Abstract

   Cardiovascular disease (CVD) remains a leading cause of global
   morbidity and mortality. Timely diagnosis is important in reducing both
   short and long-term health complications. Saliva has emerged as a
   potential source for biomarker discovery, offering a non-invasive tool
   for early detection of individuals at elevated risk for CVD, yet
   large-scale extensive proteomic analysis using saliva for a
   comprehensive biomarker discovery remains limited. In an effort to
   develop a diagnostic tool using saliva samples, our study aims to
   assess the salivary and plasma proteomes in subjects with high risk of
   developing CVD using a large-scale proteomic approach. Leveraging on
   the SOMAscan platform, we analyzed 1,317 proteins in saliva and plasma
   collected from subjects at a high risk of CVD (HR-CVD) and compared the
   profiles to subjects with low risk of CVD (LR-CVD). Our analysis
   revealed significant differences in the plasma and salivary proteins
   between the two groups. Pathway enrichment analysis of the
   differentially detected proteins revealed that the immune system
   activation and extracellular matrix remodeling are the most enriched
   pathways in the CVD-HR group. Comparing proteomic signatures between
   plasma and saliva, we found approximately 42 and 17 differentially
   expressed proteins associated with CVD-HR uniquely expressed in plasma
   and saliva respectively. Additionally, we identified eight common
   CVD-risk biomarkers shared between both plasma and saliva,
   demonstrating promising diagnostic tools for identifying individuals at
   high risk of developing CVD. In conclusion, saliva proteomics holds a
   significant promise to identify subjects with a high risk to develop
   CVD. Further studies are needed to validate our findings.

   Keywords: Proteome, Saliva, Qatar biobank, CVD, Qatar GenomeProject,
   Machine Learning

   Subject terms: Diagnostic markers, Biomarkers, Cardiology,
   Cardiovascular biology, Proteomics

Introduction

   Cardiovascular disease (CVD) is a group of conditions affecting the
   heart and blood vessels, such as heart failure, hypertension, stroke,
   coronary heart disease, and atherosclerosis^[36]1. CVD stands as a
   leading cause of mortality worldwide, responsible for approximately
   one-third of all deaths, and is recognized as the primary
   noncommunicable disease^[37]2. Furthermore, the CVD direct medical cost
   is predicted to increase to $818 billion by 2030 compared to $273
   billion in 2010 in the United States^[38]3. Various risk factors
   contribute to the development of CVD, including increased body mass
   index (BMI), smoking, diabetes, high levels of low-density lipoprotein
   cholesterol, bad dietary habits^[39]2, and inflammation^[40]4. The
   prevalence of these risk factors differs among populations^[41]5.

   Data from the planning and statistical authorities in Qatar indicate
   that CVD is among the leading causes of mortality in 2020, contributing
   to 29% of all deaths in the country^[42]6. This is mainly attributed to
   the prevalent risk factors for CVD in Qatar^[43]7. A recent study
   showed that one in every five Qatari subjects is either prediabetic or
   diabetic, and one in three is hypertensive^[44]8. These risk factors
   are predicted to substantially increase by 2050^[45]9. A recent study
   examined the expected burden of CVD on diabetes over the next 10 years
   in Qatar and predicted a direct cost of 11.40 billion US$ and an
   indirect cost to surpass 8.30 billion US$^[46]10. Despite these
   concerning statistics, published studies investigating CVD risk among
   the Qatari and Arab populations at large are still scarce^[47]11. The
   early identification of individuals who are at high risk for developing
   CVD is crucial for early and cost-effective intervention^[48]12,[49]13.
   Therefore, there is a high demand for diagnostic biomarkers as they
   help in the early detection of diseases^[50]14. In our previous study,
   we assessed the association between the salivary microbiome and CVD
   risk using a large cohort of Qatar Genome Project (QGP)
   participants^[51]11. We showed significant differences in the salivary
   microbiome composition between HR and LR CVD subjects^[52]11.

   Recent advancements in protein assays allowed high throughput proteomic
   profiling, enabling the rapid discovery of new biomarkers by examining
   large numbers of proteins involved in various biological
   pathways^[53]12,[54]15. Among the pioneering high-throughput proteomic
   platforms utilized extensively in epidemiological and clinical
   investigations is the SOMAscan platform^[55]16. This platform employs
   single-stranded RNA or DNA sequences, termed aptamers, capable of
   recognizing epitopes on folded proteins^[56]16. With the capacity to
   analyze approximately 7,000 proteins in a relatively small sample
   volume, the SOMAscan platform offers exceptional sensitivity and
   reproducibility^[57]16. Its application has proven instrumental in
   identifying protein signatures linked to various diseases and potential
   biomarkers, including those associated with glomerulonephritis^[58]17,
   cancer^[59]18, Parkinson’s disease^[60]19, asthma^[61]20, and systemic
   sclerosis^[62]21among others. In CVD, SOMAscan was used to search for
   new CVD biomarkers in population-based studies like the Heart and Soul
   study (USA)^[63]22, Framingham Heart Study (USA)^[64]15,[65]23, Jackson
   Heart Study (US)^[66]24, in addition to other cohorts from
   Iceland^[67]25and Italy (InCHIANTI study)^[68]26. These studies have
   identified new CVD biomarkers^[69]15,[70]22–[71]24but were mainly done
   on populations with European ancestry^[72]15,[73]22,[74]23,[75]25. It
   is also worth noting that in CVD, large-scale proteomic profiling is
   mainly performed in the blood, and the proteome of other body fluids
   remains lacking^[76]27.

