Authors: Shudi Pan (1.Department of Population and Public Health Sciences, Keck School of Medicine, University of Southern California, Los Angeles, CA, USA), Zhenjiang Li (1.Department of Population and Public Health Sciences, Keck School of Medicine, University of Southern California, Los Angeles, CA, USA), Juan Pablo Lewinger (1.Department of Population and Public Health Sciences, Keck School of Medicine, University of Southern California, Los Angeles, CA, USA), Jesse A. Goodrich (1.Department of Population and Public Health Sciences, Keck School of Medicine, University of Southern California, Los Angeles, CA, USA), Hongxu Wang (1.Department of Population and Public Health Sciences, Keck School of Medicine, University of Southern California, Los Angeles, CA, USA), Sarah Rock (1.Department of Population and Public Health Sciences, Keck School of Medicine, University of Southern California, Los Angeles, CA, USA), Carmen Chen (1.Department of Population and Public Health Sciences, Keck School of Medicine, University of Southern California, Los Angeles, CA, USA), Todd M. Jenkins (2.Division of Biostatistics & Epidemiology, Cincinnati Children’s Hospital Medical Center, Department of Pediatrics, University of Cincinnati College of Medicine, Cincinnati, OH, USA), Stephanie Sisley (3.USDA/ARS Children’s Nutrition Research Center, Department of Pediatrics, Baylor College of Medicine, Houston, TX, USA), Stephen Daniels (4.Department of Pediatrics, School of Medicine, University of Colorado, Aurora, CO, USA), Douglas I. Walker (5.Gangarosa Department of Environmental Health, Rollins School of Public Health, Emory University, Atlanta, GA, USA), Max T. Aung (1.Department of Population and Public Health Sciences, Keck School of Medicine, University of Southern California, Los Angeles, CA, USA), Erika Garcia (1.Department of Population and Public Health Sciences, Keck School of Medicine, University of Southern California, Los Angeles, CA, USA), Rob McConnell (1.Department of Population and Public Health Sciences, Keck School of Medicine, University of Southern California, Los Angeles, CA, USA), Sandrah P. Eckel (1.Department of Population and Public Health Sciences, Keck School of Medicine, University of Southern California, Los Angeles, CA, USA), Michele A. La Merrill (6.Department of Environmental Toxicology, University of California; Davis, CA, USA), Tanya L. Alderete (7.Department of Environmental Health and Engineering, Johns Hopkins Bloomberg School of Public Health, Baltimore, MD, USA), Zhuanghua Chen (1.Department of Population and Public Health Sciences, Keck School of Medicine, University of Southern California, Los Angeles, CA, USA), Frank D. Gilliland (1.Department of Population and Public Health Sciences, Keck School of Medicine, University of Southern California, Los Angeles, CA, USA), Thomas H. Inge (8.Department of Surgery, Northwestern University Feinberg School of Medicine and Ann & Robert H. Lurie Children’s Hospital of Chicago, Chicago, IL, USA), David V. Conti (1.Department of Population and Public Health Sciences, Keck School of Medicine, University of Southern California, Los Angeles, CA, USA), Justin R. Ryder (8.Department of Surgery, Northwestern University Feinberg School of Medicine and Ann & Robert H. Lurie Children’s Hospital of Chicago, Chicago, IL, USA), Lida Chatzi (1.Department of Population and Public Health Sciences, Keck School of Medicine, University of Southern California, Los Angeles, CA, USA)
Categories: Article, metabolomic, proteomic, biomarker, hypertension, adolescent, obesity, bariatric surgery
Source: Hypertension (Dallas, Tex. : 1979)
Authors: Shudi Pan, Zhenjiang Li, Juan Pablo Lewinger, Jesse A. Goodrich, Hongxu Wang, Sarah Rock, Carmen Chen, Todd M. Jenkins, Stephanie Sisley, Stephen Daniels, Douglas I. Walker, Max T. Aung, Erika Garcia, Rob McConnell, Sandrah P. Eckel, Michele A. La Merrill, Tanya L. Alderete, Zhuanghua Chen, Frank D. Gilliland, Thomas H. Inge, David V. Conti, Justin R. Ryder, Lida Chatzi
Emerging omics approaches, including metabolomics and proteomics, can be integrated into obesity treatment for better blood pressure management. We tested whether preoperative metabolomic and proteomic profiles predict long-term elevated blood pressure (EBP) changes better than known risk factors in adolescents undergoing bariatric surgery.
We included 108 participants from the Teen-Longitudinal Assessment of Bariatric Surgery (Teen-LABS). Plasma untargeted metabolomics (via liquid chromatography with high-resolution mass spectrometry) and Olink proteomics were assessed pre-surgery. An elastic net model with stability selection was used to identify features predictive of EBP reductions five years post-surgery. Models including identified metabolites, proteins and known-risk-factors (sex, race, clinical sites, parents’ education level, BMI, SBP and DBP at baseline) were compared to known-risk-factor-only models. For biological relevance, the consistency of association directions was examined in a multiethnic non-interventional adolescent cohort (n = 79).
