Abstract

Context

   Bariatric surgery has been shown to be effective in inducing complete
   remission of type 2 diabetes in adults with obesity. However, its
   efficacy in achieving complete diabetes remission remains variable and
   difficult to predict before surgery.

Objective

   We aimed to characterize bariatric surgery-induced transcriptome
   changes associated with diabetes remission and the predictive role of
   the baseline transcriptome.

Methods

   We performed a whole-genome microarray in peripheral mononuclear cells
   at baseline (before surgery) and 2 and 12 months after bariatric
   surgery in a prospective cohort of 26 adults with obesity and type 2
   diabetes. We applied machine learning to the baseline transcriptome to
   identify genes that predict metabolic outcomes. We validated the
   microarray expression profile using a real-time polymerase chain
   reaction.

Results

   Sixteen patients entered diabetes remission at 12 months and 10 did
   not. The gene-expression analysis showed similarities and differences
   between responders and nonresponders. The difference included the
   expression of critical genes (SKT4, SIRT1, and TNF superfamily),
   metabolic and signaling pathways (Hippo, Sirtuin, ARE-mediated
   messenger RNA degradation, MSP-RON, and Huntington), and predicted
   biological functions (β-cell growth and proliferation, insulin and
   glucose metabolism, energy balance, inflammation, and
   neurodegeneration). Modeling the baseline transcriptome identified 10
   genes that could hypothetically predict the metabolic outcome before
   bariatric surgery.

Conclusion

   The changes in the transcriptome after bariatric surgery distinguish
   patients in whom diabetes enters complete remission from those who do
   not. The baseline transcriptome can contribute to the prediction of
   bariatric surgery-induced diabetes remission preoperatively.

   Keywords: obesity, type 2 diabetes, bariatric surgery, microarray,
   transcriptomics, genetic predictors
     __________________________________________________________________

   Bariatric surgery is now considered the standard therapy for adults
   with obesity and type 2 diabetes who do not respond to a healthy
   lifestyle and optimal medical therapy [[43]1]. However, it does not
   achieve complete remission of type 2 diabetes in approximately a third
   of patients at 1 year [[44]2, [45]3]. Furthermore, the proportion of
   patients in remission can decline after several years of follow-up,
   although the rate varies between studies [[46]3, [47]4]. Therefore,
   identifying adults with obesity and type 2 diabetes who will achieve
   short- and long-term remission of diabetes after bariatric surgery is
   essential for informed decisions before surgery.

   Most standard bariatric surgery procedures, including laparoscopic
   adjustable gastric banding, sleeve gastrectomy, Roux-en-Y gastric
   bypass, and biliopancreatic diversion, promote weight loss and
   metabolic improvement, albeit with variable magnitude [[48]1]. Gastric
   bypass (Roux-en-Y gastric bypass and biliopancreatic diversion) has
   been shown to be superior to restrictive procedures (gastric banding
   and sleeve gastrectomy) in leading to diabetes remission [[49]5].
   However, the likelihood of complete remission of type 2 diabetes after
   bariatric surgery remains difficult to predict at the individual level,
   regardless of the surgical procedure [[50]2, [51]6-8]. Consequently,
   several studies have attempted to identify predictors of metabolic
   outcomes using patients' demographic, clinical, and biochemical factors
   to help clinicians select patients and surgical procedures [[52]2,
   [53]6-8]. However, despite improving the short-term prediction of
   diabetes remission at 1 year, the specificity of these scores has
   remained insufficient to accurately identify individuals who will enter
   diabetes remission [[54]9]. Furthermore, these remission scores have
   limited predictive capacity for long-term metabolic outcomes after
   bariatric surgery [[55]10, [56]11].

   Genetic variants associated with adults with obesity and type 2
   diabetes have been shown to influence weight loss and metabolic
   response to bariatric surgery [[57]12-14]. Few studies have evaluated
   the ability of single or multiple variants to predict diabetes
   remission after bariatric surgery [[58]12, [59]13]. The findings have
   shown that patients with these variants are prone to better metabolic
   outcomes than noncarriers, suggesting that genetic factors can
   contribute to improving the preoperative prediction of diabetes
   remission after bariatric surgery. However, this approach may apply
   only to selected populations with specific variants. Therefore, these
   clinical and genetic studies have led us to hypothesize that
   identifying genetic predictors that can be integrated with clinical
   parameters may improve the prediction of individual response to
   bariatric surgery and form the basis of personalized care.