   To date, approximately 3,000 human saliva proteins have been
   identified, encompassing various enzymes, immunoglobulins,
   glycoproteins, and hormones, which collectively contribute to the
   maintenance of oral cavity homeostasis^[77]28. Saliva, often regarded
   as a "mirror of the gut" holds promise for diagnostic applications due
   to its diverse molecular composition derived from local blood supply,
   microbes, and cellular constituents^[78]29. Moreover, saliva is an
   accessible and non-invasive sample that is easy to collect, and as a
   result, salivary biomarkers can be applied for developing rapid
   diagnostic tools^[79]27. Despite these advantages, aptamer-based
   methods in saliva have primarily focused on cardiac markers such as
   troponins and myoglobin, while a comprehensive proteomic analysis for
   CVD protein signatures remains sparse^[80]27. The present study aims to
   leverage the SOMAscan proteomic panel to uncover biomarkers associated
   with CVD risk in both saliva and plasma samples from the Qatari
   population, thus laying the groundwork for identifying non-invasive
   salivary biomarkers for CVD risk and enhancing our understanding of
   proteomic signatures linked to CVD risk in saliva.

Methods

Ethical statement

   Approval for the study was obtained from the Institutional Review Board
   (IRB) of Sidra Medicine under protocol #1510001907, and from Qatar
   Biobank (QBB) under protocol #E/2018/QBB-RES-ACC-0063/0022. Prior to
   sample collection, all study participants signed an informed consent,
   and the experiments were conducted in accordance with the approved
   guidelines and in accordance with the Declaration of Helsinki.

Study population and clinical data

   Cardiovascular disease (CVD) risk scores were computed to assess the
   risk of suffering a heart attack over the next 10 years, using the Cox
   proportional-hazards regression, as detailed in our previous
   report^[81]11. From the same study cohort, we randomly selected 50
   subjects with low-risk to develop CVD (CVD score < 10) (CVD-LR) and 50
   subjects with high-risk (HR) (CVD score > 20) (CVD-HR)^[82]11. The
   study included Qatari participants aged 18–64 years. Participants were
   excluded if they had recent antibiotics use (within three months) or
   suffered from chronic diseases (e.g., gastroesophageal reflux disease,
   Crohn’s disease, thyroid disease or cancer). De-identified samples,
   along with anthropometric and clinical data for all study subjects,
   were collected from QBB. In brief, enrolled subjects were advised to
   fast for at least 8 h before the collection of samples. Matched plasma
   and saliva samples were collected from the same subjects following QBB
   standardized sample collection protocol^[83]30,[84]31. Around 60 ml of
   blood was collected and used for routine blood tests. Then, the
   remaining was aliquoted and stored at −80°C^[85]30,[86]31. For saliva,
   about 5 mL of unstimulated saliva was collected in a falcon tube,
   divided into aliquots of 0.4 mL, and stored at −80°C^[87]32.

SOMAscan proteomics

   The salivary and plasma proteome was characterized using the SOMAscan
   platform, which uses single-stranded DNA-based protein affinity
   reagents called SOMAmers (Slow Off-rate Modified Aptamers), as detailed
   in previous studies^[88]33–[89]35. In essence, each SOMAmer® reagent
   selectively binds to a specific target protein, totaling approximately
   1317 proteins. The SOMAscan assay involves distributing SOMAmers into
   various sample dilution bins tailored to the analyzed matrix. These
   diverse distributions and dilution schemes are designed to ensure that
   analyte concentrations fall within the linear range of the assay for
   each SOMAmer. In conventional matrices such as plasma and serum,
   SOMAmers are split into 0.005%, 1%, and 40% dilution bins. However, for
   non-traditional matrices like saliva, specific dilution bins are not
   predefined, and samples are typically analyzed at a single dilution,
   with all SOMAmers assayed accordingly. To establish the optimal
   dilution for saliva samples, we conducted SOMAScan assays using pooled
   and individual saliva samples serially diluted from 40% to 0.3125%.
   This process facilitated the identification of the optimal saliva
   dilution, determined to be 10% diluted in assay buffer, resulting in
   average assay values falling within the mid-range of the dynamic range
   for each SOMAmer. Throughout the SOMAScan assay, adherence to the
   manufacturer’s cell and tissue protocol instructions was maintained.
   Relative fluorescence unit values obtained from SOMAscan were
   normalized against the hybridization control to correct for any
   systematic effects introduced during hybridization. The hybridization
   control factor was determined by pooling all samples from different
   plates. Median normalization was applied across all samples within the
   arrays, ensuring a successful assessment of signal intensity variance
   based on the hybridization controls.