Metabolite-based models and protein-based models with known-risk-factors each outperformed known-risk-factor–only models in predicting EBP reductions (p-value < 0.01). Higher levels of four metabolites (uric acid, taurocholic acid, nonadecanoic acid, and cystine) were consistently associated with reduced likelihood of EBP improvement or BP improvement across cohorts. SERPINA11, ICAM5 and TINAGL1 demonstrated consistent associations in both cohorts.
This is the first study identifying potential preoperative biomarkers for EBP reductions in adolescents following bariatric surgery. Integrating metabolomics and proteomics with known risk factors substantially improved predictive models compared to known-risk-factors-only models. Larger and more diverse cohorts are needed to confirm these findings and whether identified features are associated with other blood pressure treatments.
High blood pressure is a major modifiable risk factor for cardiovascular disease (CVD) events, the leading cause of mortality worldwide^1^, and contributes over 40% to the overall CVD burden among individuals aged 15–39^2^. Adolescents and young adults with high blood pressure face an elevated risk of early-onset CVD and higher long-term CVD mortality, regardless of their blood pressure levels in later life^3–5^. The substantial prevalence of high blood pressure among adolescents poses a critical public health challenge, with approximately 5.4% adolescents affected by hypertension and another 8.7% experiencing elevated blood pressure (EBP)^4,6,7^. Therefore, intervening during adolescence and young adulthood to manage blood pressure could significantly reduce CVD incidence later in life and ultimately alleviate the overall CVD burden in the general population.
Obesity is a key risk factor for elevated blood pressure^8,9^. Among adolescent populations, the Framingham offspring study found that 65–78% incidence of high blood pressure were attributed to overweight or obesity^10^. The National Health and Nutrition Examination Survey (NHANES) in the United States from 2017 to 2018 estimated that 40% children and adolescents aged 2–19 were with overweight or obesity^11,12^, highlighting the substantial public health burden of obesity on blood pressure. Weight-loss interventions targeting children and adolescents, including lifestyle changes on diet or physical activities, bariatric surgery, and pharmacotherapy are essential components of blood pressure management^13^. Of these, bariatric surgery and medications have demonstrated substantial weight loss and cardiometabolic benefits among adolescents, with bariatric surgery yielding an average of 25% weight reduction and long-term cardiometabolic benefits supported by a more robust body of evidence^14–18^. Despite its cardiometabolic benefits, bariatric surgery outcomes for EBP/hypertension remission exhibit considerable heterogeneity, with remission rates ranging from 28%–58% in adults and 46.7%–100% in adolescents^19–22^ across studies. Previous studies identified factors such as age, sex, race/ethnicity, hypertension duration, medication usage and weight loss as risk factors of high BP remissions^23–25^, however, their predictive performance of postoperative remissions remains understudied.
Emerging omics approaches, including metabolomics and proteomics, have the potential to inform personalized weight loss management and could enhance predictability of hypertensions or EBP remission resulting from weight loss, paving the way for personalized and effective interventions^26,27^. Untargeted metabolomics captures a comprehensive view of an individual’s physiological status and may uncover metabolites predictive of EBP reductions as well as provide insights on underlying mechanisms of blood pressure regulation after bariatric surgery^28,29^. Similar, proteomics can reveal preoperative protein expression alterations and provide complementary insights into the molecular pathways influencing postoperative blood pressure changes^30^. To date, no studies have investigated the role of metabolomics and proteomics in predicting EBP changes in adolescents following bariatric surgery.
We hypothesized that preoperative plasma metabolomic and/or proteomic signatures could predict long-term EBP reductions in adolescents following bariatric surgery using supervised machine learning models. Utilizing a multi-site cohort study of adolescents undergoing bariatric surgery, we aimed to identify early biomarkers predictive of EBP reductions five years after bariatric surgery. To establish the biological relevance of these identified features, we then examined their associations with blood pressure changes in both the surgical cohort and an independent, multiethnic, non-interventional cohort of overweight adolescents, , and identified features were also studied longitudinally in the surgical cohort.
Because of the sensitive nature of the data collected for this study, requests to access the dataset from qualified researchers trained in human subject confidentiality protocols may be sent to University of Southern California at Dr. Lida Chatzi.
