Møller et al. Genome Medicine (2024) 16:40 https://doi.org/10.1186/s13073-024-01313-8 RESEARCH Open Access © The Author(s) 2024. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecom- mons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data. Genome Medicine Predicting the presence of coronary plaques featuring high-risk characteristics using polygenic risk scores and targeted proteomics in patients with suspected coronary artery disease Peter Loof Møller1,2, Palle Duun Rohde2, Jonathan Nørtoft Dahl3,4, Laust Dupont Rasmussen3,10, Louise Nissen3,4, Samuel Emil Schmidt2, Victoria McGilligan5, Daniel F. Gudbjartsson6,7, Kari Stefansson6,8, Hilma Holm6, Jacob Fog Bentzon4,9, Morten Bøttcher3,4, Simon Winther3,4 and Mette Nyegaard2* Abstract Background The presence of coronary plaques with high-risk characteristics is strongly associated with adverse car- diac 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 probabil- ity 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. Measure- ments 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 (GPSMult). 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 GPSMult as input features. Results HRPs were detected in 165 (11%) patients, and 15 input features were associated with HRP presence. Predic- tion 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 GPSMult. Combining CRFs with GPSMult increased predic- tion 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 + GPSMult (AUC 74.6 ± 0.1, P = 1.00) models, respectively. *Correspondence: Mette Nyegaard nyegaard@hst.aau.dk Full list of author information is available at the end of the article Page 2 of 10Møller et al. Genome Medicine (2024) 16:40 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:// clini caltr ials. gov/ ct2/ show/ NCT02 264717 (September 2014). Keywords High-risk coronary plaque, Coronary artery disease, Prediction, Genetics, Olink proteomics Background Coronary computed tomography angiography (CCTA) is guideline-endorsed for identifying patients with suspected obstructive coronary artery disease (CAD) [1, 2]. Additionally, CCTA enables risk stratification including detection of high-risk plaques (HRP) predis- posing patients to adverse cardiac events [3–5]. HRPs are defined as having two or more characteristics on the CCTA associated with a higher likelihood of hav- ing an acute coronary syndrome (ACS) [6]. These characteristics include low attenuation plaque, spotty calcification, positive remodeling, and the napkin ring sign. In patients with stable chest pain suggestive of obstructive CAD, pre-test probability (PTP) estimation is recommended to guide decisions about downstream testing [1, 2]. Classically, PTP estimation is based on sex, age, and chest pain typicality [1]. In general, both patient stratification, diagnostic value, and risk prediction have been improved by also considering clinical risk factors (CRFs) [7, 8] in the PTP estimation. The majority of de novo chest pain patients have low PTP and are recom- mended to undergo index testing by CCTA to identify potentially obstructive lesions [1, 2, 7]. Evidence of HRP characteristics has been documented in both patients with obstructive and non-obstructive lesions which CCTA outlines [4, 9], and as early treatment of patients with HRP characteristics could improve prognosis [10], an optimized clinical tool to guide CCTA utilization for HRP identification is warranted. Based on large-scale proteomic panels and genotyping arrays yielding multi-protein models and genome-wide polygenic scores (GPSs), respectively, proteomics and genetics have been used for the prediction of obstructive CAD [11, 12], CAD-related traits [13], and general CAD risk management [14]. However, whether the combina- tion of proteomic and genetic data improves the predic- tion accuracy of HRP presence is unknown. Further, a high PRS of CAD has previously been shown to be most predictive in patients younger than 55 years of age [15]. It is currently unknown whether this is also the case for HRP presence. The primary aim of this study was to identify features (i.e., CRF, GPS, and