) priors for each GWAS summary statistic using the UK BioBank European LD reference panel (https://www.dropbox.com/s/t9opx2ty6ucrpib/ldblk_ukbb_eur.tar.gz). The inferred posterior effect sizes were then used to generate the PRS across chromosomes using PLINK’s score function and then summed for each individual. Logistic regression models were performed to assess the association between each scaled PGSMetabolite and disease phecode (case vs. control). Age, sex, age2 and the first ten global principal components were included in the model as covariates. Model structure: Disease status ~ PGS-Metabolite + age + sex + age^2 + PC1 + ⋯ + PC10

[4] Metabolite ~ disease Mendelian randomization associations

There were 178 unique significant diseases across the three fatty acid metabolites. Each of the 178 diseases was manually mapped by its name to a well-powered GWAS summary statistic available in Open GWAS. Two-sample Mendelian randomization was performed to assess the causal association between the three metabolites as exposures and the 178 diseases as outcomes. A nominally significant p-value threshold of 0.05 was applied using the Inverse Variance Weighted (IVW) MR method to determine significant associations. Because the suggestive fatty acid metabolite-disease associations were all protective, we only assessed MR associations with OR < 1. The OR for the suggestive and causal associations for each metabolite were then compared.

[5] Metabolite PRS x obesity associations

A set of disease-metabolite associations identified by the aforementioned interaction models assessing PUFA metabolites and weight measurements on disease, were further used to assess the potential events of canalization of the body weight measurements on PUFA metabolites and the impact of it on disease. Specifically, type 2 diabetes and ω3 fatty acids, obstructive chronic bronchitis and ω3FA, type 2 diabetes and DHA, chronic obstructive bronchitis and DHA, and diaphragmatic hernia and DHA. The cohort was divided into two cohorts for each body weight measurement, BMI and WHR defining obese and non-obese by standard cutoffs. The BMI and WHR obese and non-obese groupings were the same as in the previous section. The prevalence for the disease was compute for 100 bins of the PRSMetabolite for the obese and non-obese body weight measurement cohorts. To quantify the canalization, we computed delta observed, which is the difference between the right and left tail differences of the body weight obese and non-obese groups as further described in Nagpal et.al. We also computed delta expected, which is again the difference between the right and left tail differences but simulating the disease prevalence for 10 iterations, which was used in compute delta departure, a scaled estimate of the departure between observed delta and expected delta. More information on this approach is explained in Nagpal et.al.

[6] Assessing significance and concordance

A multiple testing p-value threshold of 6.28×10-6 (0.05/(1,326*6)) was applied for assessing the significance of the metabolite-disease non-genetic and genetic associations. Only the significant associations for non-genetic and genetic associations are demonstrated on this web app. A significant Mendelian randomization association was identified if there was a p < 0.05 (nominal significance) for the Inverse Variance Weighted (IVW) method. The IVW results are demonstrated on this website in the Mendelian randomization section. For the Metabolite PRS x obesity associations, only the metabolite x obesity associations are portrayed for the diseases with suggestive support (having significant p-value and concordant OR for metabolite level and metabolite PRS associations) for at least one of the three metabolites.