This current study leveraged data from the Teen-Longitudinal Assessment of Bariatric Surgery (Teen-LABS) consortium, a prospective, multi-center, observational cohort designed to assess the effectiveness and safety of bariatric surgery in adolescents with severe obesity. Participants aged <19 years who underwent bariatric surgery between 2007 and 2012 were included^14,31,32^. Participants free of EBP prior to surgery were excluded, which led to a study sample of 108 participants in total who had pre-operative EBP (Figure S1). Ethical approval was obtained from the University of Southern California Institutional Review Board (IRB protocol HS-19-00057), and participants/guardians provided written informed assent/consent.
Associations of features identified in Teen-LABS were also evaluated in the Southern California Children’s Health Study (CHS) Meta-Air cohort, a selected sub study of young adults (ages 17–24 years) with a history of having overweight/obesity during adolescence from the CHS cohort^33,34^. A total of 172 young adults of mixed race/ethnicity aged between 17 and 23 years were recruited between 2014 and 2018. Follow-up data were collected between 2020 and 2022. The Meta-Air cohort excluded participants with any medications known to influence body composition or insulin metabolism or any major illness since birth. Ethical approval was granted by the USC Institutional Review Board (IRB protocol HS-19-00338), and informed consent/assent was obtained. The CHS Meta-Air cohort was used for examining biological relevance of identified features due to its similar age and sex distribution to the Teen-LABS study. This selection also enhances the generalizability of our findings to a diverse group of adolescents of overweight/obesity who had not undergone bariatric surgery. This analysis included 79 CHS Meta-Air cohort participants who had complete omics and blood pressure data, with baseline defined at their recruitment visit. The mean follow-up time was 5.99 years after the baseline visit.
Untargeted metabolomics were measured in fasting plasma samples using liquid chromatography with high-resolution mass spectrometry (LC-HRMS, Thermo Vanquish Duo with Q-Exactive HFX) in Teen-LABS and CHS ^35^ using established protocols^36^. To maximize metabolite detection, each sample was analyzed using four different reverse-phase C18 chromatography in ESI− mode, C18 in ESI+ mode, hydrophilic interaction chromatography (HILIC) in ESI− mode, and HILIC in ESI+ mode. Metabolite profiles were then extracted with apLCMS, incorporating modifications from xMSanalyzer^37,38^. The extracted metabolites were defined by their mass-to-charge ratio (m/z), retention time, and ion intensity. The identical analytical workflow was used for both the Teen-LABS and CHS metabolomic samples, as described in our previous study^39^. Briefly, batch correction was performed using a quality control-based random forest signal correction algorithm from the statTarget R package to eliminate unwanted variability between and within batches^40^. Features detected in fewer than 20% of cases were removed, as were those with a QC sample coefficient of variation over 30%. Missing values were imputed by assigning half of the minimum observed value of each metabolite. This resulted in 569 annotated endogenous metabolites in Teen-LABS matched to an in-house database for high-confidence annotations (Level 1)^41^. For CHS, we utilized the same in-house database. Exogenous metabolites such as pesticides were excluded in feature selection.
Proteins were measured prior to surgery in fasting plasma using the Olink Explore 384 Cardiometabolic panel and Olink Explore 384 inflammation panel using the proximity extension array (Olink Proteomics, Uppsala, Sweden)^42^. Proteins that had twenty percent or more of the total samples below the detection limit were excluded. For the remaining proteins, missing values were imputed using the K-nearest neighbor (KNN) algorithm^43^, , resulting in 682 proteins for analysis. The reported values are normalized protein expression levels following log2 transformation. The proteomics data in CHS were assessed using the Olink Explore 384 cardiometabolic panel only and was followed with the same preprocessing workflow.
At each visit, average systolic blood pressure (SBP) and diastolic blood pressure (DBP) were taken from ≥2 separate measurements obtained using a Welch Allyn Spot Vital Signs Monitor (4200B, Hillrom, Batesville, Indiana) by trained study staff^32^. Blood pressure was measured with a correctly sized cuff once the patient had been sitting upright and relaxed, feet flat on the floor, for at least five minutes, and at least thirty minutes had elapsed since they last smoked or consumed any caffeine. Anti-hypertensive medication information was also collected during interviews prior to surgery and five years after bariatric surgery on medication use form (MED) or comorbidity assessment baseline (CAB) or follow up (CAF) form. We defined EBP according to the American Academy of Pediatric guidelines for children aged 13 years and older, and according to the American Heart Association blood pressure guideline^44,45^. Reductions of EBP were defined as SBP ≤ 120, DBP ≤ 80 and no current use of anti-hypertensive medications at the fifth-year follow-up visit. For each adolescent, we categorized their blood pressure status into two participants who had EBP reductions and those whose blood pressure remained elevated five years after bariatric surgery. Thus, our main outcome was EBP reduction status five years after bariatric surgery compared to before surgery.
