INTRODUCTION

Bariatric surgery results in significant, sustained weight loss in most patients. The surgery also improves most cardiovascular risk factors, including blood pressure, diabetes, dyslipidemia, fatty liver disease, inflammation, and health-related quality of life ((1-4)). Total mortality and cardiovascular mortality are reduced by 40% and 56%, respectively ((5)). Whereas voluntary, nonsurgical weight loss is difficult to maintain over the long term ((6)), surgical weight loss is generally maintained long term and, as such, it may provide an efficient way of identifying proteins and protein pathways contributing to improved health outcomes induced by bariatric surgery or weight loss in general.

Short-term studies of 1 or 2 years (i.e., the nadir of maximum weight loss) ((7)) have suggested plasma proteins significantly change after bariatric surgery ((8-11)), but it is unknown which protein changes are durable in the long term. Long-term assessment is important because short-term postsurgical changes in many proteins may be due to the invasive nature of the surgery itself and the dramatic first year of postsurgical weight loss, which is expected to affect many related physiological systems. However, once weight has stabilized and modest weight regain progresses, the underlying short-term benefits of the protein changes may disappear if these proteins return to near presurgical levels. Identification of proteins that remain changed long term may help identify biological pathways responsible for disease reduction. Therefore, a proteome-wide association study (PWAS) was conducted to test for significant differences in 12-year changes in 1,297 plasma protein levels between 234 individuals with severe obesity who had gastric bypass surgery and a nonintervened group of 144 individuals with severe obesity. Protein changes were also related to significant changes in disease risk factors.

METHODS

A longitudinal, controlled study of the risks and benefits of Roux-en-Y gastric bypass surgery was begun in 2000. After a baseline exam (exam 1), additional exams were conducted at the University of Utah at approximately 2 years (exam 2), 6 years (exam 3), and 12 years (exam 4) ((3, 12, 13)). Follow-up of major clinical variables was more than 90% at exam 4, although only 67% of individuals had their blood drawn at the University of Utah and could provide a sample for proteomic measurements. Of the 67% returning individuals, a subset of the individuals was selected who either had gastric bypass surgery after the baseline exam or who never had bariatric surgery during follow-up. The nonsurgery group did not have a study-related intervention during the follow-up period, and even though mean weight did not change, there was a wide variation in weight change in this group over the 12 years. An initial subset of 203 individuals (137 in the surgery group and 66 in the nonsurgery group) was used as a discovery set, and a second subset of 175 individuals (97 in the surgery group and 78 in the nonsurgery group) was used for replication. In addition to plasma protein measurements at baseline and 12 years, proteins were measured at the 2-year follow-up (exam 2) for 204 of the 234 individuals who had gastric bypass surgery. Proteins from exam 3 were not measured. All individuals provided written informed consent, and the study was approved by the University of Utah Institutional Review Board.

Clinical measurements

The clinical measurements shown in Table 1 were measured or calculated as previously described ((3, 12-14)). Fat-free mass and fat mass (FM) were measured by bioimpedance. Resting energy expenditure (REE) was measured after an overnight fast by indirect calorimetry (TrueOne 2400, ParvoMedics, Sandy, Utah) ((12)). Clinical biochemistries were obtained after an overnight fast. Diabetes remission at 12 years in individuals with baseline diabetes, diabetes incidence in individuals without baseline diabetes, and 10-year risk of coronary heart disease (CHD) as assessed by the Framingham Risk Score (FRS) were derived. To ensure that the FRS represented change in CHD risk only due to changes in the clinical variables and not age, age at exam 4 was used in both the baseline and 12-year follow-up risk equations. Diabetes was defined as a fasting glucose of 126 mg/dL or greater or being on antidiabetic medication. Blood pressure, lipids, and diabetes-related variables that are affected by antihypertensive, antidiabetic, or lipid medications were adjusted prior to analysis, as described previously ((3)). Briefly, individuals taking medications for each condition had their values changed to the sex-specific means of untreated individuals with the condition. Analysis with and without medication adjustment was performed.