SOMAscan data analysis

   The raw fluorescence data of 1,317 proteins were first normalized via
   quantile normalization using the “normalizeBetweenArrays” from the
   limma package (v3.56.2)^[90]36. UMAP analysis indicated that patient’s
   BMI and age have a very strong confounding effect in segregating
   samples. Consequently, a differential expression analysis was conducted
   using limma, with consideration given to age and BMI within the design
   matrix. We used the ‘lmFit’ function for multiple linear regression,
   followed by the ‘ebays’ function with parameter ‘robus = TRUE’ to
   compute moderate t-statistics, F-statistics, and log-odd ratios.
   P-values were adjusted using Benjamini & Hochberg method. Proteins
   showing p-values < 0.05 and at least 50%-fold change between low and
   high CVD risk patients in either plasma or saliva were selected.
   Additionally, differentially expressed proteins demonstrating a
   consistent trend of expression change between the two tissues and
   significant statistical changes in both were selected as the initial
   biomarker candidates. For visualization purposes, the effects of age
   and BMI were first regressed out. Samples were then clustered into two
   groups based on UMAP representations reflecting low and high age/BMI
   values. The resulting cluster IDs were utilized to regress out the
   effects of age and BMI using the ‘removeBatchEffect’ function from the
   limma package.

Functional enrichment analysis

   The enrichGO function from the R/Bioconductor clusterProfiler package
   (v4.8.3)^[91]37 was used to perform Gene Ontology (GO) and pathway
   enrichment analysis focusing on Biological Process ontologies. Only GO
   terms exhibiting an adjusted p-value < 0.05 were included in the
   analysis. Then, GO enrichment plots were generated utilizing the
   ggplot2 package.

Estimation of protein biomarkers importance using machine learning models

   The following machine learning models: Random Forest (RF)^[92]38,
   Elastic-net (eNet)^[93]39, Partial least squares via mixOmics
   (pls)^[94]40, XGBoost^[95]41, generalized linear model (GLM) and Radial
   Basis Function (RBF) kernel SVM^[96]42were respectively used to
   determine the predictive importance of each marker in classifying
   CVD-HR and CVD-LR patients. We used the tidymodels R package to train
   the different models^[97]43. Each model was trained using repeated
   cross validation (5-folds and 10 repeats in each fold). To avoid label
   unbalancing during training, the different cross-validation subsets
   were generated in a stratified manner. Hyper parameter tuning was done
   using a grid search algorithm. The RF, eNet, and pls models had the
   highest performance in all tissues (plasma, saliva) and were selected
   to calculate the mean importance of each marker in the three models.
   The variable importance of each model was scaled to be within [0,1]

Statistical analysis and visualization

   The demographic and clinical data of the study cohort were analyzed
   using GraphPad Prism (10.1.2). Mann–Whitney U tests were utilized to
   compare variables, including age, BMI, systolic and diastolic blood
   pressure, glucose level, HbA1C, lipid profile, total protein, albumin,
   urea, and creatinine. Next, the Chi-square test was employed to compare
   the impact of smoking and sex between the CVD-HR group and the CVD-LR
   group. Statistical significance was set at p-values less than 0.05. All
   statistical analysis were conducted using R version 4.3.1, with the
   limma package (version 3.56.2)^[98]36. The visualization of the results
   was carried out using ggplot2 and ComplexHeatmap R packages^[99]44.
   Venn diagrams were created using Intervene^[100]45.

Results

Characteristics of the Study Cohort

   The baseline demographic and clinical characteristics of the study
   cohort comprising a total of 100 individuals are listed in Table
   [101]1. Based on cardiovascular risk score, we selected 50 subjects
   categorized as CVD high-risk (CVD-HR) and 50 individuals classified as
   CVD low-risk (CVD-LR). Overall, the CVD-HR subjects had an average age
   of 55.32 ± 6.7 compared to 43.06 ± 7.6 in the CVD-LR group.
   Furthermore, a significantly higher proportion of smokers was observed
   in the CVD-HR group compared to the CVD-LR group. The CVD-HR group
   displayed markedly elevated systolic and diastolic blood pressure,
   glucose levels, HbA1C, lipid profile parameters, urea, and creatinine
   levels compared to the CVD-LR subjects. Our findings also revealed
   specific biases in various clinical characteristics, with systolic and
   diastolic blood pressure demonstrating a positive correlation with
   CVD-HR. Moreover, BMI and age were identified as contributors to
   patient segregation, as illustrated in Figure S1. Therefore, we
   controlled for age and BMI during the differential expression analysis
   of plasma and saliva proteomic profiles.

Table 1.

   Demographic and Clinical Characteristics of the Study Population.
   CVD high risk group (CVD-HR)
   N = 50 CVD low risk group (CVD-LR)
   N = 50 p value
   Age (years) 55.32 ± 6.7 43.06 ± 7.6  < 0.0001^a***
   Male/Female 36/14 25/25 0.024^b*
   BMI 31.15 ± 5.241 29.60 ± 4.638 0.096^a
   Systolic blood pressure 137.1 ± 17.64 113.0 ± 14.95  < 0.0001^a***
   Diastolic blood pressure 74.68 ± 10.22 66.98 ± 10.64 0.0002^a***
   Smoking, n 22^(44%) 12^(24%) 0.0348^b*
   Glucose(mmol/L) 8.938 ± 4.093 5.874 ± 2.316  < 0.0001^a***
   HBA 1C % 7.738 ± 2.117 5.750 ± 1.259  < 0.0001^a***
   HDL-Cholesterol(mmol/L) 1.496 ± 0.4493 1.135 ± 0.3237  < 0.0001^a***
   LDL-Cholesterol (mmol/L) 3.435 ± 1.230 2.899 ± 0.8931 0.0049^a**
   Cholesterol Total (mmol/L) 5.488 ± 1.243 4.972 ± 0.9963 0.0101^a*
   Triglyceride(mmol/L) 1.874 ± 0.8429 1.270 ± 0.6329  < 0.0001^a***
   Total Protein (gm/L) 72.46 ± 4.062 72.10 ± 3.655 0.59^a
   Albumin(gm/L) 44.28 ± 3.860 44.32 ± 3.560 0.66^a
   Urea(mmol/L) 5.004 ± 1.533 4.438 ± 1.786 0.0262^a*
   Creatinine(µmol/L) 74.40 ± 16.56 69.58 ± 21.78 0.0328^a*
   [102]Open in a new tab