For CHS, SBP and DBP were measured using an automatic sphygmomanometer on the left arm in the sitting position by trained staff members. We categorized SBP and DBP into quartiles for the two visits. Participants who experienced a decrease of one quartile or more in either SBP or DBP at the follow-up visit compared to baseline were categorized as improved in BP. Participants whose blood pressure either increased or remained unchanged over time were classified as “not improved.”
The known risk factors included in model comparisons were based on previous literature^23–25^. These variables were age (years) at baseline, self-identified race (non-Hispanic white, other races, Hispanic/latinos), sex (female vs male), caregiver’s education level (college grad or higher, High school or less, some college or associate degrees and missing), body mass index (BMI) at baseline, site (Cincinnati Children’s Hospital Medical Center, Baylor college of Medicine, University of Alabama at Birmingham, University of Pittsburgh Medical Center and Nationwide children’s hospital) and SBP and DBP at baseline. Participants’ body weight and height were measured in triplicate using a calibrated stadiometer and an electronic scale (TBF-310; Tanita, Tokyo, Japan), and BMI was calculated by dividing the weight in kilograms by the squared height in meters. For CHS, demographic information was collected through questionnaires. For effect estimates of identified metabolites and proteins, the covariates used in both Teen-LABS and CHS were age at baseline, self-identified race/ethnicity (non-Hispanic white, other races and Hispanic/Latino), sex, caregiver’s education level and BMI at baseline.
The analytic workflow was illustrated in Figure 1. Descriptive statistics of demographic characteristics were stratified by EBP status change between prior to surgery and five years after bariatric surgery in Teen-LABS, and by EBP improvement status in CHS. The mean and standard deviation were calculated for continuous variables, while categorical variables were summarized using counts and percentages. All analyses were performed using R (Version 4.3.2; R Foundation for Statistical Computing, Vienna, Austria). Before downstream analysis, metabolites and proteins were log2 transformed, metabolites were also scaled for better comparisons. Statistical significance was defined as a p-value less than 0.05 for all analyses, except for the interaction analysis, where a p-value threshold of 0.2 was used.
Feature selection in Teen-LABS was performed using elastic net regression with a stability selection procedure separately for both metabolomics and proteomics. Elastic net regression with a stability selection procedure is particularly suitable for small sample size and highly correlated features^46^. We randomly sampled 85% (n ~92) of the entire study population for feature selection and conducted elastic net regression 5,000 times to ensure stability of selections. During each iteration, three-fold cross-validation was employed to tune the λ parameter in the elastic net regression. The frequency of feature selection was calculated across iterations and the top ten features with the highest selection frequencies were chosen for downstream analyses. Additionally, we used the network visualization based on the correlation matrix among identified metabolites and proteins to understand the intercorrelations among features and their biological relationships.
For proteomics and metabolomics separately, we constructed three models to compare predictive performances among 1) model with known risk factors 2) model with identified features and 3) model with both known risk factors and identified features. To minimize variability in performance estimates, repeated three-fold cross-validation was conducted 100 times. The split ratio was the 80/20 rule from the Pareto Principle^47^. Model performance was compared using area under the curve (AUC) as the primary metric.
Firth’s logistic regression was used to estimate the association between identified features and EBP reductions in Teen-LABS, adjusting for covariates within a single model. Firth’s logistic regression addresses issues of small sample size and separation in logistic regression by reducing bias in maximum likelihood estimates, providing more reliable effect estimates than standard logistic regression^48^. Associations of identified features with EBP improvement were also evaluated using Firth’s logistic regression adjusting for the same set of covariates in CHS.
Six sensitivity analyses were performed. 1) We replicated the analysis with proteomics and metabolomics combined in both feature selection and model comparisons steps. 2) In the main analysis of model comparisons, preoperative anti-hypertensive medication and preoperative weight loss were added in addition to the known risk factors to examine whether these two additional preoperative predictors enhanced the prediction performance of EBP reductions. 3) For identified features that had consistent association directions between Teen-LABS and CHS, we examined their prediction performance in Teen-LABS using elastic net logistic regression. 4) We also examined the effect estimates of identified biomarkers with systolic blood pressure percent reduction from baseline to five-year or follow-up visit (SBPbaseline−SBPfive−year/followupSBPbaseline) in both Teen-LABS and CHS. 5) We used mixed-effects logistic regression models to examine time-varying associations between repeated metabolite levels before surgery, at 6 months, 12 months and 36 months, and EBP reduction status at five years with random intercepts for participants. The interaction between metabolite intensities and visit time was to assess whether metabolite effects changed over time, adjusting for the same covariates as the main analysis. 6) In the CHS cohort, we also examined whether the effect of identified metabolites vary based on weight change status by adding an interaction term between weight changes and metabolites, adjusting for sex, race/ethnicity, age and parent’s education level.
