Polygenic scores for low lung function and the future risk of adverse health outcomes

The Malmö Diet and Cancer study (MDCS) is a large prospective cohort study of men and women from Malmö, Sweden. Between 1991 and 1996, subjects aged 44–73 years living in Malmö were recruited to take part in the study. All men born between 1923 and 1945 and women born between 1923 and 1950 were invited to participate. Participation rate was approximately 40%. A total of 30,446 men and women underwent baseline examinations. All participants provided written consent, and the Ethical Committee at Lund University, Lund, Sweden approved the study (LU 51–90, LU 2009/633, LU 2011/356). Subjects with missing genotype data were excluded (n = 1151). Additionally, subjects with missing information on key variables for model adjustments were excluded (n = 1857). As exclusion of prevalent disease at baseline is not necessary for prediction of incident events if genetic risk is the exposure, we included prevalent cases where later a further event was recorded for the same subject for coronary events (CE) and chronic kidney disease endpoints but not SCD and diabetes. As SCD by definition is a sudden event in someone who is not known to previously have coronary heart disease, we excluded all prevalent cases of CE for the outcome of SCD. Prevalent cases of diabetes at baseline that occurred earlier than baseline were excluded for the purposes of this analysis to ensure that the majority of diabetes events included were type 2 diabetes mellitus (T2DM). The number of subjects examined for each outcome are illustrated in Fig. 1.

Standing height was measured using a fixed stadiometer calibrated in centimetres. Weight was measured whilst subjects wore light clothing and no shoes using a balance beam scale to the nearest 0.1 kg. Body mass index (BMI) was calculated as weight (kg)/height²(m). Blood pressure was measured using a mercury column sphygmomanometer after 10 min of rest in the supine position. Non-fasting blood samples were taking at enrolment. Serum and plasma were separated within one hour and stored at -80 degrees centigrade. Serum concentrations of Apolipoprotein A1 (Apo A-1) and Apolipoprotein B (Apo B) were measured by Quest Diagnostics (San Juan Capistrano, CA) using an immunonephelometric assay run on the Siemens BNII (Siemens, Newark, DE). Smoking status was ascertained from self-administered questionnaire responses. Higher education was defined as gaining university level education. Low physical activity was defined as the lowest quintile of the physical activity score, described in detail elsewhere [20]. Use of inhaled corticosteroids was defined as R03BA codes in the ATC (Anatomical, Therapeutic, Chemical classification) register. Baseline prevalent diabetes was defined as self-reported history of physician’s diagnosis of diabetes or use of anti-diabetic medication. Baseline prevalent CE and CKD were defined using national registers. No lung function was measured at the baseline MDCS but a subset of participants were re-examined as part of the Malmö Preventive Project cohort (MPP), where spirometry was measured. (n = 14 961 with FEV₁ (L) and 14 979 with FVC (L)).

MDCS participants were genotyped using Illumina GSA v1 genotyping array. Quality controls (QC) on the genotyped variants were conducted by removing variants for probe to genome mismatch, incorrect assignment of allelic variant, failed genotype calling, call rate less than 99%, minor allele frequency (MAF) less than 0.01 or failed Hardy-Weinberg Equilibrium test at p < 1 × 10^− 15. For sample QC, samples were excluded if genetic sex did not match pedigree sex or had an overall sample call rate less than 90%. The FlashPCA [21] was implemented for principal component analysis (PCA) on the genotyped data. Imputation was performed using the reference panel of Haplotype Reference Consortium (HRC r1.1) [22].

A Bayesian regression framework PRS-continuous shrinkage (PRS-CS) [23] was applied to build PGS for lung function. PRS-CS had shown superior performances in the prediction of common diseases and quantitative traits [23]. PRS-CS was used to infer posterior single-nucleotide polymorphism (SNP) effect sizes based on linkage disequilibrium reference panel of 1000 Genomes Project (European sample, n = 503) and GWAS summary statistics on FEV₁ (n = 400 102) and FVC (n = 400 102) from meta-analysis on the UK Biobank and SpiroMeta, respectively, where age, age squared, sex, height and smoking was adjusted in the association analyses [11]. We restricted SNPs that are biallelic, having MAF no less than 0.01 and consistent directions in sub studies of GWAS of meta-analysis. Using the PLINK (version 1.90) [24], PGS for FEV₁ and FVC on 29 295 MDC samples were then constructed by aggregating the obtained posterior effect sizes for FEV₁ (635 689 SNPs) and FVC (634 755 SNPs).