TABLE 1. Discovery and replication participant subset characteristics and winsorized variable means and standard deviations (SD) Variable Discovery Replication Groups Surgery (n = 137) Nonsurgery (n = 66) All (n = 203) Surgery (n = 97) Nonsurgery (n = 78) All (n = 175) Gender (M) 20% 12% 17% 12% 11% 12% Age (Exam 1, y) 46.4 ± 10.5 47.1 ± 11 46.6 ± 10.9 40.1 ± 10 43.9 ± 12.3 41.8 ± 11.2 BMI (exam 1, kg/m2) 46.4 ± 7.1 44.6 ± 6.4 45.8 ± 6.9 46.1 ± 6.4 45.3 ± 7.5 45.8 ± 6.9 BMI (exam 2, kg/m2; n = 204) 29.7 ± 5.5 (n = 137) NA NA 34.3 ± 9.1 (n = 67) NA NA BMI (exam 4, kg/m2) 34.6 ± 8.0 44.8 ± 7.5 37.9 ± 9.1 35.0 ± 8.8 45.0 ± 8.9 39.6 ± 10.2 Diabetes (exam 1, %) 34 38 35 8 32 19 Diabetes (exam 4, %) 17 38 24 7 54 28 Changes in clinical variables (exam 4-exam 1) FFM (kg) M ± SD −13.4 ± 6.7 −2.4 ± 6.5 −9.9 ± 8.4 −13.5 ± 7.0 −4.9 ± 7.4 −9.8 ± 8.4 n 127 59 186 91 70 161 FM (kg) M ± SD −22.2 ± 11.5 −2.1 ± 10.3 −15.8 ± 14.5 −21.0 ± 11.6 −1.2 ± 10.1 −12.4 ± 14.7 n 127 59 186 91 70 161 BMI (kg/m2) M ± SD −11.5 ± 5.4 −0.2 ± 5.6 −7.8 ± 7.6 −11.0 ± 5.5 −0.2 ± 5.3 −6.2 ± 7.6 n 137 66 203 97 78 175 REE (kcal) M ± SD −550 ± 238 −301 ± 215 −478 ± 257 −569 ± 260 −325 ± 249 −452 ± 282 n 79 32 111 50 46 96 AST (U/l) M ± SD −0.5 ± 8.1 −2.4 ± 8.5 −1.1 ± 8.3 −0.5 ± 10.0 −4.0 ± 7.8 −2.1 ± 9.2 n 136 66 202 96 77 173 ALT (U/l) M ± SD −5.0 ± 11.3 −2.7 ± 13.1 −4.3 ± 11.9 −4.4 ± 15.0 −5.4 ± 12.8 −4.8 ± 14.0 n 137 66 203 97 77 174 Uric acid (mg/dL) M ± SD −0.7 ± 1.2 0.1 ± 1.3 −0.4 ± 1.3 −0.8 ± 1.3 −0.2 ± 1.2 −0.6 ± 1.3 n 137 65 202 96 77 173 SBP (mmHg) M ± SD −4.6 ± 19.1 9.3 ± 21.5 −0.1 ± 20.9 −2.0 ± 21.5 8.8 ± 20.3 2.8 ± 21.6 n 137 66 203 97 78 175 DBP (mmHg) M ± SD −1.4 ± 13.9 5.5 ± 14.0 0.8 ± 14.3 1.1 ± 14.9 7.0 ± 14.4 3.7 ± 14.9 n 137 66 203 97 78 175 Glucose (mg/dL) M ± SD −13.3 ± 20.7 −2.5 ± 23.9 −9.8 ± 22.3 −7.5 ± 14.0 −1.9 ± 25.9 −5.0 ± 20.3 n 137 66 203 97 78 175 Insulin (μU/mL) M ± SD −10.8 ± 11.8 −6.7 ± 13.9 −9.5 ± 12.6 −11.3 ± 13.0 −6.2 ± 12.8 −9.1 ± 13.1 n 137 66 203 97 78 175 HOMA-IR M ± SD −3.1 ± 3.4 −2.0 ± 4.1 −2.8 ± 3.7 −2.9 ± 3.5 −1.5 ± 3.9 −2.3 ± 3.7 n 137 66 203 97 78 175 HOMA-B M ± SD −71 ± 147 −69 ± 177 −70 ± 157 −100 ± 179 −38 ± 166 −73 ± 175 n 133 63 196 96 76 172 HbA1c (%) M ± SD 0.01 ± 0.93 0.25 ± 1.00 0.09 ± 0.96 0.07 ± 0.76 0.44 ± 1.22 0.23 ± 1.01 n 136 66 202 97 78 175 TG (mg/dL) M ± SD −70.6 ± 59.8 −24.8 ± 56.8 −55.7 ± 62.5 −73.0 ± 67.3 −33.9 ± 69.7 −55.6 ± 70.9 n 137 66 203 97 78 175 LDL-C (mg/dL) M ± SD −2.9 ± 29.5 26.4 ± 29.7 6.6 ± 32.5 −3.4 ± 30.6 27.8 ± 28.4 10.5 ± 33.4 n 137 66 203 97 78 175 HDL-C (mg/dL) M ± SD 13.7 ± 12.6 2.2 ± 9.4 10.0 ± 12.8 15.6 ± 13.8 2.3 ± 9.0 9.7 ± 13.6 n 137 66 203 97 78 175 FRS (%) M ± SD −2.3 ± 3 −0.8 ± 2 −1.8 ± 3 −1.0 ± 2 −0.8 ± 2 −0.9 ± 2 n 137 66 203 97 78 175 DMINC % 0 0 0 3 21 24 n 91 41 132 88 51 139 DMREM % 23 0 23 4 5 9 n 46 25 71 8 25 33 Exam 1: baseline exam; exam 2: exam at 2 years; exam 4: exam at 12 years. Only 204 of the 234 participants who had gastric bypass surgery had available exam 2 proteomic measurements. Abbreviations: ALT, alanine aminotransferase; AST, aspartate aminotransferase; DBP, diastolic blood pressure; DMINC, diabetes incidence at 12 years; DMREM, diabetes remission at 12 years; FFM, fat-free mass; FM, fat mass; FRS, Framingham Risk Score for 10-year cardiovascular disease incidence; HbA1c, hemoglobin A1c; HDL-C, high-density lipoprotein cholesterol; HOMA-B, homeostatic assessment of insulin secretion; HOMA-IR, homeostatic assessment of insulin resistance; LDL-C, low-density lipoprotein cholesterol; NA, not measured at exam 2; REE, resting energy expenditure; SBP, systolic blood pressure; TG, triglycerides.