We included 108 participants with EBP prior to surgery (baseline) in Teen-LABS and 79 participants in CHS (Table 1). Both cohorts were adolescents (Teen-LABS: Mean 17.09 years; CHS: Mean 19.96 years). Both cohorts recruited participants with overweight or obesity, and participants in Teen-LABS had an average BMI of 55 kg/m^2^ (SD: 10 kg/m^2^) whereas CHS had an average BMI of 29.8 kg/m^2^ (SD: 4.9 kg/m^2^). The race composition was different in these two cohorts. In the interventional cohort Teen-LABS, more than half of the adolescents were non-Hispanic white (69%) and the non-interventional cohort CHS consisted mostly Hispanic/Latino adolescents in Southern California. For parents’ education levels, there was a higher prevalence of college graduates among CHS participants’ parents compared to Teen-LABS participants. We compared characteristics between these two cohorts using Wilcoxon rank-sum test and Fisher’s exact test and found all demographic characteristics were significantly different in these two cohorts. (Table 1). Within the Teen-LABS cohort, participants who experienced EBP reductions were more likely to be female, older, white, lower BMI and SBP at baseline compared to those who remained with EBP for the follow-up period. Within the CHS cohort, adolescents who improved in BP at their follow-up visit were also more likely to be female and younger.
The Elastic net with stability selection procedure was leveraged to identify top features identified for EBP reductions in the interventional cohort. We identified ten features based on their selection frequency in 5,000 times of subsampling as identified features used in model comparisons in metabolomics and proteomics separately. For metabolomics (Table S1), the identified metabolites were one acyl carnitine (o-acetylcarnitine), two amino acids (carnosine, cystine), one bile acid (taurocholic acid), three fatty acids (tricosanoic acid, traumatic acid, nonadecanoic acid) and three key metabolites derivatives (succinate semialdehyde, uric acid and nicotinamide). For proteomics (Table S2), the identified proteins were carcinoembryonic antigen-related cell adhesion molecule 21 (CEACAM21), renin (REN), serine protease inhibitor, clade A, member 11 (SERPINA11), intercellular adhesion molecule 5 (ICAM5), disintegrin and metalloproteinase domain-containing protein 23 (ADAM23), bcl-2-like protein 11 (BCL2L11), tyrosine-protein kinase receptor TIE-1 (TIE1), tubulointerstitial nephritis antigen-like 1 (TINAGL1), aignaling lymphocytic activation molecule family member 1 (SLAMF1) and ecto-ADP-ribosyltransferase 3 (ART3). We also computed the relative intensities of each metabolite and protein by EBP reduction status (Table S3 and Table S4).
The network visualization of the 20 identified features in the Teen-LABS cohort reveals distinct correlation patterns (Figure 2). O-acetylcarnitine, uric acid, and cystine formed a prominent cluster with strong positive correlations. The correlation between O-acetylcarnine and uric acid was 0.47, between uric acid and cystine was 0.35, and between cystine and O-acetylcarnine was 0.53. We did not find any distinct patterns for identified proteins (Figure S2).
For metabolite-based models with repeated three-fold cross validation, the mean AUC of models with known risk factors was 0.52 (IQR: 0.08), the mean AUC of models with the ten identified metabolites was 0.80 (IQR: 0.10), and the mean AUC of models with the identified metabolites and known risk factors was 0.76 (IQR: 0.13 Figure 3). From the DeLong’s test comparing ROC curves of models with known risk factors only and models with metabolites and known risk factors, the p-value was smaller than 0.01, indicating that the AUC of metabolites and known risk factors models were statistically significantly higher than the AUC of known risk factors only models. We observed similar differences between protein-based models and models with known risk factors (Figure 4). With repeated cross validation, the mean AUC of identified-protein-based models was 0.71 (IQR: 0.18) and the mean AUC of adding known risk factors in the protein-based models was 0.66 (IQR: 0.28). We also found that the models with identified proteins and known risk factors significantly outperformed the known-risk-factor-only models (p-value < 0.01).