   Clinical and experimental studies suggest that the transcriptomic
   response to bariatric surgery is critically implicated in restoring
   metabolic homeostasis [[60]15-24]. Type 2 diabetes in adults with
   obesity is characterized by low-grade inflammation, insulin resistance,
   and progressive loss of β cells [[61]25-28]. In rodent models that
   recapitulate obesity-linked type 2 diabetes, duodenal-jejunal bypass
   surgery results in reduced inflammation, improved glucose tolerance and
   insulin sensitivity, and increased regeneration and proliferation of β
   cells [[62]20-22]. These phenotypic changes have been attributed to
   decreased expression of proinflammatory cytokine genes [[63]21],
   modulation of gene expression in the insulin signaling pathway
   [[64]20], and activation of cell cycle regulator genes [[65]22] in the
   liver, jejunum, adipose tissue, and Langerhans islets. Similarly, in
   humans, gene expression studies using whole blood, mononuclear cells,
   and adipose tissue reported a reversal of chronic inflammation
   simultaneously with a reduction of proinflammatory cytokine and
   chemokine gene expression after bariatric surgery [[66]19, [67]23,
   [68]24]. Furthermore, the gene expression of SIRT1, a known regulator
   of glucose and lipid homeostasis, is suppressed in obesity and
   activated by caloric restriction and bariatric surgery concomitantly
   with weight loss and improvement in glucose metabolism [[69]29,
   [70]30]. Together, these studies suggest that the beneficial effect of
   bariatric surgery can be achieved through a rapid restoration of
   altered molecular mechanisms underlying the pathogenesis of type 2
   diabetes and their downstream effects on biological functions.

   In this prospective study, our objective was to analyze the molecular
   changes specific to adults with obesity and type 2 diabetes who enter
   complete remission after bariatric surgery and to distinguish them from
   those who do not. We hypothesized that genes that promote restoration
   of metabolic homeostasis can serve as potential predictors of the
   efficacy of bariatric surgery in inducing complete remission of type 2
   diabetes. We used whole-genome microarrays of peripheral blood
   mononuclear cells (PBMCs) and applied machine-learning models to test
   this hypothesis. We designed the study to capture early and late
   dynamic changes in gene expression that begin rapidly after surgery and
   continue for more than a year. We selected PBMCs because the cells are
   easily accessible. Furthermore, clinical and experimental studies of
   gene expression of diet modifications showed that PBMCs reflect an
   early alteration of lipid and carbohydrate metabolism and inflammation
   of white adipose and liver tissues [[71]31-35]. Gene expression changes
   in PBMCs have also been shown to reflect the pathogenesis [[72]36] and
   pathophysiology [[73]37] of type 2 diabetes associated with obesity.
   These include genes related to insulin signaling, inflammation,
   oxidative stress, DNA damage, and lipid metabolism [[74]36, [75]37].
   These studies support using PBMCs as a surrogate to investigate the
   biology of obesity and diabetes.

Materials and Methods

Patients

   The institutional review board of King Abdullah International Medical
   Research Center approved the study protocol. We performed all methods
   following the ethical principles of the Declaration of Helsinki of the
   World Medical Association for medical research involving human
   subjects.

   We estimated that the size of the study population is between 10 and 30
   patients based on similar transcriptomic studies in adults with obesity
   and type 2 diabetes treated with bariatric surgery [[76]15-17, [77]19,
   [78]38]. We conducted the study from February 2013 to April 2017 at the
   King Abdulaziz Hospital in Al Ahsa and the Imam Abdulrahman Bin Faisal
   Hospital in Dammam, Saudi Arabia. We obtained the written informed
   consent of all participants. We included adult patients (aged ≥18
   years), men and women with obesity (body mass index [BMI] ≥ 35), and
   type 2 diabetes (glycated hemoglobin A[1c] level [HbA[1c] ≥ 7.0%]) of
   less than 5 years. We excluded patients with pregnancy, malignancy,
   psychiatric disorders, and immunosuppressive therapy. Similarly, we did
   not include patients who did not consent or could not sign informed
   consent. Before surgery, we collected demographic information,
   coexisting hypertension and dyslipidemia, and medication history. We
   recorded vital signs, height, and body weight and calculated BMI.

Diabetes Outcomes After Surgery

   We considered patients in complete remission of type 2 diabetes to be
   responders (Rs) if their HbA[1c] level was 5.7% or less without
   diabetes medication 12 months after bariatric surgery. We considered
   patients as nonresponders (NRs) if they were in partial remission
   (HbA[1c] level between 5.8% and 6.4% without diabetes medication) or in
   no remission (HbA[1c] level of 6.5% or higher or remaining on diabetic
   medications). We did not choose the recommended HbA[1c] threshold of
   less than 6.5% for diabetes remission [[79]39] as it falls within the
   classification of a prediabetic state [[80]40]. However, the strict
   criteria chosen here reflect a true remission of diabetes, increasing
   the likelihood of identifying an accurate gene expression response.

Biochemical Analysis

   We obtained blood samples at baseline before surgery (T0), at 2 months
   (T1), and 12 months (T2) after bariatric surgery. We drew blood by
   venipuncture in sterile BD vacutainer EDTA tubes (BD Biosciences) at
   each time point. We measured fasting glucose, liver, kidney, lipid
   profiles, and complete blood count using the automated Alinity-CI
   series system (Abbott). We determined plasma levels of HbA[1c] by
   high-performance liquid chromatography using the D-100 system (Bio-Rad
   Laboratories). We measured C-peptide and insulin levels using a
   sandwich chemiluminescent immunoassay. We determined C-reactive protein
   (CRP) activity using a turbidimetric method using the Alinity-CI series
   (Abbot). We isolated PBMC from whole blood using Ficoll-Paque
   (Ficoll-Paque PLUS; GE Healthcare, Bio-Sciences AB) according to the
   manufacturer's standard protocol. We stored the PBMCs at −80 °C until
   analysis.