   ^a Mann–Whitney U was used, ^b Chi-square was used. *P-value < 0.05,
   **P-value < 0.01, ***P-value < 0.001. Data were shown as mean ± SD or
   n, (%).

The plasma and salivary proteomes show differentially expressed proteins in
CVD-HR and CVD-LR subjects

   Differential expression analysis using plasma samples of 50 CVD-HR
   subjects in comparison to 50 CVD-LR subjects encompassing a total of
   1,317 proteins detected using SOMAscan. Among these proteins, in the
   plasma, a subset of 207 proteins exhibited significant differences
   (p-value < 0.05) between CVD-HR and CVD-LR (Figure S2). Subsequently,
   proteins displaying both p-values < 0.05 and at least a 50% fold-change
   (|FC|> 1.5) between the CVD-HR and CVD-LR groups were selected for
   visualization in a heatmap, as depicted in Fig. [103]1a. Approximately
   44 plasma proteins (21 increased and 23 decreased) demonstrated a
   significant differential expression with at least a 50% ( |FC|> 1.5)
   fold-change in plasma of CVD-HR and CVD-LR groups (Fig. [104]1a).

Fig. 1.

   [105]Fig. 1
   [106]Open in a new tab

   Heat maps of differentially expressed proteins in CVD-HR group.
   Hierarchical clustering heatmaps of proteins that are differentially
   expressed between the CVD-HR and CVD-LR groups in plasma (a) and saliva
   (b). The cohort age, BMI, systolic blood pressure (BP), diastolic blood
   pressure (BP), smoking, and gender are shown. Samples are clustered
   using ward. D2 hierarchical clustering method. Key: red = up-regulated,
   blue = down-regulated. Proteins with p-values < 0.05 and at least
   50%-fold change are shown. CVD-HR = 50 subjects, CVD-LR = 50 subjects.

   On the other hand, a total of 94 proteins exhibited significant
   differences (p-value < 0.05) in saliva samples when comparing the two
   groups (Figure S2). Among these, 25 salivary proteins demonstrated
   significant differential expression with at least a 50%-fold-change (18
   increased and 7 decreased) between the CVD-HR and CVD-LR groups, as
   illustrated in Fig. [107]1b.

Identification of common CVD-risk biomarkers in plasma and saliva

   The differentially expressed proteins in CVD-HR group were further
   examined to search for common CVD-risk biomarkers between plasma and
   saliva. We found eight proteins that showed correlated enrichment in
   both the plasma and saliva of CVD-HR subjects (Figs. [108]2 and
   [109]3). These potential biomarkers include Plexin B2 (PLXNB2), LDL
   receptor-related protein 1B (LRP1B), GDNF Family Receptor Alpha 1
   (GFRA1), acid phosphatase 5, tartrate resistant (ACP5), Chemokine (C–C
   motif) ligand 15 (CCL15), Complement Component 1, R Subcomponent (C1R),
   proteasome activator subunit 3 (PSME3) and kallikrein 5 (KLK5). The
   PLXNB2, LRP1B, GFRA1, ACP5, C1R, and CCL15 were upregulated in both
   saliva and plasma of CVD-HR compared to the CVD-LR groups (Figs. [110]2
   and [111]3). On the other hand, PSME3 was the only downregulated
   protein in the CVD-HR group in both plasma and saliva (Figs. [112]2 and
   [113]3). Interestingly, KLK5 showed a difference in the direction of
   change between saliva and plasma, as shown in Figs. [114]2 and [115]3.
   We next examined whether taking anti-diabetic, antihypertensive, or
   antilipidemic medications will influence the shared CVD-risk biomarkers
   (Figure S3). All the shared CVD-risk biomarkers showed significant
   differential expression in saliva and plasma samples of CVD-HR after
   correction for treatment, except plasma KLK5, which was significantly
   different in saliva but not in plasma samples (Figure S3).

Fig. 2.

   [116]Fig. 2
   [117]Open in a new tab

   Venn diagram depicting shared and distinct CVD-risk biomarkers between
   plasma and saliva in CVD-HR group. CVD-HR Plasma = 50 subjects, CVD-HR
   Saliva = 50 subjects.

Fig. 3.