In the CHS cohort, five identified metabolites and three identified proteins were found to have consistent associations with those in Teen-LABS, including cystine, nonadecanoic acid, taurocholic acid, tricosanoic acid, uric acid, SERPINA11, ICAM5 and TINAGL1. Among these eight features, four metabolites and two proteins had non-statistically significant negative associations with BP improvement. Only tricosanoic acid and ICAM5 were trending positive with BP improvement in CHS (Figure 3 and Figure 4). In Teen-LABS, cystine exhibited the strongest association, with an odds ratio (OR) of 0.51 (95% CI: 0.31, 0.81). The odds of EBP reductions had a 44% decrease with 1 SD increased in log2- nonadecanoic acid intensity (OR: 0.56, 95% CI: 0.34, 0.90) and taurocholic acid (OR: 0.65, 95% CI: 0.40, 0.96). Uric acid (OR: 0.59, 95% CI: 0.33, 1.01) also demonstrated similar directions. None of these five metabolites were statistically significant associated with blood pressure improvement in the non-interventional cohort CHS. For proteomics, SERPINA11, ICAM5 and TINAGL1 had consistent association direction in the both cohorts (Figure 4). SERPINA11 was significantly associated with EBP reduction in teen-LABS (OR: 0.34, 95% CI: 0.14, 0.79) and trended negative with BP improvement in CHS (OR: 0.91, 95%CI: 0.40, 2.07). ICAM5 was positively associated with EBP reduction in Teen-LABS (OR: 2.51, 95% CI: 1.05–6.93), and trended toward higher odds of BP improvement in CHS (OR: 1.20, 95% CI: 0.48–3.14). TINAGL1 showed a significant negative association in Teen-LABS (OR: 0.18, 95% CI: 0.03–0.83) and was trended toward lower odds of BP improvement in CHS (OR: 0.69, 95% CI: 0.14–3.06). In addition, we found REN (OR: 0.61, 95% CI: 0.36–0.95) and TIE1 (OR: 0.33, 95%CI: 0.11–0.82) were significantly negatively associated with EBP reductions in Teen-LABS.
In sensitivity analyses, 1) The elastic net regression with stability procedure of both metabolites and proteins features identified five proteins and five SERPINA11, CEACAM21, REN, tricosanoic acid, succinate semialdehyde, nicotinamide, cystine, ICAM5, TIE1 and O-acetylcarnitine, which were also identified in feature selections conducted separately in metabolomics and proteomics. Models with combined omics did not improve the prediction performance (AUC: 0.78, IQR: 0.11) compared to models with metabolomics or proteomics only (Figure S3) 2) The model with additional known risk factors slightly improved the prediction performance (AUC: 0.59, IQR: 0.14). After adding the two preoperative known risk factors, the models with identified metabolites (AUC: 0.76, IQR: 0.13) or proteins (AUC: 0.71, IQR: 0.13) with risk factors remained the same (Figure S4). 3) For features shown consistent association directions, models with the seven features (six metabolites and one protein) combined with risk factors (AUC: 0.70, IQR: 0.14) outperformed the risk-factor-only models. 4) We observed consistent directions of effect for cystine and tricosanoic acid on percent SBP reduction in both the Teen-LABS and CHS cohorts (Figure S3). Although higher baseline tricosanoic acid was associated with greater SBP reduction in Teen-LABS (β = 0.01; 95% CI −0.02 to 0.04) and CHS (β = 0.004; 95% CI −0.013 to 0.02), these estimates were not statistically significant. By contrast, elevated baseline cystine was significantly linked to smaller SBP reductions in Teen-LABS (β = −0.03; 95% CI −0.05 to −0.004) and CHS (β = −0.04; 95% CI −0.08 to −0.006). 5) We then assessed whether the same association direction held for repeated measured metabolites at 6 months, 12 months and 36 months post-surgery with the metabolite×visit interaction. Likelihood ratio tests for all interactions were non-significant (p-value > 0.05), indicating that the effects of these identified metabolites did not change over the three year follow-up period. Uric acid, taurocholic acid and cystine all demonstrated to have negative associations with EBP reductions across follow-up visits (Table S5). In the non-interventional CHS cohort, nicotinamide shows a nominally negative interaction with weight change (OR=0.85, P=0.19 < 0.2; Table S6). This may explain the negative association of nicotinamide with BP reduction in Teen-LABS, but the opposite association with BP improvement in CHS.
To our knowledge, this is the first study to examine preoperative prognostic biomarkers for predicting long-term EBP reductions among adolescents undergoing bariatric surgery. By leveraging untargeted metabolomics and Olink proteomics data through feature selection analyses in adolescents undergoing bariatric surgery, we identified predictive models whose identified features with known risk factors consistently outperformed those based on only previously known risk factors. For biological relevance, effect estimates were evaluated in both Teen-LABS and CHS, an independent non-interventional cohort. Four metabolites (uric acid, taurocholic acid, nonadecanoic acid and cystine) demonstrated consistent association directions across both cohorts. For uric acid, taurocholic acid, nonadecanoic acid and cystine, higher levels of them were associated with lower likelihood of EBP reductions five years following bariatric surgery, indicating their potentially harmful effects on blood pressure regulation. Tricosanoic acid showed a positive association with reductions in EBP, though the association was not statistically significant. Two proteins SERPINA11 and TINAGL1 showed consistent negative association in both cohorts, and ICAM5 showed consistent positive associations. Higher levels of SERPINA11 and TINAGL1 were significantly negatively associated with EBP reductions in Teen-LABS. Among metabolites that did not show consistent associations across two cohorts, o-aceylcarnitine was significantly negatively associated with EBP reductions, while carnosine showed a significant positive association in Teen-LABS. For proteomics, although not consistent across cohorts, REN/Renin, the hormone which increases blood pressure) were significantly negatively associated with EBP reductions in the Teen-LABS study.
