Proteomic Signatures as Biomarkers of Atherosclerosis Burden

TL;DR

This study identifies blood-based protein patterns that accurately measure atherosclerosis severity and predict future heart and stroke risks better than existing methods.

Contribution

The study introduces four novel proteomic signatures derived from machine learning to quantify atherosclerosis burden and improve cardiovascular risk prediction.

Findings

Four proteomic signatures achieved strong discrimination of atherosclerotic disease with ROC-AUC up to 0.92.

The signatures improved cardiovascular risk prediction beyond SCORE2 with a ΔC-index of +0.036.

Signature levels correlated with plaque burden and predicted future myocardial infarction and stroke in external cohorts.

Abstract

Atherosclerosis progresses silently over decades before manifesting clinically as myocardial infarction or stroke. Currently, no circulating biomarker reliably quantifies the burden of atherosclerosis beyond imaging techniques. Here, we sought to define plasma proteomic signatures that reflect the systemic burden of atherosclerosis. Using CatBoost machine learning applied to plasma proteomes (Olink Explore 3072; 2,920 proteins) from 44,788 UK Biobank participants, we derived four proteomic signatures which robustly discriminated individuals with known atherosclerotic disease from propensity score-matched controls (ROC-AUC up to 0.92, 95% CI: 0.90–0.94 in the test set). Each signature was based on distinct protein sets: the whole proteome (WholeProteome; n = 2920), proteins associated with genetic predisposition to atherosclerosis (Genetic; n = 402), those implicated in atherogenesis (Mechanistic; n = 680), and proteins enriched in arterial tissue (Arterial; n = 248). Among 41,200 individuals without atherosclerosis at baseline, all four signatures were strongly associated with future major adverse cardiovascular events over a median follow-up of 13.7 years (HR per SD increase in WholeProteome signature: 1.70, 95% CI: 1.64–1.77), providing significant improvements in risk discrimination (ΔC-index: +0.036; p <0.0001) and reclassification (Net Reclassification Index: 0.085–0.135 at a 10% risk threshold) beyond SCORE2. Signature levels increased with the number of clinically affected vascular beds, correlated with carotid ultrasound–measured plaque burden, and predicted future myocardial infarction and stroke in the external KORA S4 (n=1,361) and KORA-Age1 (n=796) cohorts with a median follow-up period of 15.1 and 6.8 years, respectively. Longitudinal analyses across three serial assessments showed that all signatures followed distinct trajectories, with significantly steeper annual increases among individuals with a higher burden of vascular risk factors. These findings demonstrate that proteomic signatures effectively capture atherosclerotic burden and improve cardiovascular risk prediction in asymptomatic individuals. Plasma proteomics may serve as a scalable and accessible alternative to imaging for identifying subclinical atherosclerosis, thereby supporting prevention strategies for cardiovascular disease.

Figures