proteins) associated with HRP pres- ence and evaluate a combination of clinical, proteomic, and genetic data to predict the trait in de novo chest pain patients with suspected obstructive CAD. Secondly, we aimed to uncover any age-related variation in predictive features. Methods Study population This is a sub-study of the Danish study of Non-Invasive testing in Coronary Artery Disease (Dan-NICAD) 1 study. The study protocol and main results have previ- ously been described [16, 17]. In short, Dan-NICAD 1 was a prospective, multicenter cross-sectional study of 1675 patients with no previous history of CAD, low- intermediate PTP, and symptoms suggestive of obstruc- tive CAD referred for initial testing by CCTA. On the day of CCTA, patients also underwent blood sampling, while symptoms and cardiac risk factors were registered. EDTA plasma was isolated from blood samples and stored at -80 °C. CCTA add HRP definition CCTA was performed on a 320-slice volume CT scan- ner (Aquilion One; Toshiba Medical Systems, Japan) following usual clinical guidelines. The presence and location of coronary plaques and CAD severity assess- ment were evaluated by a cardiologist [16]. CCTA plaque analysis was performed blinded to clinical risk factors, proteomics, and genetics, using the previously validated software Qangio CT (Research Edition ver. 3.1.4.2, Medis NL [18]). Obstructive CAD was defined as a 50% diameter stenosis by visual assessment at CCTA. Plaques were evaluated by dedicated and trained core laboratory personnel (Cardiac Imaging Center, Department of Cardiology, Goedstrup Hospital, Denmark) for four qualitative high-risk characteristics: (1) positive remod- eling (remodeling index > 1.10 calculated as vessel area at the location of plaque divided by reference vessel area in adjacent normal segments), (2) low-attenuation (non- calcified plaque with a plaque volume of > 1 mm [3] containing voxels with CT attenuation < 30 Hounsfield Units), (3) spotty calcification (< 3 mm calcium encased in non-calcified plaque), and (4) napkin ring sign (ring- like structure with lowest CT attenuation in the center of the plaque). The presence of HRP was defined as having coronary lesions with two or more high-risk Page 3 of 10Møller et al. Genome Medicine (2024) 16:40 characteristics, as recommended by the CAD-RADS 2.0 system [19]. Clinical risk factor model The clinical risk factor model was based on the fea- tures incorporated in the risk factor-weighted clinical likelihood (RF-CL) model of obstructive CAD [7]. The model is based on sex, age, angina typicality (typical, atypical, non-specific, and dyspnea), and the number of risk factors (family history of early CAD, smoking [never vs. active/former], dyslipidemia [receiving cho- lesterol-lowering medication], hypertension [receiv- ing antihypertensive medication], and type 2 diabetes) ranging from 0 to 5. Models utilizing CRFs included all of the above features in their input. Protein model In total, 368 proteins were measured in EDTA plasma samples at Olink proteomics AB (Uppsala, Sweden) as previously described [20]. The protein analysis was done using four Olink Target 96 panels: the Cardiovas- cular II, Cardiovascular III, Inflammation, and Immune Response panel. The Cardiovascular II and III panels were chosen for their relevance to cardiovascular pro- cesses, while the Inflammation and Immune Response panels were included to broaden the search for associa- tions. Samples marked in Olink’s internal quality con- trol steps were removed. Also, if the fraction of samples below the limit of detection (LOD) for a protein exceeded 20%, the protein was excluded from further analyses. For all included proteins, values were rank- normalized before adjusting for the collection box. A total of 300 proteins passed quality control and were included in the input to all models utilizing proteomic features. Genetic model Genotyping, quality control, and imputation have pre- viously been described [20]. In short, genotyping used the Illumina Global Screening Array, followed by impu- tation using the Michigan Imputation Server. Single nucleotide