We found in the network analysis that cystine, uric acid and o-acetylcarnitine were closely correlated with each other, and both cystine and uric acid had consistent negative association directions in the surgical and non-interventional observational cohort, and in the surgical cohort over time. In the sensitivity analysis, cystine had statistically significant associations with SBP percent reduction across cohorts. Previous animal and epidemiological studies have shown extensive evidence for a causal link between elevated uric acid levels and EBP or primary hypertension in adolescents^49,50^. Uric acid is the end product of the purine metabolism, which is involved in breaking down acids, alcohol, fructose and DNA/RNA. The proposed mechanism involves activation of the rein-angiotensin system, reactive oxygen specific, thromboxane and cyclooxygenase (COX) generation^49,51^. Cystine, an oxidized dimeric form of cysteine, has been established as a marker of extracellular oxidative stress. Increased cystine levels were cross-sectionally associated with CVD-related risk factors such as arterial stiffness^52^, wall thickness^53^ and pulmonary artery systolic pressure in humans^54^. Elevated uric acid and cystine levels were both associated with lower likelihood of having EBP reductions or BP improvement in the Teen-LABS and CHS cohorts, which is consistent with the association directions reported in previous literature. In fact, a clinical trial in adolescents with elevated blood pressure showed that lowering uric acid can reverse elevated blood pressure^55^. O-aceylcarnitine is a saturated fatty acyl-l-carnitine and is functionally related to carnitine. Although o-aceylcarnitine did not show consistent associations in both cohorts, higher levels of plasma aceylcarnitines were associated with higher blood pressure in adult population^56–60^, as well as children and adolescents^61^. In a large Swedish population-based cohorts, metabolite-based predicted aortic stiffness, included both acylcarnitine and cystine was significantly associated with CVD incidence and mortality after 23 years follow up^62^. Therefore, we hypothesize that these three closely related metabolites (cystine, uric acid and o-aceylcarnitine) are all involved in oxidative stress and fatty acid degradation to regulate blood pressure levels among adolescents and may participate in irreversible vascular or renal injury resulting in persistent, surgical treatment hypertension. Supporting this hypothesis, in our sensitivity analysis, postoperative uric acid and cystine levels remained, although not significantly, negatively associated with EBP reductions from the longitudinal analysis. Since we found effects of uric acid and cystine did not change across three-year follow-up visits, targeting these biomarkers, with for instance uric acid lowering therapy, may still be beneficial for blood pressure management even after surgery. Tricosanoic acid and nonadecanoic acid were two novel identified metabolites, which showed consistent association directions. These two metabolites are long-chain fatty acids involved in fatty acid metabolism^63^. To our knowledge, no previous studies have identified them as predictors of EBP reductions or blood pressure levels.
Taurocholic acid, a taurine-conjugated bile acid, was identified in our study as inversely associated with remission of EBP, suggesting that higher circulating levels may be linked to persistent hypertension following bariatric surgery in adolescents. This finding contrasts with recent experimental models that suggest a potential antihypertensive role for taurocholic acid. In a rat model of hypertension, taurocholic acid supplementation restored bile acid conjugation and reduced systemic blood pressure, potentially via metabolic reprogramming through the gut-liver axis^64,65^. One possible explanation for this discrepancy lies in the context-specific roles of bile acids, where systemic levels may reflect compensatory or dysregulated phenomena in metabolic disease rather than direct causal effects. More studies are needed to understand the effects of taurocholic acid on long-term blood pressure changes.
Elevated protein levels in SERPINA11, REN, TIE1 and TINAGL1 were significantly associated with a lower likelihood of EBP reductions five years following bariatric surgery in the Teen-LABS cohort. While SERPINA11 itself has not been directly linked to hypertension, its location on chromosome 14q32.13, a region studied for blood pressure associations, might suggest its indirect involvement^66^. The proximity to other serpin genes implicated in vascular processes suggests that SERPINA11 could play a role in extracellular matrix remodeling or inflammatory responses, both of which are relevant to hypertension pathophysiology. REN is a key enzyme in the renin–angiotensin–aldosterone system (RAAS), converting angiotensinogen into angiotensin and thereby promoting vasoconstriction and fluid retention^67^, and this finding underscores the possibility that in certain individuals, pathophysiologic changes may be occurring preoperatively, rendering surgical weight loss therapy less effective for reversal of elevated blood pressure. TIE1 plays an important role in angiogenesis and blood vessel stability and may regulate blood pressure via RAAS^68^. We also found that ICAM5 had a significant positive association in Teen-LABS. ICAM5 is a neuronal type I transmembrane glycoprotein that is involved in neuron-neuron mediation^69^, and TINAGL1 is a secreted lipocalin that regulates lipid metabolism, inflammation, oxidative stress and angiogenesis^70^. The lipocalin family has accumulating evidence showing its associations with CVD. However, as far as we know, no prior research has identified these proteins as predictors of EBP reduction or blood pressure levels.
