High-throughput mediation analysis of human proteome and metabolome identifies mediators of post-bariatric surgical diabetes control

Abstract

To improve the power of mediation in high-throughput studies, here we introduce High-throughput mediation analysis (Hitman), which accounts for direction of mediation and applies empirical Bayesian linear modeling. We apply Hitman in a retrospective, exploratory analysis of the SLIMM-T2D clinical trial in which participants with type 2 diabetes were randomized to Roux-en-Y gastric bypass (RYGB) or nonsurgical diabetes/weight management, and fasting plasma proteome and metabolome were assayed up to 3 years. RYGB caused greater improvement in HbA1c, which was mediated by growth hormone receptor (GHR). GHR's mediation is more significant than clinical mediators, including BMI. GHR decreases at 3 months postoperatively alongside increased insulin-like growth factor binding proteins IGFBP1/BP2; plasma GH increased at 1 year. Experimental validation indicates (1) hepatic GHR expression decreases in post-bariatric rats; (2) GHR knockdown in primary hepatocytes decreases gluconeogenic gene expression and glucose production. Thus, RYGB may induce resistance to diabetogenic effects of GH signaling.Trial Registration: Clinicaltrials.gov NCT01073020.

Conflict of interest statement

For the parent study (SLIMM-T2D) from which samples were derived for analysis in the current study, D.S., A.V., F.H., and A.B.G. received unrestricted funding from the Herbert Graetz Fund and from Covidien, to support surgical procedures for research participants with BMI 30-35 (not covered by insurance). Some supplies for the SLIMM-T2D clinical study were received from Lifescan, a division of Johnson and Johnson, Nestle Inc, and Novo Nordisk. M.E.P. received unrestricted investigator-initiated research grant funding from Medimmune to support the present work, and funding to support assay costs from SomaLogic. ABG served on advisory boards for Baranova, and Kowa. DCS received funds from PCORI and is a stock/shareholder of GI Windows. A.K., B.H., C.M.R., J.G., and Y.S. were employees of Medimmune when the work was initiated. Y.Y. is now employed at Vertex, F.H. at Form Health, A.K. at Sanofi, B.H. at Immunocore, J.G. at AstraZeneca, and A.B.G. at Novartis Institutes of Biomedical Research. All other authors declare no competing interests.

© 2021. The Author(s).

Figures

Fig. 1. Graphical overview shows study flow including clinical study schematic (top) and analysis (bottom).

Differential analyte and pathway analysis (bottom left) identified a network composed of top analytes from Valine, Leucine, and Isoleucine Degradation and Beta-Alanine Metabolism. Mediation analysis (bottom right) identified GHR as mediator of decrease in HbA1c. Triangles denote measured proteins or metabolites, while circles denote proteins or metabolites in the pathway not measured; symbols are colored according to between-group Z scores, using heat map indicated, with the orientation of triangle also indicating up- or downregulation, respectively. Source data are provided as a Source Data file.

Fig. 2. Differentially abundant analytes with time.

Heatmaps of a proteins and b metabolites. Heatmap of log2(RYGB/DWM) at all time points (post-baseline log2 abundance values are baseline-corrected) for analytes that are differentially abundant at any time point (FDR < 0.15 and fold-change absolute value > 1.5). The range of colors show log2(RYGB/DWM) from −1.6 to 1.6, which corresponds to fold changes from −3 to 3. Analytes with a larger absolute value of fold-change are shown as only having fold change absolute value of 3, so that weaker fold changes are easily observed. Source data are provided as a Source Data file.

Fig. 3. Phospholipid biosynthesis and other top-ranking differential pathways.

a Top-ranking differential pathways. Graph illustrates –log10(p-values) and false discovery rates (FDRs), with bars colored by FDR, from Limma Roast’s Mixed statistic (which is one-sided) adjusted with FDRs. b Phospholipid biosynthesis network. Nodes are colored by between-group z-score, whereas unmeasured nodes are colored dark gray. Orientation of triangle also indicates the directionality of regulation. Connections are from the Pathway Commons network. Source data are provided as a Source Data file.

Fig. 4. Branched chain amino acids (BCAA) and downstream metabolites.

a Schematic of BCAA metabolic pathway showing measured analytes. KIC, KMV, and KIV indicate the branched chain ketoacids ketoisocaproate, ketomethylvalerate, and ketoisovalerate, respectively. b Relative abundance of measured analytes, aligned to position in pathway in (a). Data were analyzed by two-sided moderated t-tests; post-baseline time points were analyzed using change from baseline. Data are reported as mean ± SEM on the log2 scale. Analysis was derived from samples from independent human participants. RYGB at 0, 3, 12, and 36 months: n = 19, 18, 19, and 13; DWM: n = 19, 18, 16, 9. Nominal p-values are as follows: Leucine: 3 mo **= 0.001, 12 mo **= 0.0041, 36 mo *= 0.03; Isoleucine: 12 mo *= 0.017, 36 mo *= 0.036; Valine: 3 mo **= 0.0019, 12 mo **= 0.0013, 36 mo *= 0.016; KIC: 3 mo ***= 0.00016; KIV: 3 mo: ***= 0.00026, 12 mo: **= 0.0045. 3-OH-isobutyrate: 3 mo: #<0.0001; isovalerylcarnitine: 3 mo: #<0.0001, 12 mo: **= 0.002; 2-methylbutyrylcarnitine: 0 mo: *= 0.025; propionylcarnitine: 3 mo: #<0.0001, 12 mo: *= 0.016. isobutyrylcarnitine: 0 mo: *= 0.024. Source data and FDRs are available in Supplementary Data 2.