Through the study, PGS for FEV₁ or FVC are a relative measure of lung function in arbitrary units. High PGS represent high measures of FEV₁ or FVC, implying high lung function and vice versa.

SCD was defined as a fatal CE where death took place within the first 24 h, in individuals without a previous CE. This included those cases where death occurred outside of hospital. International Classification of Diseases (ICD) 9 codes included 410, 412 and 414; and ICD 10 codes I21, I24 and I25 [25, 26]. Major adverse cardiovascular events (MACE) were defined as CE, coronary artery bypass graft (CABG) or percutaneous coronary intervention (PCI) with ICD 9 codes 410, 412, 414 and ICD 10 codes I20, I21, I24 and I25. For CE, data linkage with the National Cause of Death Registry, Swedish Hospital Discharge registry and the Malmö Myocardial Infarction Register was used to retrieve cases [27]. Incident diabetes was defined using the Malmö HbA1c register (MHR), the Swedish national Diabetes register (NDR), the Swedish hospital discharge register, the Swedish outpatient register, the cause of death register, the Swedish drug prescription register, the regional Diabetes 2000 register of the Scania region and the All New Diabetics in Scania (ANDIS) registry. Incident cases of diabetes were also retrieved from re-examination of MDCS subjects in the MPP (2002–2006), MDCS baseline screening in the cardiovascular cohort (1992–1994), MDCS 5 year rescreening (1997–2001) and MDC cardiovascular rescreening (2007–2012). Incident CKD events were defined as ICD-10 codes N18 or N19. The Swedish inpatient registry had been in operation during the entire follow-up period and data from this registry has been found to have acceptable validity for epidemiological research [28]. All subjects were followed from the baseline examinations until the event of interest, death from other causes, emigration or last follow-up date (31st December 2019), whichever came first.

All statistical analyses were carried out using SPSS V.26 (IBM, Armonk, New York, USA) and STATA v14.0. Cox regression models were implemented to obtain hazard ratios (HR) for all endpoints by quartiles of FEV₁-PGS and FVC-PGS (Q1 highest PGS –reference) and by 1-SD increase of FEV₁-PGS and FVC-PGS. Adjustments were kept to a minimum but were made for potentially key confounders known from the literature (Model 1: unadjusted, Model 2: age, sex, height, weight, smoking, sBP, prevalent diabetes (except for the outcome of diabetes), and the first 5 principal components (PCs) of population structures. The outcomes of diabetes and hypertension were additionally adjusted for inhaled corticosteroids. PC allow us to adjust for underlying population structures, where genetic ancestry may explain associations between variants and a specific phenotype. As Apo A-1 and Apo B levels were available for fewer subjects, a further analysis was carried in these subjects after adjustment for Apo A-1 and Apo B (number of subjects by outcome with apolipoproteins available= SCD = 26,035, MACE = 26,273, diabetes = 25,376, CKD = 26,534, all-cause mortality = 26,536 and sBP = 26,536). Time dependent covariate analysis and Kaplan Meier curves were used to assess the proportional hazards assumptions. Using the time-dependent covariate analysis, proportional hazards assumptions were fulfilled for all adjusted outcomes. Univariate regression models were used to assess mean sBP by quartiles of FEV₁-PGS and quartiles of FVC-PGS, adjusted for potential confounders known from the literature. A p-value of < 0.05 was considered statistically significant, however, to take into account the role of multiple testing, a Bonferroni adjusted p-value of < 0.01 was used to assess p-trend across quartiles.

We assessed the improvement of prediction models of coronary heart disease and diabetes after the addition of PGS for lung function to the prediction models. Performance of the models was assessed using log-likelihood ratio tests and a category-free net reclassification index (NRI) [29]. For incident MACE, risk prediction was carried out in non-diabetic (at baseline) subjects, with no known intermittent claudication and no known previous CE (n = 22,065 subjects). Covariates in the risk prediction model included age, sex, smoking status, sBP, ApoA1 and ApoB and BP medication. For incident diabetes, risk prediction was carried out in non-diabetic subjects (at baseline) (n = 25,655 subjects). Covariates in the risk prediction model included age, sex, sBP, parental history of diabetes, ApoA1, ApoB and BMI.