Microarray Analysis

   According to the manufacturer's instructions, we used the SV Total RNA
   Isolation Kit (Promega) to extract the total RNA from the PBMCs. We
   measured the quantity and quality of the RNA using the NanoDrop ND-1000
   spectrophotometer (NanoDrop Technologies Inc). We performed
   gene-expression profiling of the samples using the Gene Chip Human
   Genome U133 Plus 2.0 Array (Thermo Fisher Scientific). The array
   includes 54,000 probe sets to analyze 47,000 transcripts and variants
   [[81]41]. Briefly, we reverse-transcribed 500 ng of total RNA into
   double-stranded complementary DNA (cDNA) and then labeled with biotin
   in cRNA (IVT PLUS reagent kit, Thermo Fisher Scientific). We hybridized
   approximately 11 µg of labeled and fragmented cRNA in the arrays, then
   stained and washed the sample on the GeneChip Fluidics Station 450
   (Thermo Fisher Scientific) using the GeneChip Hybridization, Wash, and
   Stain Kit (Thermo Fisher Scientific). We scanned the arrays using an
   Affymetrix 7000 G scanner (Thermo Fisher Scientific).

Validation of Microarray Data by Quantitative Real-time Polymerase Chain
Reaction

   We reverse-transcribed 1000 ng of RNA to cDNA using the high-capacity
   cDNA reverse-transcription kit (Thermo Scientific) from 59 samples,
   consisting of T0 (n = 21), T1 (n = 17), and T2 (n = 21), in good
   quantity and quality following the manufacturer's procedure. We used
   PowerUp SYBR Green reagent (Thermo Scientific) and the FAST
   QuantStudioTM 12K Flex system from Applied Biosystems to validate 12
   selected genes identified in the transcriptomic analysis according to
   company instructions (Supplementary Table 1, data set 2 [[82]42]). We
   used the 18S RNA to normalize the results and mean fold changes in each
   gene. We calculated the mean fold changes in each gene for each sample
   using the 2^−ΔΔCt method as previously described [[83]43].

Bioinformatics Analysis

Quality control and preprocessing of HG-U133 Plus 2.0 microarray data

   Before starting the downstream analysis of the microarray oligo-chip
   data, we first analyzed the data to assess the relative quality of the
   array chips used in the experiment. We used the pseudo-image plots to
   evaluate the spatial distribution of the data on the chips. All HG-U133
   Plus 2.0 chips were good in terms of the spatial distribution of the
   data. We evaluated the quality of the CEL files for each time point
   using the Bioconductor array Quality Metrics program [[84]44]. We used
   the program for outlier detection in MA plots and differences between
   arrays at the two time points. None of the arrays was considered an
   outlier. Moreover, we used other quality measures such as relative log
   expression and normalized unscaled standard errors to ensure the
   quality of the chips and probes therein (Supplementary data set 1
   [[85]42]).

Downstream analysis of the microarray

   After quality control, we processed the raw CEL files using freely
   available updated chip definition files (CDF) for HG-U 133 Plus 2.0
   arrays based on Entrez genes (Supplementary data set 1 [[86]42]). We
   adopted 2-step processes to identify gene sets differentially expressed
   between 2 time points, using time T0 as a baseline. In the first step,
   we calculated the variance of gene expression values at all time points
   (T0, T1, and T2) for each gene in the gene-expression matrix. We
   analyzed the differentially expressed genes between 2 time points in
   the second step, considering T0 as a baseline, using the significance
   analysis of microarrays (SAM) method [[87]45]. We also compared the
   gene expression data between R and NR patients at the start of T0. The
   filtered data were analyzed using a linear model to identify genes
   differentially expressed in R in contrast to NR. We kept the false
   discovery rate (FDR) threshold at less than 0.05 (Supplementary data
   set 1 [[88]42]).

Gene enrichment analysis

   We uploaded to the IPA online application
   ([89]www.qiagen.com/ingenuity) the differentially expressed genes
   (DEGs) after bariatric surgery in T1 and T2 for the entire cohort (n =
   26 patients) and for R (n = 16) and NR (n = 10). We analyzed the
   association of DEGs with canonical pathways and biological functions
   and their predicted activation states. IPA calculates an overlap P
   value associated with each function or pathway using a right-tailed
   Fisher exact test. This estimates the probability that the association
   between our data set of DEGs and a given pathway is due to random
   chance. A P value less than .05 indicates a statistically significant
   nonrandom association. IPA also predicts the activation state of
   canonical pathways by calculating a Z score independent of the overlap
   P value. A Z score greater than 1 indicates that a canonical pathway is
   significantly activated, and a Z score less than −1 designates
   inhibition. We used the application's comparison tool to identify
   similarities and differences between Rs and NRs. Comparison analysis
   allows for the visualization of the canonical pathways side by side
   across time points.