   [118]Fig. 3
   [119]Open in a new tab

   Expression level of proteins with correlated enrichment in both plasma
   and saliva of CVD-HR compared to CVD-LR. The Wilcoxon test was used to
   compare the two groups. p-value < 0.05 was considered significant.
   CVD-HR = 50 subjects, CVD-LR = 50 subjects.

The prediction performance of CVD-risk biomarkers using machine learning
models

   To accurately examine the diagnostic ability of the selected biomarkers
   to distinguish CVD-HR from CVD-LR and identify a non-invasive biomarker
   that can be measured either in saliva or plasma we run an unbiased
   machine learning (ML) analysis. First, we started by identifying the
   best performing ML models on our data. Hence, we compared six ML
   models: Random Forest (RF), Elastic-net (eNET), mixOmics (pls),
   XGBoost, generalized linear model (GLM), and SVM (RBF) using 1,317
   proteins or the 8 selected markers as input and using 70% of samples
   for training and the other 30% as testing set (see methods).

   In plasma, the unbiased model trained on the 1,317 proteins (all
   features) gave better predictive power compared to the model trained
   using only the eight shared markers (selected features), as illustrated
   in Fig. [120]4a. However, interestingly, the restricted model (selected
   features) is still highly accurate and shows an AUC > 0.75, indicating
   that the previously selected markers still hold a very strong
   predictive power.

Fig. 4.

   [121]Fig. 4
   [122]Open in a new tab

   Estimation of CVD-risk biomarkers importance using machine learning
   models. Random Forest (RF), Elastic-net (eNET), mixOmics (pls),
   generalized linear regression model (GLM), XGBoost, and (RBF) SVM
   models were used to estimate the predive power markers in classifying
   CVD-HR and CVD-LR groups in plasma (a) and saliva (b). The model, all
   features, indicates the model learned using all the 1,317 proteins. The
   “selected features” model was built using the eight common markers. The
   average importance between RF, eNet, and pls models were used to rank
   the markers in plasma (c) and saliva (d). The variable importance of
   each model was scaled to be within [0,1]. CVD-HR = 50 subjects,
   CVD-LR = 50 subjects. Measured proteins = 1,317.

   Alternatively, the selected features model demonstrated better
   predictive power to the models using all the features, as shown in
   Fig. [123]4b. This suggests that the selected markers have a better
   predictive potential in saliva tissue, potentially indicating their
   suitability as non-invasive biomarkers.

   Moreover, the RF, eNet, and PLS models had the highest performance in
   both plasma and saliva, with an average AUC > 0.8 in plasma
   (Fig. [124]4a) and AUC > 0.7 in saliva (Fig. [125]4b). Next, we
   calculated the mean importance of each marker in the unbiased version
   of these three models (Table [126]S1). The variable importance of each
   model was scaled to be within [0,1] before averaging. Among the eight
   common CVD-risk biomarkers, the top three predictive biomarkers in
   plasma were LRP1B (median importance = 0.876309), PLXNB2 (median
   importance = 0.352254), and CCL15 (median importance = 0.328339),
   respectively (Fig. [127]4c and Table S1). Meanwhile, in saliva, the top
   three predictive biomarkers were C1R (median importance = 0.387032),
   LRP1B (median importance = 0.375685), and PLXNB2 (median
   importance = 0.266522) (Fig. [128]4d and Table S1).

   Across all the differentially expressed proteins in plasma and saliva
   of the CVD-HR group, plasma LRP1B (median importance = 0.876309) was
   the strongest CVD-risk predictive biomarker, followed by 14–3-3 protein
   (YWHAE) (median importance = 0.804179) and saliva Protein S100-A7
   (S100A7) (median importance = 0.541921) (Table S1).

Pathway enrichment analysis for the differentially expressed proteins

   To get insights into pathways involved in the differentially expressed
   proteins, the gene Ontology (GO) gene sets were analyzed by the
   enrichGO function from the R/Bioconductor cluster Profiler. The
   analysis revealed eight pathways enriched in plasma proteins of the
   CVD-HR group, as shown in Fig. [129]5a. The extracellular matrix
   organization and the extracellular structure organization were the most
   enriched pathways for the differentially expressed biomarkers in the
   plasma of the CVD-HR group. Similarly, for the saliva differential
   proteins, ten pathways were also enriched in the CVD-HR group
   (Fig. [130]5b), namely, the humoral immune response.

Fig. 5.

   [131]Fig. 5
   [132]Open in a new tab

   Pathways enrichment analysis for the differential proteins in the
   CVD-HR and CVD-LR groups. The 10 most enriched Gene Ontology (GO) terms
   in plasma (a) and saliva (b) are listed. The circle size represents the
   ratio of proteins/genes. The enrichment level is indicated by the color
   bar (blue to red) representing the p-value (0.050 to 0.000).
   CVD-HR = 50 subjects, CVD-LR = 50 subjects.