Plasma aliquots were stored at −80°C. Prior to assay, the plasma was thawed, and approximately 70 μL was used for the proteomics assay. Proteins were measured by a Slow Off-rate Modified Aptamer (SOMAmer)-based protein array (SomaLogic, Boulder, Colorado) ((15)). A total of 1,297 proteins were measured and are referred to by their gene names. Normalization and calibration of protein levels were done using SomaScan proprietary software (SomaLogic). Sample data were normalized to remove hybridization variation within a run, followed by median normalization across all samples to remove other assay biases within the run and, finally, calibration to remove assay differences between runs. Calibrator coefficients of variation on each plate were calculated; at least 50% of SOMAmer reagents had to have coefficients of variation less than 0.1, and 95% had to have coefficients of variation below 0.2. Any flagged samples by SomaScan software were removed from the analysis. Protein abundance was calculated as log of the relative fluorescence units.

Statistical methods

Outliers were removed at four standard deviations above or below the mean value of each protein. Within-individual changes in protein levels (protein at exam 4 − protein at exam 1) were calculated for all 1,297 proteins. Clinical data were winsorized at 5% to 95% quantiles.

β coefficients from linear regression models were used to test for protein changes versus surgery status (surgery group compared with nonsurgery group) or versus BMI changes (exam 4 − exam 1) after adjusting for age, sex, and baseline BMI, first in the discovery batch and then in the replication batch. Additional regression models were run after adding change in BMI as a covariate when comparing the surgery and nonsurgery groups. All statistical tests were two-sided tests and were adjusted for multiple comparisons using the Bonferroni correction.

A functional enrichment analysis was performed in R using the fgsea (Korotkevich G et al, bioRxiv doi:10.1101/060012) and clusterProfiler ((16)) packages (R Foundation, Vienna, Austria) for gene set enrichment analysis (GSEA). The analysis was performed using pathway annotation from human Reactome (version 73; June 2020 release), WikiPathways (July 2020 release), and Kyoto Encyclopedia of Genes and Genomes (KEGG; July 7, 2020, release) databases. The full SomaScan panel of proteins was used as the background set of proteins. GSEA results were filtered for redundant entries based on semantic similarity between protein groups. UniProt keyword (October 15, 2019, release) annotations were used for classification of proteins. Significantly overrepresented diseases were assessed using the Genetic Association Database ((17)). Clustering in heat maps was performed using the complete linkage Euclidean distance method.

RESULTS

Baseline and 12-year change variables are presented for the discovery and replication subsets in Table 1. For the gastric bypass surgery group, unadjusted protein means at baseline, 2 years, and 12 years for all 1,297 proteins are color coded for those that increased or decreased over 12 years (Supporting Information Table S1). The unadjusted protein means at baseline and at the 12-year follow-up in the comparison nonsurgery group are listed in Supporting Information Table S2.

Long-term effect of surgery and weight loss on the proteomic profile

A total of 58 adjusted protein changes differed between the gastric bypass surgery and nonsurgery groups in the discovery set at a Bonferroni significance level (p < 0.05/1,297 = 3.85 × 10−5), among which 51 were replicated (p < 0.05/58 = 8.6 × 10−4; Figure 1A, Supporting Information Tables S3-S5 for discovery, replication, and combined analysis, respectively).