A predictive model of response to bariatric surgery

   We developed a machine-learning-based model to predict the outcome of
   bariatric surgery preoperatively. We formulated the gene expression
   data at baseline for Rs and NRs into a supervised 2-class
   classification problem. We performed all data training and testing
   performance procedures using the Python module called “Scikit-learn”
   [[90]46] (Supplementary data set 1 [[91]42]). This model development
   aimed to select a few genes (or features) that may predict the outcome
   of bariatric surgery in obese patients before surgery. We adopted a
   combined approach to feature selection using i) wrapper-based and ii)
   model-based methods simultaneously. Recursive Feature Elimination with
   Cross-Validation (RFECV) from Scikit-learn was used as a wrapper method
   around logistic regression (LR) as a classifier with a default
   parameter (penalty “l2”; solver = “lbfgs”). RFE initially trains the
   model on all the features, and the coefficient or feature importance of
   each feature is obtained by the classifier or estimator (here, LR). The
   least important features are then eliminated from the feature set, and
   this procedure is repeated until the desired number of features are
   selected. The “SelectFromModel” meta-transformer in Scikit-learn was
   used to select model-based features, along with an estimator that can
   produce sparse solutions. Linear models, especially “logistic
   regression” penalized with L1 norm, is the method of choice to reduce
   the dimensionality of the data due to its ability to produce sparse
   solutions, that is, shrink the feature coefficients to zero and select
   the nonzero coefficients. Here, we used LR (penalty = ‘L1', solver =
   ‘liblinear’) as an estimator with “SelectFromModel” to choose the
   nonzero coefficients and informative features or genes.

   To adopt a combined approach of developing a predictive model to
   classify Rs and NRs using baseline gene expression data, we applied two
   separate feature selection methods in parallel (as described earlier)
   along with the computation of DEGs in Rs (vs NRs). The final gene set
   was selected based on the consensus of these three methods.

Statistical Analysis

   We reported continuous variables as means ± SD and summarized
   categorical variables using frequencies. We modeled the primary end
   points to analyze the change over time and to compare the R and NR
   groups using mixed linear models with the main effects of group, time,
   and interaction terms. The reported P values were based on the
   time-effect and interaction terms from the linear mixed-effects model.
   Comparisons between groups at a single time point were made using the t
   test for continuous variables with normal distribution or Wilcoxon rank
   sum tests for skewed distribution and the chi-square or Fisher exact
   test for categorical variables. We performed all analyses using SAS
   software, version 9.4 (SAS Institute Inc).

Results

Baseline Characteristics

   Fifty-three patients met the inclusion criteria and signed the informed
   consent to participate in the study ([92]Fig. 1). Twelve patients
   withdrew after signing their consent and 5 did not have surgery. We did
   not include 10 of the 36 remaining patients in the final analysis.
   Eight did not complete their 12-month follow-up, 1 patient developed
   multiple strokes 6 months after bariatric surgery, and 1 had
   insufficient RNA. Therefore, information on the primary end point was
   available before and 2 and 12 months after bariatric surgery in 26
   patients. Three samples were missing in T1 and 2 in T2.

Figure 1.

   Figure 1.
   [93]Open in a new tab

   Flow of patients during the study period. CONSORT diagram of 53
   patients who signed informed consent and were enrolled. Twenty-nine
   patients completed all procedures, and 26 completed follow-up at 1
   year.

   We show the patients’ characteristics at baseline in [94]Table 1. There
   were 11 men and 15 women with obesity, a BMI of 48.8 ± 8.4, and
   diabetes (fasting blood sugar 8.9 ± 2.2 and HbA[1c] 7.8 ± 1.3), all
   except one receiving 1 to 4 oral antidiabetic medications combined with
   insulin in 4 patients. Other clinical and metabolic alterations
   included hypertension, dyslipidemia, inflammation, and hyperuricemia
   ([95]Tables 1 and [96]2). The patients underwent various bariatric
   laparoscopic surgical procedures. These consisted of Roux-en-Y gastric
   bypass (n = 14), biliopancreatic diversion (n = 6), sleeve gastrectomy
   (n = 5), and mini-gastric bypass (n = 1) (see [97]Table 1).

Table 1.

