Predicting the presence of coronary plaques featuring high-risk characteristics using polygenic risk scores and targeted proteomics in patients with suspected coronary artery disease

Abstract

Background: The presence of coronary plaques with high-risk characteristics is strongly associated with adverse cardiac events beyond the identification of coronary stenosis. Testing by coronary computed tomography angiography (CCTA) enables the identification of high-risk plaques (HRP). Referral for CCTA is presently based on pre-test probability estimates including clinical risk factors (CRFs); however, proteomics and/or genetic information could potentially improve patient selection for CCTA and, hence, identification of HRP. We aimed to (1) identify proteomic and genetic features associated with HRP presence and (2) investigate the effect of combining CRFs, proteomics, and genetics to predict HRP presence.

Methods: Consecutive chest pain patients (n = 1462) undergoing CCTA to diagnose obstructive coronary artery disease (CAD) were included. Coronary plaques were assessed using a semi-automatic plaque analysis tool. Measurements of 368 circulating proteins were obtained with targeted Olink panels, and DNA genotyping was performed in all patients. Imputed genetic variants were used to compute a multi-trait multi-ancestry genome-wide polygenic score (GPS_Mult). HRP presence was defined as plaques with two or more high-risk characteristics (low attenuation, spotty calcification, positive remodeling, and napkin ring sign). Prediction of HRP presence was performed using the glmnet algorithm with repeated fivefold cross-validation, using CRFs, proteomics, and GPS_Mult as input features.

Results: HRPs were detected in 165 (11%) patients, and 15 input features were associated with HRP presence. Prediction of HRP presence based on CRFs yielded a mean area under the receiver operating curve (AUC) ± standard error of 73.2 ± 0.1, versus 69.0 ± 0.1 for proteomics and 60.1 ± 0.1 for GPS_Mult. Combining CRFs with GPS_Mult increased prediction accuracy (AUC 74.8 ± 0.1 (P = 0.004)), while the inclusion of proteomics provided no significant improvement to either the CRF (AUC 73.2 ± 0.1, P = 1.00) or the CRF + GPS_Mult (AUC 74.6 ± 0.1, P = 1.00) models, respectively.

Conclusions: In patients with suspected CAD, incorporating genetic data with either clinical or proteomic data improves the prediction of high-risk plaque presence.

Trial registration: https://clinicaltrials.gov/ct2/show/NCT02264717 (September 2014).

Keywords: Coronary artery disease; Genetics; High-risk coronary plaque; Olink proteomics; Prediction.

Fig. 1

Study design. 1462 patients underwent coronary computed tomography angiography (CCTA), followed by image analysis of high-risk plaque (HRP) characteristics. Finally, nine clinical risk factors, one multi-trait multi-ancestry genome-wide polygenic score (GPS_Mult), and 300 proteins were used to predict HRP presence

Fig. 2

Predictive performance of single features. A Area under the curve (AUC) for individual features, grouped by feature type. Error bars indicate a 95% confidence interval (CI). Asterisks indicate statistical significance after Bonferroni correction for multiple testing. CRF, clinical risk factors; GPS_mult, multi-trait multi-ancestry genome-wide polygenic score. Only single features with a lower limit of 95% CI above 50% are shown. B Prevalence of high-risk plaque (HRP) stratified by GPS_Mult quintiles. Odds ratios are calculated using the first quintile as a reference. Asterisk indicates a statistically significant difference in the estimated odds ratio compared to the reference quintile

Fig. 3

Predictive performance of individual and combined models. Models included features from up to three feature groups, as shown on the x-axis, revealing the CRF + GPS_Mult model to be the most predictive in the full cohort. Stratifying patients by age group resulted in improved GPS_Mult performance in the group ≤ 55 years of age, while the CRF + GPS_Mult models were best in both groups. AUC, area under the curve; CRF, clinical risk factor; GPS_Mult, multi-trait multi-ancestry genome-wide polygenic score; Full, CRF + protein + GPS_Mult

Fig. 4

Feature importance of the CRF + GPS_Mult models across cohort groups. A Mean absolute SHAP values, representing the average importance of input features on model prediction, with blue features leading to lower risk and red features leading to higher risk. B Patient-level SHAP values, showing how impactful some features can be in extreme cases. SHAP, Shapley additive explanation; T2D, type 2 diabetes mellitus; GPS_Mult, multi-trait multi-ancestry genome-wide polygenic score