Discussion

   In summary, the pursuit of a reliable CVD biomarker detectable in
   bodily fluids presents a substantial promise for cardiovascular risk
   assessment, diagnostic accuracy, management strategy guidance, and
   prognosis prediction^[133]46,[134]47. Nonetheless, there remains an
   urgent requirement for non-invasive biomarkers capable of accurately
   predicting CVD risk^[135]12. Current biomarkers such as troponin,
   creatinine kinase, and myoglobin primarily rely on antibody-based
   detection methods^[136]27. Despite their high sensitivity and
   selectivity, antibody-based diagnostics are often costly and subject to
   batch-to-batch variability^[137]48,[138]49. Aptamer-based platforms
   have emerged as a compelling alternative to address the limitations
   inherent in antibody-based detection^[139]27. While aptamer-based
   technology has identified novel CVD biomarkers from plasma or serum
   samples, data pertaining to saliva-based CVD protein
   signatures^[140]15,[141]22,[142]23,[143]25,[144]26, particularly
   employing high-throughput proteomic methods like SOMAscan, are
   currently scarce^[145]14,[146]27. In the present study, we utilized the
   aptamer-based SOMAscan platform to analyze plasma and saliva samples
   obtained from the QGP participants. Through this approach, we assessed
   1,317 proteins and identified unique protein signatures associated with
   increased CVD risk in the Qatari population. Then, using machine
   learning models we evaluated the predictive power of the identified
   CVD-risk biomarkers. These findings hold considerable potential for the
   advancement of promising non-invasive CVD biomarkers and furnish
   invaluable insights into the proteomic alterations observed in plasma
   and saliva concerning CVD risk.

   In plasma, a larger number of proteins (207 proteins) showed
   association with CVD-HR compared to saliva (94 proteins) (Figure S2).
   Notably, upon comparing the CVD-HR proteomic signatures between plasma
   and saliva, distinct proteins linked with CVD risk were identified in
   each fluid (Figure S2). Predominantly, the differentially expressed CVD
   markers in both plasma and saliva belonged to the category of
   inflammatory proteins or were implicated in inflammatory processes
   (Table S1). Inflammation stands as a recognized risk factor in CVD
   pathogenesis^[147]4, and inflammatory proteins, such as cytokines, are
   present in the saliva^[148]50. However, it is plausible that the
   secretory function of the salivary glands may regulate the levels of
   inflammatory markers within saliva^[149]50,[150]51. Interestingly,
   saliva protein markers seem to more prominently reflect local
   inflammation compared to systemic inflammation, as evidenced by plasma
   markers^[151]52. This discrepancy may account for the distinct CVD-HR
   signatures observed in plasma and saliva as previously
   highlighted^[152]53.

   In our study, we identified eight candidate CVD-HR protein biomarkers
   shared between plasma and saliva, capable of distinguishing between
   CVD-HR and CVD-LR groups (Figs. [153]2 and [154]3). These shared
   biomarkers include LRP1B, C1R, CCL15, KLK5, GFRA1, PLXNB2, ACP5, and
   PSME3 (Fig. [155]3). Among these candidates, LRP1B or LDL
   receptor-related protein 1B emerges as the most promising biomarker,
   exhibiting the best predictive value (median importance = 0.876309)
   (Table S1). LRP1B belongs to the LDL receptor family and is prominently
   expressed in the brain, thyroid, and salivary glands^[156]54. LRP1B
   exhibits binding affinity to various extracellular proteins implicated
   in blood coagulation and lipoprotein metabolism, such as fibrinogen and
   lipoproteins carrying apoE^[157]55. Interestingly, our study
   underscores a notable elevation in saliva fibrinogen levels and plasma
   ApoE protein within the CVD-HR group (Table S1). Numerous
   investigations have elucidated the association of the LRP1B gene with
   obesity^[158]56. Interestingly, the LRP1B gene was reported in some
   genome-wide association studies (GWAS), with findings linking it to
   systolic blood pressure, particularly in Chinese and sub-Saharan
   African populations^[159]57,[160]58. LRP1B harbors an intronic single
   nucleotide polymorphism (SNP) linked to blood pressure regulation, and
   has a notable interaction effect with smoking^[161]59. LRP1B is
   abundantly expressed in the medial layer of coronary arteries, and
   genetic variations in LRP1B have been linked to the risk of coronary
   artery aneurysms in Kawasaki disease among Taiwanese cohorts^[162]60.
   Additionally, LRP1B protein plays a role in Alzheimer’s disease by
   modulating the cellular trafficking and localization of the amyloid
   precursor protein^[163]61. Furthermore, a significant increase in LRP1B
   protein level was reported in the serum of women with systemic
   sclerosis^[164]21. In the InCHIANTI population study from Italy, LRP1B
   was inversely associated with cardiovascular health^[165]26. The
   current study reports the association of LRP1B protein expression in
   saliva and plasma with CVD-HR (Fig. [166]3). Moreover, our data
   highlights LRP1B as a potential biomarker for CVD risk, boasting the
   highest predictive accuracy (Fig. [167]4).

   C1R or the Complement Component 1, R subcomponent is a proteolytic
   subunit in the C1 complex^[168]62, an integral initiator of the
   classical pathway of the complement system^[169]62. The complement
   system plays a key role in the immune system^[170]19,[171]63. It’s
   involved in the inflammatory mechanism leading to the development of
   atherosclerosis^[172]64. Expression levels of complement proteins,
   including C1R, have been found to be elevated in atherosclerotic
   plaques^[173]64. Furthermore, local complement system activation can
   lead to neutrophil chemotaxis towards clot formation sites in acute
   myocardial infarction, with C1 protein detected within plasma
   clots^[174]65. Activation of C1 has been associated with remote
   ischemic conditioning in animal models of ischemic stroke^[175]66. C1R
   was also found to increase in circulating exosomes from ischemic stroke
   patients^[176]67. Our findings reveal the upregulation of C1R in both
   plasma and saliva samples from individuals at high risk for CVD
   (Fig. [177]3). Moreover, employing machine learning models, we observed
   that C1R exhibited superior predictive capabilities in saliva compared
   to plasma (Fig. [178]4), thereby positioning it as a potential
   non-invasive biomarker for CVD risk assessment.