   Demographic data, weight, and glucose metabolism variables at baseline
   (before surgery) and at 2 and 12 months (after surgery)^[98]a
   Variables Total
   n = 26 R
   n = 16 NR
   n = 10
   Age, y 40.2 ± 7 40.4 ±7.6 39.8 ± 6.2
   0.842^[99]b
   Sex, M/F 11/15 8/8 3/7
   0.315^[100]b
   Height, cm 162 ± 9.7 163 ± 9.7 159.4 ± 9.6
   0.383^[101]b
   Weight, kg
    Baseline 128 ± 25.9 133 ± 28 121 ± 19.8
    2 mo 109 ± 21.5 112 ± 23.9 104 ± 17.7
    12 mo 88 ± 18.8 88 ± 19.2 87 ± 19.3
    P <.0001^[102]c .048^[103]d
   BMI
   Baseline 48.8 ± 8.4 49.8 ± 8.6 47.4 ± 8.2
    2 mo 41.4 ± 6.7 41.9 ± 6.9 40.7 ± 6.7
    12 mo 33.9 ± 6.3 33.2 ± 5.3 35.4 ± 7.8
    P <.0001^[104]c .146^[105]d
   Fasting blood glucose mmol/L
    Baseline 8.9 ± 2.2 7.9 ± 2.2 8.7 ± 2.3
    2 mo 6.3 ± 0.9 6.0 ± 0.7 6.7 ± 0.9
    12 mo 5.2 ± 0.9 4.9 ± 0.4 5.8 ± 1.1
    P <.0001^[106]c .942^[107]d
   HBA[1c], %
    Baseline 7.8 ± 1.3 7.7 ± 1.4 7.9 ± 1.2
    2 mo 6.2 ± 0.8 5.9 ± 0.7 6.7 ± 0.7
    12 mo 5.5 ± 0.7 5.3 ± 0.6 6.1 ± 0.7
    P <.0001^[108]c .491^[109]d
   Insulin level, mmol/L
    Baseline 18.8 ± 11.8 15.9 ± 10.6 23.6 ± 12.9
    2 mo 9.6 ± 8.2 8.9 ± 8.2 10.6 ± 8.6
    12 mo 5.3 ± 3.1 5.6 ± 3.6 4.9 ± 2.4
    P <.0001^[110]c .102^[111]d
   C-peptide, ng/mL
    Baseline 3.3 ± 1.4 3.3 ± 1.3 3.5 ± 1.5
    2 mo 3.0 ± 1.8 3.0 ± 2.2 2.9 ± 0.9
    12 mo 1.7 ± 0.9 1.5 ± 0.8 1.9 ± 1.1
    P <.0001^[112]c .619^[113]d
   Medications, No.
    Antidiabetic
     Baseline 25 16 9
     2 mo 14 7 7
     12 mo 6 0 6
   .385^[114]b
   Type of surgery^e
    RYGB 14 8 6
    BPD 6 5 1
    SLG 5 3 2
    LGB 1 0 1
   .428^[115]b
   [116]Open in a new tab

   Abbreviations: BMI, body mass index; BPD, biliopancreatic diversion
   without duodenal switch; F, female; HbA[1c], glycated hemoglobin A[1c];
   LGB, laparoscopic mini gastric bypass; M, male; NR, nonresponders; R,
   responders; RYGB, Roux-en-Y gastric bypass; SLG, sleeve gastrectomy.

   ^a Mean ± SD.

   ^b P values are determined by the t test for continuous variables and
   the chi-square or Fisher exact test for categorical variables.

   ^c P values for the main effect of time (baseline, 2, and 12 months
   after bariatric surgery).

   ^d P values of the interaction effect of time (baseline, 2, and 12
   months after bariatric surgery) and groups, R (responders) vs NR
   (nonresponders).

Table 2.

   Effects of bariatric surgery on metabolic syndrome variables at
   baseline (before surgery) and 2 and 12 months (after surgery)^[117]a
   Variables Total
   n = 26 R
   n = 16 NR
   n = 10
   Triglycerides, mmol/L
    Baseline 1.75 ± 1.5 1.83 ± 1.88 1.61 ± 0.56
    2 mo 1.50 ± 0.8 1.53 ± 0.90 1.44 ± 0.69
    12 mo 1.05 ± 0.5 0.99 ± 0.48 1.16 ± 0.43
    P .036^[118]b .485^[119]c
   HDL, mmol/L
    Baseline 1.06 ± 0.3 1.05 ± 0.4 1.07 ± 0.2
    2 mo 0.96 ± 0.2 0.92 ± 0.3 1.02 ± 0.3
    12 mo 1.19 ± 0.3 1.16 ± 0.4 1.23 ± 0.3
    P <.0001^[120]b .751^[121]c
   Uric acid, mmol/L
    Baseline 347 ± 112 346 ± 113 347 ± 115
    2 mo 361 ± 136 361 ± 127 358 ± 155
    12 mo 295 ± 98 296 ± 93 292 ± 109
    P <.0001^[122]b .924^[123]c
   CRP, mg/L
    Baseline 13.4 ± 9.4 12.1 ± 8.6 16 ± 10.9
    2 mo 8.2 ± 5.2 7.4 ± 3.6 9.8 ± 7.6
    12 mo 3.7 ± 3.8 3.5 ± 2.8 4.1 ± 5.3
    P <.0001^[124]b .846^[125]c
   Medications, No.
    Antilipid
     Baseline 14 10 4
     2 mo 13 8 5
     12 mo 10 4 6
   .422^[126]d
    Antihypertensive
     Baseline 9 4 5
     2 mo 8 4 4
     12 mo 5 3 2
   .234^[127]d
   [128]Open in a new tab

   Abbreviations: CRP, C-reactive protein; NR, nonresponders; R,
   responders.