polymorphisms (SNPs) with minor allele frequencies ≤ 1% were excluded. The genotype weights of the multi-trait multi-ancestry genome-wide  polygenic score (GPSMult, https:// www. pgsca talog. org/ score/ PGS00 3725/) by Patel et  al. [21] were used to estimate genetic risk for CAD, as no dedi- cated HRP polygenic score exists. GPSMult aggregates multiple polygenic scores from CAD, BMI, ischemic stroke, diabetes mellitus, peripheral artery disease, glomerular filtration rate, systolic and diastolic blood pressure, LDL and HDL cholesterol, and triglycerides utilizing the 1,296,172 variants included in HapMap3. Combined models For combined models, the prediction was based on three groups of input features: (1) clinical risk fac- tors (nine features), (2) proteins (300 features), and (3) GPSMult (one feature), with the name of each model reflecting the included feature groups. Data analysis The primary outcome was defined as the presence of HRP. For individual features (i.e., individual clinical risk factors, individual proteins, and GPSMult), performance was measured using the receiver operating characteris- tic (ROC) with the area under the curve (AUC) and 95% confidence interval (CI) implemented in the pROC [22] R package. Significance testing was performed using the Wilcoxon rank sum test and a Bonferroni corrected sig- nificance threshold of 0.05/315. Models using more than one input feature were con- structed using logistic regression with elastic net regular- ization [23, 24] implemented in the glmnet and caret [25] R packages, using weights to compensate for case–con- trol imbalance. To limit the impact of random sampling noise, all models were based on 100 repeats of fivefold cross-validation (CV), foregoing a completely independ- ent test set. ROC was used as a summary metric to select the optimal model. Hyperparameters were chosen from a combination of three alphas (0.1, 0.55, 1) and 10 lambdas ranging from 0.0001 to 1. For all models, performance in every fold and repeat was stored, resulting in 500 AUC estimates for each model. Per- formance was reported as the mean ± standard error of all AUC estimates. As the AUC distributions for some mod- els were not Gaussian, comparisons of model performance used the Kruskal–Wallis rank sum test to compare all mod- els followed by Bonferroni corrected Dunn’s test for individ- ual post-hoc comparisons, reporting the adjusted P value. Estimation of predictive performance in plaque sub- types was performed for all models. For each model, the HRP probability was averaged per individual across the 100 repeats, before calculation of the AUC with 95% CI for single features. All analyses were performed within R [26] version 4.2.1. SHapley Additive explanation (SHAP) values were estimated using the fastshap [27] R package, using 100 Monte Carlo repetitions. Plotting was performed using ggplot2 [28]. Results Study population Of the 1675 eligible patients, 1462 (87%) had complete data on clinical risk factors, proteomics, genomics, and CCTA. Baseline demographics are shown in Table  1. Page 4 of 10Møller et al. Genome Medicine (2024) 16:40 Baseline demographics of excluded patients are shown in Table S1, while the reason for exclusion is shown in Fig S1. High-risk characteristics including positive remod- eling, low-attenuation, spotty calcification, and nap- kin-ring sign were present in 309 (21%), 144 (10%), 181 (12%), and 36 (2%) patients, respectively (Fig.  1). In total, HRP presence was identified in 165/1462 (11%) patients, and 341/1462 patients (23%) had obstructive CAD. HRP presence was significantly correlated with obstructive CAD presence (Pearson’s correlation = 0.49, P < 0.001), with 134 patients having both. Single clinical, protein, and genetic feature association with high‑risk plaque A total of 84 individual clinical, protein, and genetic features had a 95% CI lower limit of the AUC estimate above 50% (Fig. 2A). Among those, 14/84 (17%) features were significantly associated with HRP presence when accounting for multiple testing: three clinical risk fac- tors (age, sex (male) and number of