   Another promising biomarker to predict high CVD-risk identified by the
   current study is kallikrein 5 or kallikrein-related peptidase 5 (KLK5).
   KLK5 is a member of the Kallikrein-related peptidases (KLKs) family
   which comprises highly conserved serine proteases^[179]68. KLKs, along
   with the complement system and the renin-angiotensin system (RAS)
   pathway, play crucial roles in cardiovascular disease by initiating
   vascular inflammation, leading to hypertension and subsequent clot
   formation^[180]19. KLK5 is mainly expressed in the skin, brain, breast,
   and testis^[181]69and plays a key role in skin homeostasis^[182]70.
   Additionally, KLK5 is also involved in the thrombolytic
   system^[183]71as it binds and modifies plasminogen, kininogen, and
   fibrinogen^[184]72. On the other hand, KLK5 can be inhibited by
   antiplasmin and antithrombin^[185]72. Interestingly, our study observed
   a significant decrease in KLK5 levels in the plasma of individuals at
   high CVD risk, whereas a marked increase was noted in saliva samples
   from subjects with high CVD risk compared to those at low risk
   (Fig. [186]3and Table S1). It’s important to note that KLKs proteins in
   plasma differ from tissue KLKs, exhibiting distinct enzymes and
   releasing different kinins^[187]19. This difference in enzymatic
   component, activation, and effect might explain the interesting
   variation in KLK5 level between saliva and plasma. Additionally, some
   of the proteins implicated in blood coagulation, such as plasma
   thrombin and the Integrin alpha-IIb: beta-3 complex (platelet
   receptor), demonstrated decreased levels in plasma, while fibrinogen
   exhibited a significant increase in saliva samples from the CVD-HR
   group (Table S1). This distinctive pattern mirrors the observed
   variation in KLK5 levels between plasma and saliva. Moreover, KLK5 is
   recognized as a promising early biomarker in cancer^[188]73. In a
   recent study, KLK5 was found to be associated with T2D in an African
   American population^[189]74. Here, we suggest the potential use of
   salivary KLK5 as a noninvasive and an early biomarker for predicting
   high CVD-risk.

   Another candidate for high CVD-risk markers is the Chemokine (C–C
   motif) ligand 15 (CCL15), also known as Macrophage inhibitory protein-5
   (MIP-5) or leukotactin-1 (Lkn-1)^[190]75. This pro-inflammatory
   chemokine plays a pivotal role in activating and recruiting leukocytes
   into the blood vessel wall^[191]75. In a study involving a South
   African population, CCL15 emerged as a valuable indicator of vascular
   health, demonstrating a positive association with Carotid intima media
   thickness (cIMT), an early marker of atherosclerotic changes^[192]75.
   Plasma CCL15, along with 70 proteins, was part of a protein risk-score
   that was associated with atherosclerotic cardiovascular disease
   incidence^[193]25. CCL15 was found increased in patients with
   myocardial infarction with non-obstructive coronary arteries in
   comparison to acute myocardial infarction with obstructive coronary
   arteries patients^[194]76. Besides, recent evidence from a large-scale
   study conducted in cohorts from Norway and the USA has highlighted the
   link between plasma CCL15 and heart failure incidents^[195]77. Our data
   confirm the association of increased levels of CCL15 with high CVD-risk
   in the plasma and report similar findings in the saliva of the Qatari
   subjects (Fig. [196]3).