   ^a Mean ± SD.

   ^b P values for the main effect of time (baseline, 2, and 12 months
   after bariatric surgery).

   ^c P values of the interaction effect of time (baseline, 2, and 12
   months after bariatric surgery) and groups, R (responders) vs NR
   (nonresponders).

   ^d P values determined by t test.

Metabolic Outcomes

   One year after bariatric surgery, we classified n = 16 (61.5%) patients
   as Rs and n = 10 (38.4%) as NRs ([129]Fig. 2). We did not find
   statistically significant differences in demographic, clinical, and
   biochemical variables, antidiabetic medications, and type of surgery
   between Rs and NRs before surgery (see [130]Tables 1 and [131]2).

Figure 2.

   [132]Figure 2.
   [133]Open in a new tab

   Graphical summary of the study design. Twenty-six patients with obesity
   and type 2 diabetes who underwent bariatric surgery were studied.
   Plasma and peripheral blood mononuclear cells before surgery and at 2
   and 12 months after surgery were used to determine fasting blood sugar,
   glycated hemoglobin A[1c] (HbA[1c]), C-peptide, and transcriptome. Body
   mass index, fasting blood sugar, HbA[1c], and C-peptide before and
   after surgery are shown as violin plots. The responders vs
   nonresponders phenotypes were based on HbA[1c] and antidiabetic
   medications 12 months after surgery. The machine-learning method
   applied to a pool of 20,183 genes at baseline after classification into
   responders and nonresponders was used to identify genes that can
   predict the metabolic outcome preoperatively.

Weight and Metabolic Changes After Bariatric Surgery

Weight and glucose metabolism

   Bariatric surgery reduced the average weight by 15% and 31% at 2 and 12
   months from baseline. The weight change was reflected by changes in
   BMI, which decreased by 15% and 34%, respectively (see [134]Table 1).
   Fasting blood sugar, plasma HbA[1c], C-peptide, and insulin levels
   decreased significantly with time, indicating a progressive improvement
   in glucose metabolism and restoration of insulin sensitivity. The
   reduction in body weight was significantly different between Rs and
   NRs, but not for HbA[1c] and fasting blood sugar (see [135]Table 1).
   The improvement in glycemic control was paralleled by a progressive
   decrease over time in antidiabetic medications until their
   discontinuation in Rs and in 4 of 10 NRs.

Metabolic syndrome

   Bariatric surgery progressively reversed all biomarkers of metabolic
   syndrome, including a significant decrease in triglycerides and
   high-sensitivity CRP levels, an increase in high-density lipoprotein
   cholesterol, and a reduction in the number of medications needed to
   treat hypertension (see [136]Table 2). However, antilipid medications
   decreased in Rs but not in NRs. There were no significant differences
   in any metabolic syndrome variables examined in this study between Rs
   and NRs (see [137]Table 2).

Transcriptional Changes After Bariatric Surgery

   The analysis identified 2150 DEGs in T1 and 2711 in T2, compared to
   baseline (T0), using a P value less than .05 and an FDR of less than
   0.05 after Benjamini-Hochberg correction for multiple tests. The
   percentage of DEGs in T1 and T2 was 3.98% and 5.02% of the total genes
   included in the GeneChip array, respectively [[138]41]. Pathway
   enrichment analysis showed that DEGs were related to proteolysis and
   stress response, senescence, and apoptosis (ubiquitination signaling,
   sumoylation signaling, and EIF2 signaling), immune response, and
   cytokine signaling (T-cell receptor signaling, Cdc 42, Th1 signaling,
   and antigen presentation signaling) both in T1 and T2 after surgery
   (Supplementary Fig. 1, data set 1 [[139]42]). In the context of type 2
   diabetes, ubiquitination, sumoylation, Cdc 42, and immune signaling are
   involved in glucose metabolism, insulin regulation, and the development
   of insulin resistance in obesity [[140]47-50].

Transcriptional Changes in Responders and Nonresponders After Bariatric
Surgery

   The analysis identified 1217 DEGs in Rs and 1648 in NRs in T1 compared
   to their respective baseline expression data. In T2, there were 1191
   DEGs in Rs and 1650 in NRs. The heat map representing an unsupervised
   hierarchical cluster analysis ([141]Fig. 3A) and principal component
   analysis ([142]Fig. 3B) revealed different responses of gene expression
   to bariatric surgery between Rs and NRs, supporting the hypothesis that
   molecular alterations over time distinguish patients in whom diabetes
   enters complete remission vs those in which it did not. The Venn
   diagram ([143]Fig. 3C) displays the number of distinct and shared genes
   between Rs and NRs in T1 and T2.

Figure 3.