(A) Plot of 12-year protein changes’ normalized β values (β/SE β) vs. nominal −log10 p values. Left plot from the surgery vs. nonsurgery analysis of the combined discovery and replication samples. Right plot from the Δ BMI (exam 4 − exam 1) analysis. Red color identifies the 78 replicated and Bonferroni-corrected significant protein changes. (B) Overlap of 71 significant 12-year protein changes associated with BMI change (blue oval) and 51 associated with gastric bypass surgical status without correcting for change in BMI in the model (red oval). Twelve of the fifty-one proteins were still significantly associated with surgery status after correcting for changes in BMI (purple oval). Four proteins were no longer associated with surgery status after adjusting for BMI changes even though they were not found to be associated with BMI changes

Because the 12-year change in weight varied within both the surgery and nonsurgery groups, additional PWAS was conducted to find the association of 12-year changes in proteins with the quantitative change in BMI, adjusting for age, sex, and baseline BMI. In the discovery set, 12-year changes in levels of 99 proteins were associated with changes in BMI (p < 3.85 × 10−5). Of these 99 protein changes, 71 were replicated (p < 0.05/99 = 5.1 × 10−4; Figure 1A). Sixty of the seventy-one identified proteins increased as BMI decreased over the 12 years, including insulinlike growth factor binding protein 2 (IGFBP2); apolipoprotein M (APOM); and adiponectin, C1Q and collagen domain containing (ADIPOQ); whereas 11 proteins decreased, including leptin (LEP), C-reactive protein (CRP), growth hormone receptor (GHR), afamin (AFM), and myeloperoxidase (MPO; Figure 1A).

Combining the significant results from the two PWAS analyses (Figure 1B), 78 long-term protein changes were associated with change in BMI (n = 71) or between surgery and nonsurgery groups (n = 51). All but seven protein changes that were significantly different between surgery groups were also associated with change in BMI (Figure 1B). Therefore, the protein change differences between groups were further adjusted for change in BMI to see whether some proteins changed independently of BMI change. Twelve of the fifty-one replicated protein changes that were significantly associated with surgery/nonsurgery status remained significantly associated with surgery status even after further adjustment for BMI change (Figure 1B).

Short- versus long-term effect of surgery and weight loss on the proteomic profile

Short-term protein change associations with BMI change were maintained in the long term (Figure 2A, Supporting Information Tables S6-S7). There was a strong correlation of the 12-year follow-up versus the 2-year follow-up standardized β coefficients (β/SE) associated with the p values shown in Figure 2A for the 71 proteins (r = 0.93).

(A) Comparison of −log10 p values from the association of short-term (exam 2 – exam 1) vs. long-term (exam 4 − exam 1) protein change associations with BMI change for all 1,297 proteins in the combined discovery and replication groups. Red dots indicate the 71 replicated proteins for BMI change. (B) Patterns of protein changes per unit of BMI change after substantial weight loss (exam 2 − exam 1) and after some weight regain (exam 4 − exam 2) obtained from the surgery group (N = 204 who have measurements at all three time points) for the 78 proteins in Figure 1. Ratio of β coefficients derived from the regression of the protein change on BMI change for the 2-year follow-up (exam 2 − exam 1; β1-2) divided by the β coefficient using changes from the subsequent 10 years of follow-up (exam 4 − exam 2, β2-4). Ratios greater than one suggest greater protein change per unit of BMI change during the weight-loss period compared with the weight-gain period. Ratios less than one suggest greater protein changes per unit of BMI change in the weight-gain period compared with the weight-loss period. Proteins shown in red increased with increasing BMI and proteins in black decreased with increasing BMI. Blue lines represent ±50% increase or decrease in the ratio. FCN2, with a ratio of 5.0, is not shown to improve the readability of the figure. Serpin family A member 4 (SERPINA4) and cadherin 1 (CDH1) had greater changes during weight loss but also had negative ratios of −7.0 and −4.8, respectively, indicating that the protein change association with BMI change did not reverse from the weight-loss to the weight-gain period, as did the other proteins (points not shown). The β coefficients used in the ratios can be found in Supporting Information Table S6

All protein changes except regenerating family member 4 (REG4) had the same direction of association with BMI change during weight-loss (exam 2 - exam 1) and weight-regain (exam 4 - exam 2) time periods (Supporting Information Table S6). If the protein change was positively associated with BMI change from exam 1 to 2 (both decreased over time), it was positively associated from exam 2 to 4 (both increased). If the protein change was inversely associated from exam 1 to 2 (BMI decreased but the protein increased), the protein change was inversely associated with BMI change from exam 2 to 4 (BMI increased but the protein decreased). Therefore, protein changes consistently reflected both BMI decreases and BMI increases.