risk factors (0–5)), ten proteins (MMP12, TREM1, MMP-3, KIM1, CDCP1, PRSS8, LEP, GDF-15, PIgR, and COL1A1), and GPSMult. Stratifying patients by GPSMult quintiles revealed increased HRP prevalence with increasing polygenic Table 1 Baseline information Abbreviations: CAD coronary artery disease, CCTA coronary computed tomography angiography, LDL low-density lipoprotein, HDL high-density lipoprotein Values are listed as mean ± standard deviation for normally distributed data; otherwise, the median and interquartile range are used a Missing values were observed in total cholesterol n = 41, LDL cholesterol n = 41, HDL cholesterol n = 38, triglyceride n = 46, systolic and diastolic blood pressures n = 3, body mass index n = 8, coronary artery calcium score n = 1 Overall (n = 1462) High‑risk plaque (n = 165) No high‑risk plaque (n = 1297) Demographics Age, years 57 ± 9 60 ± 8 57 ± 9 Males 699 (48%) 122 (74%) 577 (44%) Risk factors, n (%) Family history 527 (36%) 63 (38%) 464 (36%) Current/former smoker 778 (53%) 104 (63%) 674 (52%) Dyslipidemia 349 (24%) 42 (25%) 307 (24%) Hypertension 517 (35%) 69 (42%) 448 (35%) Type 2 diabetes 85 (6%) 16 (10%) 69 (5%) Type of chest pain, n (%) Typical angina 393 (27%) 63 (38%) 330 (25%) Atypical angina 497 (34%) 47 (28%) 450 (35%) Non-specific chest discomfort 271 (19%) 23 (14%) 248 (19%) Dyspnea 301 (21%) 32 (19%) 269 (21%) Laboratory tests Cholesterol medication 349 (24%) 42 (25%) 307 (24%) Yes No Yes No Yes No Total cholesterol, mmol/La 4.9 ± 1.2 5.5 ± 1.0 5.2 ± 1.6 5.8 ± 1.4 4.9 ± 1.1 5.5 ± 1.0 LDL cholesterol, mmol/La 2.8 ± 1.1 3.4 ± 0.9 3.1 ± 1.2 3.7 ± 0.9 2.8 ± 1.0 3.4 ± 0.9 HDL cholesterol, mmol/La 1.4 ± 0.4 1.5 ± 0.5 1.4 ± 0.4 1.4 ± 0.5 1.5 ± 0.4 1.5 ± 0.4 Triglyceride, mmol/La 1.5 [1.0–2.1] 1.3 [0.9–1.9] 1.6 [1.0–2.3] 1.5 [1.1–2.3] 1.5 [1.0–2.1] 1.3 [0.9–1.9] Measurements Blood pressure medication 517 (35%) 69 (42%) 448 (35%) Yes No Yes No Yes No Systolic blood pressure, mm Hga 143 ± 19 136 ± 18 144 ± 18 146 ± 19 143 ± 19 135 ± 18 Diastolic blood pressure, mm Hga 84 ± 11 82 ± 11 83 ± 11 86 ± 11 85 ± 11 81 ± 11 Body mass indexa 26.8 ± 4.2 26.3 ± 3.6 26.8 ± 4.3 Obstructive CAD at CCTA 341 (23%) 134 (81%) 207 (16%) Coronary artery calcium scorea 0 [0–81] 210 [52–624] 0 [0–42] Page 5 of 10Møller et al. Genome Medicine (2024) 16:40 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 (GPSMult), 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; GPSmult, 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 GPSMult 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 Page 6 of 10Møller et al. Genome Medicine (2024) 16:40 score, with the highest quintile having an odds ratio of 6.03 (P = 0.02) relative to the lowest quintile (Fig. 2B). Combined prediction models of high‑risk plaque Compared to the model based on CRF features, the prediction accuracy of HRP presence by the protein model was lower (AUC 73.2 ± 0.1 vs. 69.0 ± 0.1, P < 0.001) (Fig.  3A). The GPSMult model had an AUC of 60.1 ± 0.1, which was inferior to both the CRF and protein models. Combining GPSMult with CRFs or protein fea- tures increased prediction accuracy of HRP presence (CRF + GPSMult: 74.8 ± 0.1, P = 0.004; Pro- tein + GPSMult = 71.0 ± 0.1, P < 0.001). Combining CRF and protein features leads to the inclusion of eight features in total, four of these being proteins, but no improvement in the accuracy of prediction was observed compared to the CRF model (73.2 ± 0.1, P = 1.00). Simi- larly, a full model including all feature groups did not improve the CRF + GPSMult model (74.6 ± 0.1, P = 1.00). Across all models, using each model for predict- ing individual high-risk plaque characteristics showed the lowest predictive performance for low attenuation plaques and the highest for napkin ring sign (Table 2). Supplementary analysis defining HRP as 1- or 3-feature positive plaques (instead of 2-feature) found CRF to have the best predictive ability in 3-feature HRP (n = 29), while CRF + GPSMult had the best predictive ability for 1-fea- ture HRP (n = 388, Table S2). Age stratification Stratifying the cohort into patients above and below 55  years