   [144]Figure 3.
   [145]Open in a new tab

   Transcriptomic changes in adults with obesity in complete (r) and
   incomplete (NR) diabetes remission after bariatric surgery. A, Heat map
   of unsupervised hierarchical clustering (only on rows, ie, genes)
   showing the differentially expressed genes (DEGs) in R and NR
   separately across different time points (T0, T1, and T2). Each column
   represents the gene expression profile of a patient. Horizontal color
   bars at the top indicate the study time points T0 (baseline, before
   surgery), T1 (2 months after surgery), and T2 (12 months after surgery)
   for each group responders (R) or nonresponders (NR). The color scale is
   the normalized expression values from microarray data. B, Principal
   component analysis of gene expression data (containing approximately
   20,000 genes) in R and NR. The first two principal components
   (explaining ∼75% of variance) were able to separate R and NR samples.
   C, The Venn diagram shows the number of different and shared DEGs in R
   (n = 16) and NR (n = 10) in T1 and T2 after bariatric surgery.

Difference in Gene and Canonical Pathways Between Responders and
Nonresponders

   The DEGs were significantly associated with 55 and 230 canonical
   pathways in Rs and NRs in T1 (Supplementary Tables 2A and 2B, data set
   2 [[146]42]). In T2, DEGs were associated with 73 pathways in Rs and
   184 pathways in NRs (Supplementary Tables 3A and 3B, data set 2
   [[147]42]) using a threshold value of P less than .05. Comparison
   analysis revealed that Rs and NRs shared the same predicted activation
   or inhibition patterns in the most enriched metabolic and signaling
   canonical pathways. However, 5 canonical pathways differed between Rs
   and NRs in directionality, expression changes in critical genes, and
   functional biological end points ([148]Fig. 4). These pathways are
   related to the regulation of β-cell survival, proliferation, and
   regeneration (Hippo signaling pathway), regulation of energy metabolism
   and insulin signaling (sirtuin and MSP [macrophage-stimulating
   protein]-RON [recepteur d’origine nantais] signaling pathways),
   inflammation and oxidative stress (sirtuin, inhibition of ARE [AU-rich
   elements]-mediated messenger RNA (mRNA) degradation, and MSP-RON
   signaling pathways), and neurodegeneration and neuronal cell death
   (Huntington disease signaling pathways).

Figure 4.

   Figure 4.
   [149]Open in a new tab

   Differential metabolic and signaling pathways between complete and
   incomplete diabetes remission after bariatric surgery. Comparison
   analysis of canonical pathways based on the Z score in responders (R)
   and nonresponders (NR) at 2 months (T1) and 12 months (T2) after
   bariatric surgery relative to before surgery (T0). The yellow and blue
   colors indicate the pathways that are predicted to be activated (Z
   score > 1) or inhibited (Z score < −1), respectively.

   The heat map of the Hippo signaling pathway gene revealed that 18 genes
   were differentially regulated between R and NR, leading to increased
   cell proliferation (Supplementary Table 4, data set 2 [[150]42]). Among
   these genes, SKT4 (serine/threonine kinase 4), also known as mammalian
   sterile 20-like kinase (MST1) 1, a critical mediator of β-cell
   apoptotic death and failure in diabetes, was strongly upregulated in
   NRs but not in Rs ([151]Fig. 5).

Figure 5.

   [152]Figure 5.
   [153]Open in a new tab

   Hippo signaling pathway with prediction of overlaid molecular activity
   in A, responders, and B, nonresponders after bariatric surgery. A,
   Diagram of the canonical Hippo signaling pathway showing 6 genes
   upregulated (red) and 12 genes that did not reach statistical
   significance (gray) 12 months after bariatric surgery (T2) in
   responders (R), and B, 13 genes upregulated (red) and 5 genes that did
   not reach statistical significance (gray) in nonresponders (NR), along
   with predictions of biological functions. Cell proliferation (colored
   orange) is predicted to increase both in R and NR. The inset square
   shows that the STK4 gene is upregulated in T1 (3) and T2 (4) in NR but
   not in R (1 and 2). Pathway and molecular activity prediction analyses
   were generated using QIAGEN Ingenuity Pathway Analysis (IPA, QIAGEN,
   [154]www.qiagen.com/ingenuity).

   Seventy-two genes, including SIRT1 (sirtuin 1), were differentially
   regulated between Rs and NRs in the sirtuin signaling pathway, leading
   to opposite downstream biological functions (Supplementary Table 5,
   data set 2 [[155]42]). SIRT1 was upregulated in Rs but not in NRs,
   leading to a signaling cascade with the activation of glycolysis,
   lipolysis, adipogenesis, oxidative stress, and Alzheimer disease
   ([156]Fig. 6).

Figure 6.