Fig. 5. Beta-alanine metabolism.

a Network nodes are colored by between-group z-score, whereas unmeasured nodes are colored gray. Orientation of triangle also indicates the directionality of regulation. Connections are from the Pathway Commons network. b Log2 abundance of histidine measured by metabolomics in samples from independent human participants at baseline, 3, 12, and 36 months from 19, 18, 16, and 9 DWM and 19, 18, 19, and 13 RYGB participants. c Log2 abundance of CNDP1 measured by SOMAscan, measured in samples from independent human participants at baseline, 3, 12, 24, and 36 months from 19, 19, 16, 10, and 9 DWM and 19, 19,19, 15, and 14 RYGB participants. d CNDP1 plasma levels measured by ELISA at baseline, 3, 12, 24, and 36 months from 10, 10, 10, 6, and 4 DWM participants and 10, 10, 9, 5, and 6 RYGB participants, respectively. Data were analyzed by two-sided moderated t-tests; post-baseline time points were analyzed using change from baseline. Data in b−d are reported as mean ± SEM. Nominal p-values are as follows: b histidine: 3 mo: #<0.0001, 36 mo: *=0.019; c CNDP1 (SOMAscan): 3 mo: #<0.0001, 12 mo: ***=0.00014; 24 mo: **=0.0054; 36 mo: **=0.0067; d CNDP1 ELISA: 3 mo: **=0.0036. Supplementary Data 2 includes FDRs and source data for (b/c); source data for d is provided as a Source Data file.

Fig. 6. GHR is a consistent mediator of change in HbA1c.

Retinol is an inconsistent mediator (i.e., it suppresses or inhibits the causal effect; left side) whereas GHR is a consistent mediator of HbA1c change at 1 year (right side). aX-axis represents log2 abundance change from baseline to the 3 month time point, and Y-axis represents HbA1c change from baseline at 1 year. b Shows putatively causal links with arrows based on (A) and direction of change with triangles pointed up (increase) or down (decrease). Source data are provided as a Source Data file.

Fig. 7. GHR is reduced after bariatric surgery in humans and rodents and regulates hepatic glucose metabolism.

a GHR in plasma is reduced after RYGB in humans, measured by SOMAscan in samples from independent human participants at baseline, 3, 12, 24, and 36 months from 19, 19, 16, 10, and 9 DWM and 19, 19,19, 15, and 14 RYGB participants. bGhr expression in liver, assessed by qRT-PCR, from diet-induced obese rats, 8 weeks post-VSG bariatric surgery (VSG) or sham controls (Sham) (n = 13 VSG, n = 15 sham). c–g Expression of GH signaling/axis genes, assessed by qRT-PCR, in cells treated with siRNA targeting Ghr (si-Ghr) or non-targeting scrambled control siRNA (NT-ctrl) (n = 6 per group). c Ghr, d Igf-1, and e Socs1/2 in mouse primary hepatocytes treated with siRNA-Ghr or NT-control. f Glucose production in response to insulin, cAMP, and insulin/cAMP after si-Ghr or NT-ctrl. g Expression of gluconeogenic genes G6pc and Pepck in response to si-Ghr or NT-ctrl. Data are reported as mean ± SEM. For a the p-values are nominal, using a two-sided T-test; FDRs are reported in Supplementary Data 2. For b–e and g two-sided t-tests were applied; for f two-way ANOVA with Bonferroni’s multiple comparison test was applied in GraphPad Prism. P values are as follows: a: GHR: 3 mo: ***= 0.00099; 12 mo: #<0.0001, 24 mo: **= 0.0024, 36 mo: **= 0.0013. b: *= 0.0155, c: # <0.0001, d: # <0.0001; e: Socs1: *= 0.017, Socs2: ***= 0.0002; f: basal vs. insulin in NT control: *= 0.0322; basal vs. cAMP in NT control: ***= 0.0004; NT control vs. siGHR in cAMP + insulin: *= 0.0328. g: G6pc and Pepck both #<0.0001. Source data for a is provided in Supplementary Data 2. Source data for panels b–g are provided in the Source Data file. For expression analysis, two independent experiments were performed, while glucose production data represent a single experiment.