of age showed an HRP prevalence of 8.0% (50/627) in patients ≤ 55  years and 13.8% (115/835) in patients > 55 years. For all non-protein models, this strati- fication improved HRP prediction in patients ≤ 55  years (Fig.  3) relative to patients > 55  years. Especially the GPSMult among patients ≤ 55 displayed improved predic- tion accuracy significantly (AUC ≤55 = 66.5 ± 0.4 vs. AUC 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 + GPSMult model to be the most predictive in the full cohort. Stratifying patients by age group resulted in improved GPSMult performance in the group ≤ 55 years of age, while the CRF + GPSMult models were best in both groups. AUC, area under the curve; CRF, clinical risk factor; GPSMult, multi-trait multi-ancestry genome-wide polygenic score; Full, CRF + protein + GPSMult Page 7 of 10Møller et al. Genome Medicine (2024) 16:40 >55: 58.3 ± 0.2, P < 0.001). The best-performing model in both age groups was based on CRF and GPSMult, yield- ing improved prediction in younger patients (AUC ≤55 = 77.3 ± 0.3 vs. AUC >55 = 70.7 ± 0.02 (P < 0.001)). Predicting individual high-risk plaque characteristics in patients ≤ 55 years of age showed similar results as in the full cohort (Table S3). Comparing the impact of individual features (utiliz- ing SHAP values) on model output in the CRF + GPSMult models across the three groups (full cohort, ≤ 55 years of age, and > 55 years of age), sex was consistently the most influential single feature on HRP presence (Fig.  4A). Comparing the impact of CFRs to GPSMult revealed GPSMult to be equally important to the sum of risk fac- tors. For individual risk factors, family history of early CAD appeared protective in the full cohort, smoking had an impact in patients > 55  years of age, and choles- terol medication was protective in patients > 55  years of age and the full cohort (Fig. 4A). Finally, the use of anti- hypertensive medication and having diabetes both had a low impact on all model predictions. Typical chest pain symptoms indicated increased HRP risk in all groups, while nonspecific chest pain in patients ≤ 55 years of age and other chest pain in patients > 55 years of age and the full cohort indicated reduced HRP risk. Examining the individual SHAP values (Fig.  4B) revealed multiple features (age, number of risk factors, and GPSMult) with the potential to be more impactful than sex for a single patient although they were inferior from a cohort perspective. Discussion The main finding of the study was that HRP presence can be predicted using clinical risk factors, whereas genetic data primarily serves as a strong source of com- plementary information,  resulting in improved HRP prediction in the complete cohort and especially among patients below 55  years of age the prediction accuracy was significantly improved. Finally, we did not find that our proteomic data improved the overall discrimination, despite having better individual prediction accuracy than genetic data. Looking at individual input features, 14 features were significantly associated with HRP presence, with well- known clinical risk factors and proteins showing superior discrimination. Our findings are not surprising as several proteins are correlated with sex and age which indepen- dently correlates to HRP presence (Fig. 2A). In our study, we used protein measurements without adjusting the protein level for age and sex, as we wanted to retain these signals for models which did not incorporate age and sex directly. Regarding the combined models, CRF features were the most predictive feature group for HRP presence. In contrast, Bom et al. [13] reported 196 patients with 22% HRP prevalence and utilized four Olink panels, three of which were also analyzed in our study, and found pro- teins to be predictive of HRP presence (AUC  of 79 ± 1) with impaired prediction accuracy by CRFs alone (AUC of 65 ± 4). The difference in CRF predictive performance between our studies could be explained by our study having more detailed CRF data, e.g., the number of risk factors as implemented in the RF-CL model. Further- more, the higher