Graphical abstract

   graphic file with name fx1.jpg
   [41]Open in a new tab

Highlights

     * •
       Exercise causes changes in hundreds of serum proteins in overweight
       individuals
     * •
       Gut immunity-related proteins are associated with metabolic
       adaptations to exercise
     * •
       Exercise responders and non-responders exhibit differential
       proteomic changes
     * •
       Baseline protein signatures accurately predict the metabolic
       outcomes of exercise
     __________________________________________________________________

   Diaz-Canestro et al. identify circulating proteins that predict the
   metabolic responsiveness to exercise training in individuals with
   prediabetes. These findings may facilitate the clinical implementation
   of personalized exercise interventions for diabetes prevention.

Introduction

   Exercise training (ET) is a cost-effective intervention for the
   prevention and management of obesity and type 2 diabetes mellitus
   (T2DM).[42]^1^,[43]^2 In patients with T2DM, regular exercise increases
   insulin sensitivity and secretion, leading to improved glucose
   tolerance. While insulin sensitivity can be enhanced by a single bout
   of exercise,[44]^3 long-term exercise interventions are required to
   improve pancreatic beta cell function in patients with
   T2DM.[45]^4^,[46]^5 It has been proposed that metabolic adaptations to
   exercise involve complex organ crosstalk mediated by
   exercise-responsive factors, known as exerkines, secreted from skeletal
   muscle, adipose tissue, liver, gut, and other organs.[47]^6 In recent
   years, research efforts have been focused on exerkines secreted by
   skeletal muscle, referred to as myokines. The established myokine par
   excellence, i.e., interleukin-6 (IL-6), is released during muscle
   contractions, leading to improved glucose homeostasis via autocrine,
   paracrine, and endocrine effects.[48]^7 In addition, skeletal muscle
   releases other factors including, but not limited to, interleukin-15
   (IL-15), interleukin-8 (IL-8), growth/differentiation factor 15,
   myostatin (MSTN), apelin, and irisin, which have been implicated in
   promoting lipolysis and browning of white adipose tissue, increasing
   muscle mass, and improving glucose tolerance. Molecules secreted by
   other organs, such as adipose tissue and liver, also play critical
   roles in mediating metabolic adaptations.[49]^8 For example, the
   effects of ET on glucose and lipid homeostasis as well as insulin
   sensitivity were partly attributed to increased circulating levels of
   adipocyte-secreted adiponectin via enhancing fibroblast growth factor
   21 (FGF21) sensitivity in adipose tissue.[50]^9 Interestingly, FGF21
   levels increase in response to acute exercise but decrease after
   chronic exercise in obese and prediabetic individuals, possibly due to
   improved FGF21 sensitivity.[51]^10^,[52]^11 Indeed, the response to
   acute and chronic exercise differs in a large number of circulating
   proteins,[53]^10 and this inconsistency may partly reflect differences
   in metabolic adaptations. A recent multi-omics study has described the
   acute exerkine response in healthy and insulin-resistant
   individuals.[54]^12 However, how exerkines coordinate spatially and
   temporally to confer the chronic metabolic benefits of exercise remains
   poorly understood. A targeted analyses of potential exerkines mediating
   the effects of chronic exercise have yet to be performed in humans,
   notably in individuals with impaired glucose tolerance.

   Large inter-individual variability in the metabolic response to
   exercise has been established by previous
   studies.[55]^13^,[56]^14^,[57]^15 From 7% up to 44% of individuals with
   prediabetes do not respond favorably to exercise interventions in terms
   of metabolic outcomes.[58]^16 Such heterogeneous results have been
   hypothetically attributed to genetics, epigenetics, and physiological
   factors; however, the underlying mechanisms remain elusive.[59]^17 A
   recent study demonstrated that microbial species and associated
   metabolites were differently modulated by a 12-week high-intensity
   interval training (HIIT) program in metabolic “responders” and
   “non-responders,” while baseline microbiome features accurately
   predicted metabolic responsiveness to the exercise intervention.[60]^11
   These findings disclosed a key role for gut microbiota in conferring
   the metabolic benefits of exercise. However, additional investigations
   are required to further elucidate the molecular mechanisms that drive
   the lack of metabolic response to ET in individuals with prediabetes,
   as well as to discover baseline circulating biomarkers that may
   discriminate between those most likely to benefit from those who do
   not.