   Additional CVD-risk biomarkers identified in our study include GFRA1,
   PLXNB2, ACP5, and PSME3. A recent investigation among the African
   American population reported associations between plasma levels of
   GFRA1 and Plexin B2 with type 2 diabetes^[197]74. GFRA1, formally known
   as Glial cell line-derived neurotrophic factor Family Receptor Alpha 1,
   has been implicated in numerous studies investigating modifiable
   lifestyle risk factors. For example, within the Framingham Heart Study,
   GFRA1 exhibited a significant association with alcohol consumption.
   Furthermore, a study involving Saudi women with gestational diabetes
   mellitus (GDM) linked the GFRA1 gene with this condition^[198]78.
   PLXNB2, also recognized as Plexin B2, is expressed in human monocytes,
   macrophages, and foam cells, and has been observed to play a role in
   monocyte binding to endothelial cells in vitro. Additionally, Plexin B2
   has been associated with heightened diabetes risk within the
   Cardiovascular Health Study population^[199]79,[200]80. ACP5, or
   tartrate-resistant acid phosphatase type 5, serves as an enzyme
   involved in bone metabolism and immune system response against
   bacteria^[201]81,[202]82. ACP5 is primarily expressed by osteoclasts,
   dendritic cells, and activated macrophages. Serum ACP5 has been
   suggested as a potential biomarker for detecting bone metastasis in
   prostate cancer patients^[203]83. In a previous study that addressed
   the effect of magnesium on cardiovascular disease blood biomarkers,
   ACP5 level was found to be affected by magnesium
   supplementation^[204]84. Moreover, serum ACP5 levels have been observed
   to rise in chronic kidney disease patients with vascular calcification
   and undergoing hemodialysis^[205]85. Our data reveal a significant
   elevation in ACP5 levels in the plasma and saliva of individuals at
   high risk for cardiovascular disease compared to those at low risk.
   PSME3, or Proteasome activator complex subunit 3 serves as a pivotal
   regulator in protein degradation by acting as a regulatory protein for
   the 20 S proteasome^[206]86,[207]87. Within the cell, PSME3
   predominantly exists as a homodimer within the nucleus. Its role in
   macrophages has been noted for its significant contribution to
   bolstering protection against bacterial infections^[208]88,[209]89.
   Previous research has indicated an elevated level of PSME3 in
   pancreatic cancer. However, our findings demonstrate a reduction in
   PSME3 expression in both plasma and saliva samples from individuals
   classified in the high-risk group for cardiovascular disease (CVD-HR)
   (see Figs. [210]2 and [211]3). Notably, PSME3 has been associated with
   obesity and insulin resistance^[212]90. Additionally, it has
   implications in cell proliferation and fosters glycolysis in pancreatic
   cancer^[213]86. After correction for treatment (Figure S3), we observed
   a consistent level in most of CVD-risk shared biomarkers with our
   previous results (Fig. [214]4 c-d), except plasma KLK5. Suggesting
   treatment didn’t influence the validity of these markers, especially in
   the saliva of CVD-HR.

   Moreover, we conducted a pathway enrichment analysis for the
   differentially expressed proteins in CVD-HR group to explore the
   pathways associated with the identified CVD-risk biomarkers
   (Fig. [215]5). Our analysis revealed extracellular matrix organization
   and disassembly as the two shared enriched pathways among the protein
   biomarkers associated with CVD risk, identified in both plasma and
   saliva samples (depicted in Fig. [216]5). This aligns with previous
   research indicating extracellular matrix organization as a primary
   mechanism associated with proteins linked to early death in heart
   failure^[217]91. In saliva, KLK5 was significantly increased in CVD-HR
   (Fig. [218]3). KLK5, with its trypsin-like activity, can digest
   components of extracellular matrix like collagen (I, II, III, and IV),
   fibronectin, and laminin^[219]72. Our data also reports a significant
   increase in plasma collagen alpha-1(VIII) chain (CO8A1) along with
   differential expression of Matrix metalloproteinase (like MMP9, MMP3,
   and MMP12) in CVD-HR group (Table S1). This might explain the
   enrichment of extracellular matrix organization, and disassembly in
   CVD-HR (Fig. [220]5).

   The pathway enrichment analysis also gave an insight into the distinct
   CVD protein signature in saliva as it is uniquely enriched for pathways
   related to antibacterial functions (Fig. [221]5b). This can be
   explained by the significantly high level of KLK5 and KLK7 proteins
   only in saliva from the CVD-risk group (Table S1). Enzymes in the
   saliva are involved in key roles, including antimicrobial
   function^[222]28. KLK5 and KLK7, by their proteolytic activity, mediate
   the antimicrobial activity of antimicrobial peptides like
   cathelicidin^[223]92. Besides, other proteins involved in defense
   against bacteria, like MRC1, and S100A7, were also significantly
   increased only in the saliva of CVD-HR group (Table S1). Interestingly,
   all the shared CVD-risk biomarkers are linked to the oxidized LDL,
   which builds up very early in atherosclerosis development. Suggesting
   the relevance of the identified CVD-risk biomarkers in reflecting CVD
   development at an early stage^[224]79,[225]93–[226]99. Looking ahead,
   longitudinal studies will be needed to follow up on these subjects and
   observe their progression to CVD.

   Our study has some limitations. First, the sample size used on the
   SOMAscan platform for CVD-risk biomarker discovery is relatively small.
   Since SOMAscan provides a relative quantification instead of an
   absolute quantification^[227]100, a complementary quantitative
   proteomic method like immunoassay to validate the biomarkers identified
   is needed. Second, applying SOMAscan technology to complex samples like
   plasma and saliva can result in nonspecific protein detection, as a
   single aptamer may bind to multiple targets. Third, the study analyzed
   samples from only Qatari subjects, limiting its generalizability to
   other population. Further validation of CVD risk biomarkers in a larger
   mutli-ethnic cohort is needed.

   In conclusion, this study marks the first attempt, to identify a
   protein signature associated with CVD-risk in saliva samples using a
   large-scale proteomic approach (SOMAscan) in the Qatari population. Our
   results unveil the presence of eight potential CVD-risk protein
   biomarkers with promising diagnostic accuracy, providing a valuable
   tool for identifying individuals at risk of CVD development.
   Consequently, both plasma proteomics and saliva present as promising
   avenues for predicting CVD risk.

Supplementary Information

   [228]Supplementary Information 1.^ (14.5KB, docx)
   [229]Supplementary Information 2.^ (1.7MB, pdf)

Acknowledgements