predictive performance of proteomics observed by Bom et al. may be explained by their inclu- sion of the targeted Cardiometabolic Olink panel which was not included in our study. This highlights the possi- bility of future discoveries as proteomic panels include an increasing number of proteins. Due to the rarity of cohorts with detailed high-risk plaque characteristics, no suitable GWAS summary sta- tistic exists for the calculation of a traditional polygenic score for HRP. Instead, we leveraged GPSMult, an exist- ing polygenic score for CAD, which incorporates genetic information about multiple known CAD risk factors and from multiple ancestries. GPSMult showed the greatest Table 2 Plaque subtype prediction CRF clinical risk factor, GPSMult multi-trait multi-ancestry genome-wide polygenic score Numbers are the area under the curve with a 95% confidence interval Models Low attenuation (n = 144) Spotty calcification (n = 181) Positive remodeling (n = 309) Napkin ring sign (n = 36) CRF 71.1 [66.8–75.4] 72.7 [69.0–76.3] 71.9 [68.7–75.0] 76.9 [69.4–84.3] Protein 67.4 [62.9–71.8] 69.3 [65.3–73.4] 69.0 [65.6–72.3] 74.2 [66.5–81.9] GPSMult 59.2 [54.3–64.1] 60.2 [55.9–64.5] 62.7 [59.2–66.1] 61.1 [51.6–70.6] CRF + GPSMult 72.8 [68.6–77.0] 75.1 [71.7–78.5] 74.9 [71.9–77.8] 78.0 [70.0–85.9] Protein + GPSMult 69.1 [64.9–73.3] 71.5 [67.7–75.3] 71.9 [68.8–75.0] 76.3 [69.3–83.3] CRF + Protein 71.1 [66.6–75.6] 73.0 [69.3–76.8] 71.5 [68.3–74.7] 79.3 [72.7–85.9] Full 72.3 [68.3–76.4] 75.0 [71.5–78.6] 74.9 [72.0–77.9] 79.5 [72.8–86.2] Page 8 of 10Møller et al. Genome Medicine (2024) 16:40 prediction accuracy in patients below 55  years of age. This is consistent with previous findings by Mars et  al. reporting PRS for CAD prediction to have lower per- formance in people above 55  years of age [15] and also explains the somewhat low predictive ability of GPSMult by itself in our study (mean age of 60 years). Interestingly, we found that GPSMult was able to improve prediction when combined with either CRFs or proteins, indicating that genetics carry additional infor- mation about existing or novel aspects of HRP risk. Meanwhile, the inclusion of proteins with CRFs did not improve prediction, suggesting that the proteins assessed in our study do not contain additional information about HRP risk. It is possible that future improvements in HRP prediction could arise from increasing the number of included plasma proteins, using, e.g., SomaScan 7K or Olink Explore HT. One of the strengths of this study is the consecutive enrollment of nearly 1500 patients with detailed CCTA readings allowing for the discovery of HRP presence and specific plaque characteristics. Additionally, the Danish healthcare system is likely to reduce referral bias, as no direct payment is needed for citizens to undergo diag- nostic testing and therefore our findings represent an all- comer population. The primary limitation of the study is the lack of exter- nal validation. Instead, this study utilized internal testing through repeated cross-validation, enabling visualization of the uncertainty of our performance estimates (Fig. 3). Additionally, this study investigated only 300 proteins, meaning that additional studies, using larger, more explorative panels, are required to investigate the remain- ing plasma proteome. Fig. 4 Feature importance of the CRF + GPSMult 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; GPSMult, multi-trait multi-ancestry genome-wide polygenic score Page 9 of 10Møller et al. Genome Medicine (2024) 16:40 Conclusion Genetic data can be used to improve both clinical risk factors and proteomic models for the prediction of HRP presence, especially in patients below 55  years of age. However, a model combining both clinical risk factors and proteomics did not improve high-risk plaque iden- tification, despite both types of data being predictive on their own. Supplementary Information The online version contains supplementary material available at https:// doi. org/ 10. 