Although we have interpreted the preoperative elevation of these metabolites that might have harmful effects on EBP, it is also possible that their elevated levels might reflect compensatory protective mechanisms in response to increased vascular stress among patients who had preoperative EBP. Bariatric surgery may therefore reduce not only pathogenic but also beneficial metabolites. For example, we observed significant negative association of o-acetylcarnitine, REN and TIE1 with EBP reductions in Teen-LABS; However, these findings were not replicated in the CHS cohort. This might suggest that these elevated omics features levels represent a compensatory mechanism aimed at countering blood pressure dysregulation prior to surgery. Supporting this hypothesis, both L-carnitine and its ester acetyl-L-carnitine administrated either intravenously or orally have been shown to ameliorate arterial hypertension in randomized controlled trials. A recent meta-analysis of 22 randomized control trials found that L-carnitine supplementation lowers DBP levels in overweight participants^71^. Although no clinical trials were found targeted REN and TIE1, these proteins might also serve compensatory roles among patients with EBP, potentially explaining the inconsistent association directions we observed in a non-interventional cohort. Nicotinamide, which was not statistically significant in either the Teen-LABS or CHS cohorts and showed trends in opposite directions, reached statistical significance for the effect modification analysis between weight changes and metabolite intensities. This finding suggests that nicotinamide may help identify individuals whose blood pressure is more likely to improve if they lose weight. However, due to limited statistical power in our current analysis, the specific role of nicotinamide in the interaction between weight loss and blood pressure improvement needs further investigation in a larger cohort.
This study has several strengths. First, it is the first study to show that metabolite- or protein-based models can prospectively predict long-term EBP reductions after obesity interventions such as bariatric surgery. Second, although our interventional cohort was predominantly non-Hispanic white, we evaluated these findings in an independent, multiethnic, non-interventional adolescent cohort with a similar age and sex composition, suggesting broader generalizability in the etiological relevance of these biomarkers. Furthermore, we used only confirmed metabolites with confidence level 1 from untargeted metabolomics, increasing the reliability of our annotations, and metabolomics of these two cohorts were measured at the same facility, reducing the potential misclassification bias and batch effects. The study also has limitations. Our sample size was small, which may explain the inconsistent directions of association and limited power for detecting significant effects. However, we applied a rigorous analytical workflow to minimize the potentials of overfitting, and our identified features are biological plausible. Additionally, the feature selection frequency using elastic net regression was low for both metabolite- and protein-based models, possibly reflecting weak underlying associations or strong intercorrelations among the metabolites. Third, metabolites can vary substantially over time; only a single preoperative measurement was assessed, limiting our ability to assess temporal changes prior to the surgery. However, we conducted a sensitivity analysis on the associations between metabolites measured post-surgery, which might suggest the stability of identified metabolites levels over time. Fourth, blood pressure in the two cohorts was also measured using different devices, which could potentially introduce measurement bias in EBP definition. However, we conducted a sensitivity analysis using SBP percent reduction as an alternative outcome and found similar association directions for identified metabolites. Finally, although we examined the biological relevance in the independent CHS cohort for generalizability, CHS is neither a weight-loss intervention cohort nor focuses on adolescent EBP. Therefore, the biological relevance of these identified biomarkers needs to be interpreted with caution and replication in larger, adolescent-specific intervention studies is needed to confirm the robustness of effect estimates. To address this limitation, we conducted a sensitivity analysis in the CHS cohort to evaluate potential interactions between metabolites and weight changes on blood pressure improvement, and found no significant interactions, which suggests the effects were not conditioned on weight changes in the CHS cohort.
In conclusion, we found that omics-based models outperformed known-risk-factor-only models for predicting long-term EBP reductions prior to bariatric surgery, and the identified preoperative metabolites might serve as prognostic biomarkers for tailored blood pressure management prior to obesity treatment. Future research should replicate these findings with larger sample sizes and investigate their predictive power in broader populations.
Our study is the first to leverage preoperative metabolomic and proteomic profiling to predict long-term blood pressure reductions in adolescents undergoing bariatric surgery. These findings highlight the potential of integrating high-dimensional omics data to enhance risk stratification of blood pressure management. Identified biomarkers may serve as actionable biomarkers for tailored preoperative interventions of blood pressure among adolescents with obesity treatment. Future studies should validate these biomarkers in larger and more diverse populations.
Tables S1–S6,
Figures S1–S5.