   [157]Figure 6.
   [158]Open in a new tab

   Diagram of the SIRT1 signaling pathway with predicted molecular
   activity overlaid on A, responders, and B, nonresponders after
   bariatric surgery. A, Diagram of the canonical SIRT1 signaling pathway
   showing 23 genes upregulated (red) and 1 downregulated (green) and 48
   genes that did not reach statistical significance (gray) 12 months
   after bariatric surgery (T2) in responders (R). B, In nonresponders
   (NR), 50 genes were upregulated and 22 genes did not reach statistical
   significance, along with predictions of biological function. Cell
   proliferation, inflammation, neurogenesis, and elevated high-density
   lipoprotein (HDL) cholesterol (colored orange) are predicted to
   increase, while apoptosis, oxidative stress, hypertension, Alzheimer
   disease, and renal anemia (colored blue) are predicted to decrease in
   A, responders, and B, the inverse in NR. The inset square shows that
   SIRT1 is upregulated in T1 (1) and T2 (2) in R but not in NR (3 and 4).
   Pathway and molecular activity prediction analyses were generated using
   QIAGEN Ingenuity Pathway Analysis (IPA, QIAGEN,
   [159]www.qiagen.com/ingenuity).

   Inhibition of ARE-mediated mRNA degradation and MSP-RON signaling in
   macrophage pathways revealed different regulatory mechanisms of
   inflammation after bariatric surgery (Supplementary Tables 6 and 7,
   data set 2 [[160]42]). For example, genes in the tumor necrosis factor
   (TNF) superfamily were upregulated in NRs but not in Rs, leading
   through two signaling cascades to enhance mRNA stability and thus
   sustained inflammation, suggesting a role for regulated mRNA decay in
   homeostasis after bariatric surgery (Supplementary Fig. 2, data set 1
   [[161]42]). However, the signaling pathway of MSP-RON in macrophages
   predicted a dual response of increased anti-inflammatory response. It
   decreased inflammation mediated by upregulation of signal transducer
   and activator of transcription 3 (STAT3) and downregulation of nuclear
   factor kappa B (NFκB), respectively (Supplementary Fig. 3, data set 1
   [[162]42]).

   Huntington signaling pathways were predicted to be upregulated in NRs
   but not in Rs in T1 and T2 (Supplementary Table 8, data set 2
   [[163]42]). This led to a signaling cascade that included an increase
   in the calpain, heat shock protein 70 (HSP70), ubiquitin, and
   proteasome 26 genes, resulting in a predicted increase in neuronal cell
   death and neurodegeneration (Supplementary Fig 4, data set 1
   [[164]42]).

Difference in Transcriptome Between Responders and Nonresponders at Baseline

   Using the characterized metabolic response phenotypes after bariatric
   surgery, we compared the baseline transcriptome of Rs and NRs before
   surgery (see [165]Fig. 2). We identified a significant difference in
   DEGs and their predicted downstream biological functions between Rs and
   NRs. Notably, genes associated with DNA replication, recombination, and
   repair and involved in various cancer signaling pathways were distinct
   (Supplementary Fig. 5, data set 1 [[166]42]).

Early Prediction of Response to Bariatric Surgery

   With baseline gene expression data for Rs and NRs, we developed
   predictive models to identify Rs and NRs before surgery. We selected
   the best set of genes that could contribute the most to this
   classification ([167]Fig. 7). Model-based LR with L1 regularization
   (feature selection 1) yielded 35 genes (GeneSet A) with nonzero
   coefficients. RFECV with LR (feature selection 2) selected 10 genes
   (GeneSet B). The differential analysis of Rs vs NRs identified 4200
   DEGs (GeneSet C) with an FDR cutoff of 0.05. The consensus of these
   sets of genes produced the final set of genes, which included 10 genes
   (ARAF [A-Raf proto-oncogene, serine/threonine kinase], DNMT3A [DNA
   methyltransferase 3 alpha], GMDS [GDP-mannose 4,6-dehydratase], HUWE1
   [HECT, UBA and WWE domain containing E3 ubiquitin protein ligase 1],
   KLF7 [KLF transcription factor 7], LGMN [legumain], PEX14 [peroxisomal
   biogenesis factor 14], PPIE [peptidylprolyl isomerase E], RNF157 [ring
   finger protein 157], and STX6 [syntaxin 6]) ([168]Table 3). The feature
   selection was applied to the training set (containing 60% of the
   samples). The performance of these selected genes was evaluated on an
   independent test set (containing 40% of the samples), resulting in a
   precision, recall, F1 score, and accuracy of 1.0 ([169]Fig. 7A). In
   conclusion, the model based on these final 10 genes was able to
   correctly predict the response to bariatric surgery.

Figure 7.

   [170]Figure 7.
   [171]Open in a new tab

   Modeling of baseline transcriptome using machine learning. A, A
   predictive model to classify responders (R) and nonresponders (NR)
   using baseline gene expression data. Two different selection methods
   were applied in parallel using machine learning and computation of
   differentially expressed genes in R vs NR. Data training and testing
   performance were performed using the Python module “Scikit-learn.” The
   final gene set was selected based on the consensus of these methods. B,
   Heat map shows normalized gene expression profiles for the 10 most
   significant genes selected by the machine-learning approach (left) and
   for 10 randomly sampled genes (right) from a pool of 20,183 genes in 26
   patients before bariatric surgery. A blue horizontal line separates
   responders (R) from nonresponders (NR).

Table 3.

   Genes established at baseline differentiate responders vs nonresponders
   after bariatric surgery
   Gene Association with type 2 diabetes and obesity References