1186/ s13073- 024- 01313-8. Additional file 1: Table S1. Baseline information. Table S2. Analysis of 1, 2, and 3 feature plaques. Table S3. Plaque subtype prediction in patients ≤55 years of age. Fig S1. Patient inclusion. Additional file 2: Table S4. Summary level data of number of risk factors, proteomics and GPSMult stratified by case/control status. Acknowledgements The authors would like to thank the Dan-NICAD study collaborators, including study nurses and clinical staff at the enrollment centers. Authors’ contributions The authors confirm the following contributions to the paper: study concep- tion and design: PLM, PDR, MN, MB, SW; data collection: LN, MB, SW, VM, DFG, KS, HH; analysis and interpretation of results: JND, LDR, PLM, PDR, MN; draft manuscript preparation: PLM, PDR, MN, JFB, SES. All authors reviewed the results and read and approved the final manuscript. Funding VM was supported by funding from the European Union Regional Develop- ment Fund (ERDF) EU Sustainable Competitiveness Programme for Northern Ireland; Northern Ireland Public Health Agency (HSC R&D). SW was sup- ported by the Novo Nordisk Foundation Clinical Emerging Investigator grant (NNF21OC0066981). MN was supported by the Novo Nordisk Foundation Start Package grants for faculty recruitment (NNF0071050). Availability of data and materials Summary level data of number of risk factors, proteomics, and GPSMult are available in Table S4. Individual level data cannot be made publicly available. Declarations Ethics approval and consent to participate The study follows the principles outlined in the Declaration of Helsinki. The study was approved by the Central Denmark Regional Committee on Health Research Ethics (record number: 1–10-72–190-14) and the Danish Data Protec- tion Agency (record number: 1–16-02–345-14). Informed written consent was obtained from all patients prior to enrollment. The trial was registered at https:// www. clini caltr ials. gov (unique identifier: NCT02264717). Consent for publication Not applicable. Competing interests MB discloses advisory board participation for NOVO Nordisk, AstraZen- eca, Pfizer, Boehringer Ingelheim, Bayer, Sanofi, Novartis, AMGEN, CLS-Behring, and Acarix. SES is a part‐time consultant in Acarix and minor shareholder of Acarix. DFG, KS, and HH are employees of deCODE genetics/Amgen Inc. The remaining authors have no competing interests. Author details 1 Department of Biomedicine, Aarhus University, Aarhus, Denmark. 2 Depart- ment of Health Science and Technology, Aalborg University, Aalborg, Denmark. 3 Department of Cardiology, Gødstrup Hospital, Herning, Denmark. 4 Department of Clinical Medicine, Aarhus University, Aarhus, Denmark. 5 Personalized Medicine Centre, School of Medicine, Ulster University, Derry, Northern Ireland. 6 deCODE Genetics/Amgen, Inc, Reykjavik, Iceland. 7 School of Engineering and Natural Sciences, University of Iceland, Reykjavik, Iceland. 8 Faculty of Medicine, University of Iceland, Reykjavik, Iceland. 9 Centro Nacional de Investigaciones Cardiovasculares, Madrid, Spain. 10 Department of Cardiol- ogy, Aalborg University Hospital, Aalborg, Denmark. Received: 29 August 2023 Accepted: 12 March 2024 References 1. Neumann FJ, et al. 2019 ESC Guidelines for the diagnosis and manage- ment of chronic coronary syndromes: the task force for the diagnosis and management of chronic coronary syndromes of the European Society of Cardiology (ESC). Eur Heart J. 2020;41:407–77. 2. Gulati M, et al. 2021 AHA/ACC/ASE/CHEST/SAEM/SCCT/SCMR guideline for the evaluation and diagnosis of chest pain: a report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. J Am Coll Cardiol. 2021;144:e368–454. 3. Shaw LJ, et al. Society of cardiovascular computed tomography / North American Society of Cardiovascular imaging - expert consensus docu- ment on coronary CT imaging of atherosclerotic plaque. J Cardiovasc Comput Tomogr. 2021;15:93–109. 4. Williams MC, et al. Coronary artery plaque characteristics associated with adverse outcomes in the SCOT-HEART study. J Am Coll Cardiol. 2019;73:291–301. 5. Ferencik M, et al. Coronary atherosclerosis, cardiac troponin, and interleukin-6 in patients with chest pain: the PROMISE trial results. JACC Cardiovasc Imaging. 