In addition to the consistent direction of the association of protein changes with BMI change, the amount of protein change per unit of BMI change was assessed to see whether the period of rapid weight loss had a differing effect on a protein than the period of slow weight regain. All but 16 proteins had a protein change per unit of BMI change during weight regain that was within 50% (0.5 ≤ ratio of β coefficients ≤ 1.5) of the change per unit of BMI during weight loss (Figure 2B, Supporting Information Table S6). Proteins that changed more per unit of BMI change during weight loss than weight regain included leptin receptor (LEPR), insulinlike growth factor binding protein 1 (IGFBP1), APOM, and three serpin proteins. Only four proteins had greater than a 50% change per unit of BMI change during the weight-regain period than the weight-loss period.

Individuals who regained less than 10% of their baseline presurgical weight from exam 2 to exam 4 were compared with individuals who regained more than 10% of their baseline weight ((18)). Six proteins were significantly associated with weight regain versus weight maintenance (p < 0.05/1,297; Supporting Information Figure S1, Supporting Information Table S8). LEP and amyloid P component, serum (APCS) were higher at exam 4 in those who regained more than 10% of their baseline weight, whereas IGFBP2; WAP, follistatin/kazal, immunoglobulin, kunitz and netrin domain containing 2 (WFIKKN2); HtrA serine peptidase 2 (HTRA2); and sex hormone binding globulin (SHBG) were lower. The clinical characteristics of the two groups are presented in Supporting Information Table S9.

Clinical variable associations with the proteome

A heat map of the regression β-coefficients and the cluster patterns of the 71 protein and 20 clinical variable 12-year changes showed the strongest protein associations with lipids (58 proteins with high-density lipoprotein cholesterol [HDL-C], 47 with low-density lipoprotein cholesterol [LDL-C], and 44 with triglycerides [TG]), REE (36 proteins), and uric acid (23 proteins; Figure 3, Supporting Information Table S10). Surprisingly few associations were seen with variables in the glucose/insulin pathways (fasting glucose with 8; insulin, homeostatic assessment of insulin resistance, and homeostatic assessment of insulin secretion with 0; hemoglobin A1c [HbA1c] with 4, which included ADIPOQ and SHBG; or blood pressure [systolic blood pressure with 4, which included APOM, and diastolic blood pressure with 1, which included AFM]). Eight protein changes were associated with 12-year remission of diabetes (including APOM, ADIPOQ, and SHBG), and eight proteins increased when BMI decreased and were associated with a decreased 10-year risk of CHD (Supporting Information Table S10).

Heat map of clinical variable associations with the 71 protein changes that were associated with change in BMI from a regression model using combined surgery and nonsurgery groups. Colors represent (sign[β] × [−1]log10[p value]). Significance level is Bonferroni-corrected p value (p ≤ 0.05/71/20 = 3.5 × 10−5) for 71 proteins and 20 clinical variables. Red indicates high significance with negative β (the protein increases with a decrease in the clinical variable), and blue indicates high significance with positive β

Because so few glucose metabolism-related proteins were found among the protein changes that were related to BMI change, a PWAS analysis was done for all 1,297 proteins in the surgical group (Supporting Information Figure S2, Supporting Information Table S11). Longitudinal changes in glucose, insulin, homeostatic assessment of insulin resistance, homeostatic assessment of insulin secretion, HbA1c, and blood pressure still did not show highly significant associations with protein changes. The strongest associations of non-BMI-related protein changes were with liver function (particularly protein cluster 3; Supporting Information Figure S2). HDL-C change was associated with ADIPOQ, APOM, WFIKKN2, ghrelin and obestatin prepropeptide (GHRL), SHBG, sonic hedgehog signaling molecule (SHH), and complement C3 (C3). Lower TG were associated with lower AFM, GHR, apolipoprotein A1 (APOA1), apolipoprotein E (APOE), cysteine rich with EGF like domains 1 (CRELD1), and crystallin zeta like 1 (CRYZL1) and with higher tissue inhibitor of metallopeptidase inhibitor 2 (TIMP2). All significant proteins remained significant after repeating the analyses without adjusting the clinical variables for medication use.

Pathway analysis

Overexpressed pathways using GSEA on all 78 proteins in Figure 1 were identified using the SomaScan panel (SomaLogic) of proteins as background (Table 2). Pathways overrepresented by the proteins altered after weight loss through bariatric surgery involved inflammation, cellular growth, and apoptosis. These pathways were represen