Genomic risk score offers predictive performance comparable to clinical risk factors for ischaemic stroke

Abstract

Recent genome-wide association studies in stroke have enabled the generation of genomic risk scores (GRS) but their predictive power has been modest compared to established stroke risk factors. Here, using a meta-scoring approach, we develop a metaGRS for ischaemic stroke (IS) and analyse this score in the UK Biobank (n = 395,393; 3075 IS events by age 75). The metaGRS hazard ratio for IS (1.26, 95% CI 1.22-1.31 per metaGRS standard deviation) doubles that of a previous GRS, identifying a subset of individuals at monogenic levels of risk: the top 0.25% of metaGRS have three-fold risk of IS. The metaGRS is similarly or more predictive compared to several risk factors, such as family history, blood pressure, body mass index, and smoking. We estimate the reductions needed in modifiable risk factors for individuals with different levels of genomic risk and suggest that, for individuals with high metaGRS, achieving risk factor levels recommended by current guidelines may be insufficient to mitigate risk.

Fig. 1. Study design.

a Individual GRSs were derived in the UK Biobank training set (n = 11,995) using GWAS summary statistics for individual traits. b The metaGRS for ischaemic stroke was then derived by integrating individual GRSs using elastic-net cross-validation. c Validation of the metaGRS for ischaemic stroke was performed in the UK Biobank validation set (n = 395,393). UKB UK Biobank, GWAS genome-wide association study, GRS genomic risk score.

Fig. 2. Individual GRSs for stroke-related phenotypes and stroke outcomes correlate in several distinct clusters.

Shown is the partial Pearson correlation plot of individual GRSs in a random sample of 20,000 UK Biobank individuals. Estimates are from linear regression of each pair of standardised GRSs, adjusting for genotyping chip (UKB/BiLEVE) and 10 PCs. Stars indicate Benjamini–Hochberg false discovery rate < 0.05 (adjusting for 171 tests). GRSs were ordered via hierarchical clustering of the absolute correlation. Anthrop anthropometric, cardio cardiovascular (other than CAD), SBP systolic blood pressure, DBP diastolic blood pressure, Height measured height, BMI body mass index, T2D type 2 diabetes, 1KGCAD coronary artery disease from 1000 Genomes, 46K coronary artery disease from Metabochip, FDR202 coronary artery disease from 1000 Genomes (top SNPs), CES cardioembolic stroke, AS any stroke, IS ischaemic stroke, LAS large artery stroke, SVS small vessel stroke, TC total cholesterol, LDL low-density lipoprotein cholesterol, HDL high-density lipoprotein cholesterol, TG triglycerides, AF atrial fibrillation, Smoking cigarettes per day.

Fig. 3. The metaGRS identifies individuals at increased risk of ischaemic stroke.

Shown is the distribution of the metaGRS for ischaemic stroke in the UK Biobank validation set (n = 395,393), and corresponding hazard ratios. Hazard ratios are for the top metaGRS bins (stratified by percentiles) vs. the middle metaGRS bin (45–55%).

Fig. 4. The metaGRS for ischaemic stroke has comparable or higher predictive power than established risk factors.

Shown are the C-indices for incident stroke in the UKB validation set comparing the metaGRS with established risk factors. The reference model included the genotyping chip and 10 genetic PCs. Results are for the UKB validation set, excluding prevalent stroke events (n = 390,849). Red circles represent genetic/genomic scores; black circles represent non-genetic scores. Error bars represent 95% confidence intervals.

Fig. 5. Predicted cumulative incidence of ischaemic stroke.

Shown is the predicted cumulative incidence of IS in subjects with either (a) high levels of the metaGRS along with different risk factor levels (red: outside the guidelines; cyan: within the guidelines); or (b) risk factors within accepted guidelines along with different levels of the metaGRS (cyan: top 1% of the metaGRS; grey: middle 50% of the metaGRS; dark blue: bottom 1% of the metaGRS). Results are based on the UKB validation set, excluding prevalent stroke events (n = 390,849). Error bars represent 95% confidence intervals.