2022;15:1427–38. 6. Motoyama S, et al. Plaque characterization by coronary computed tomography angiography and the likelihood of acute coronary events in mid-term follow-up. J Am Coll Cardiol. 2015;66:337–46. 7. Winther S, et al. Incorporating coronary calcification into pre-test assess- ment of the likelihood of coronary artery disease. J Am Coll Cardiol. 2020;76:2421–32. 8. Winther S, et al. Coronary calcium scoring improves risk prediction in patients with suspected obstructive coronary artery disease. J Am Coll Cardiol. 2022;80:1965–77. 9. Senoner T, et al. Added value of high-risk plaque criteria by coronary CTA for prediction of long-term outcomes. Atherosclerosis. 2020;300:26–33. 10. Dawson LP, Lum M, Nerleker N, Nicholls SJ, Layland J. Coronary ath- erosclerotic plaque regression: JACC state-of-the-art review. J Am Coll Cardiol. 2022;79:66–82. 11. Klarin D, Natarajan P. Clinical utility of polygenic risk scores for coronary artery disease. Nat Rev Cardiol. 2021;19(5):291–301. 12. Ferrannini E, et al. Differential proteomics of cardiovascular risk and coro- nary artery disease in humans. Front Cardiovasc Med. 2022;8:790289. 13. Bom MJ, et al. Predictive value of targeted proteomics for coronary plaque morphology in patients with suspected coronary artery disease. EBioMedicine. 2019;39:109–17. 14. Roberts R, Campillo A, Schmitt M. Prediction and management of CAD risk based on genetic stratification. Trends Cardiovasc Med. 2020;30:328–34. 15. Mars N, et al. Polygenic and clinical risk scores and their impact on age at onset and prediction of cardiometabolic diseases and common cancers. Nat Med. 2020;26:549–57. 16. Nissen L, et al. Danish study of Non-Invasive testing in Coronary Artery Disease (Dan-NICAD): study protocol for a randomised controlled trial. Trials. 2016;17:262. 17. Nissen L, et al. Diagnosing coronary artery disease after a positive coro- nary computed tomography angiography: the Dan-NICAD open label, parallel, head to head, randomized controlled diagnostic accuracy trial of cardiovascular magnetic resonance and myocardial perfusion scintigra- phy. Eur Heart J Cardiovasc Imaging. 2018;19:369–77. 18. De Graaf MA, et al. Automatic quantification and characterization of coro- nary atherosclerosis with computed tomography coronary angiography: Page 10 of 10Møller et al. Genome Medicine (2024) 16:40 cross-correlation with intravascular ultrasound virtual histology. Int J Cardiovasc Imaging. 2013;29:1177–90. 19. Cury RC, et al. CAD-RADSTM 2.0 - 2022 Coronary artery disease-reporting and data system: an expert consensus document of the Society of Cardiovascular Computed Tomography (SCCT), the American College of Cardiology (ACC), the American College of Radiology (ACR), and the North America Society of Cardiovascular Imaging (NASCI). J Cardiovasc Comput Tomogr. 2022;16:536–57. 20. Møller PL, et al. Sortilin as a biomarker for cardiovascular disease revisited. Front Cardiovasc Med. 2021;8:652584. 21. Patel AP, et al. A multi-ancestry polygenic risk score improves risk predic- tion for coronary artery disease. Nat Med. 2023;29(7):1793–803. 22. Robin X, et al. pROC: an open-source package for R and S+ to analyze and compare ROC curves. BMC Bioinformatics. 2011;12:77. 23. Friedman J, Hastie T, Tibshirani R. Regularization paths for generalized linear models via coordinate descent. J Stat Softw. 2010;33:1–22. 24. Simon N, Friedman J, Hastie T, Tibshirani R. Regularization paths for Cox’s proportional hazards model via coordinate descent. J Stat Softw. 2011;39:1–13. 25. Kuhn M. Building Predictive Models in R Using the caret Package. J Stat Softw. 2008;28(5):1–26. 26. R Core Team. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria. Version 4.2.1. 2023. https:// www.R- proje ct. org/. 27. Greenwell B. fastshap: Fast Approximate Shapley Values. R package ver- sion 0.1.0. 2023. https:// CRAN.R- proje ct. org/ packa ge= fasts hap. 28. Wickham H. ggplot2: Elegant Graphics for Data Analysis. New York: Springer-Verlag; 2016. Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in pub- lished maps and institutional affiliations.