   To address the above questions, we performed a targeted proteomic
   analysis (688 inflammatory and metabolic proteins) in serum samples
   from well-characterized medication-naive overweight and obese Chinese
   men with prediabetes before and after chronic (12-week HIIT) exercise.
   We also determined the relationship between serum proteins and clinical
   parameters to identify potential mediators and predictors of metabolic
   outcomes. Furthermore, molecular responses were compared between
   responders and non-responders in terms of insulin resistance. Finally,
   a machine-learning (ML) algorithm that integrated baseline proteomics
   was developed to predict individual responsiveness to ET.

Results

Cohort characteristics

   Thirty-six medication-naive overweight and obese men with prediabetes
   underwent 12-week HIIT ([61]Figure 1A). After the exercise
   intervention, body weight and adiposity were reduced, together with
   significant improvements in insulin sensitivity, glucose tolerance, and
   lipid profiles ([62]Table S1). Considering the important role of
   circulating proteins as potential transducers of the metabolic benefits
   of exercise, we next explored the proteomic changes in response to
   12-week HIIT in our cohort.

Figure 1.

   [63]Figure 1
   [64]Open in a new tab

   Proteomic changes in response to 12-week high-intensity interval
   training (HIIT)

   (A) Schematic diagram of the study design.

   (B and C) Volcano plots of the change in individual proteins after (B)
   4 and (C) 12 weeks of HIIT in overweight and obese men with prediabetes
   (n = 36).

   Orange and blue dots indicate proteins that are significantly
   upregulated or downregulated, respectively, as determined by paired
   Wilcoxon rank-sum test, FDR < 0.05. Gray dots indicate proteins without
   significant changes.

Serum proteomic changes in response to 12-week HIIT

   To identify plasma proteins that were altered in response to HIIT, we
   compared the serum levels of 688 proteins at baseline versus 4 and
   12 weeks after HIIT using Olink’s antibody-based proteomics platform
   ([65]Table S2). HIIT induced changes in 89 circulating proteins (22
   upregulated and 67 downregulated) after 4 weeks (12.9% of total
   proteins, false discovery rate [FDR] < 0.05; [66]Figure 1B) and 247
   proteins (75 upregulated and 173 downregulated) after 12 weeks (35.9%
   of total proteins, FDR < 0.05; [67]Figure 1C) compared with baseline.
   Growth hormone 1 (GH1) exhibited the highest upregulation (week 12,
   log2 fold change [FC] = 1.03, FDR < 0.05), whereas pro-apoptotic
   factors such as BH3-interacting domain death agonist (BID) were among
   the proteins with the highest downregulation (week 12, log2 FC = −1.30,
   FDR < 0.05). Next, to assess the specificity of this technology, we
   tested the reproducibility of 15 inflammatory proteins assessed by
   Olink technology in 19 individuals using a different platform (i.e.,
   Mesoscale Discovery [MSD]). Eleven out of 15 assays were highly
   correlated, and 4 out of 15 were moderately correlated, suggesting a
   high reproducibility between the two technologies ([68]Figure S1).

Different trajectories of serum proteomic changes in response to 12-week HIIT

   To better understand the changes induced by 12-week HIIT, proteins that
   significantly changed during 12-week HIIT (FDR < 0.05) were further
   examined using cluster analysis. Six main clusters of longitudinal
   trajectories were identified that delineated different patterns in
   response to the exercise intervention ([69]Figure 2A) and a subset of
   significant proteins (FDR < 0.05) that showed the highest change (in
   terms of log2 FC) in each cluster are shown in [70]Figure 2B.

Figure 2.

   [71]Figure 2
   [72]Open in a new tab

   Longitudinal trajectories of proteomic changes in response to 12-week
   HIIT

   (A) Clustering of longitudinal trajectories using circulating proteins.

   (B) A subset of significant proteins (FDR < 0.05) that showed the
   highest change in response to 12-week HIIT (in terms of log[2] fold
   change [FC]) in overweight and obese men with prediabetes (n = 36) are
   shown in the heatmap.

   Data are represented as mean log[2] FC relative to baseline. Proteins
   were grouped by clusters. Significant proteins were determined by
   paired Wilcoxon rank-sum test, FDR < 0.05.

Cluster 1

   Proteins in cluster 1 (n = 55, FDR < 0.05 at week 12) increased in the
   first 4 weeks of HIIT and continued to increase until the end of the
   intervention, although to a lesser extent. GH1 showed the highest
   upregulation, followed by cluster of differentiation 93 (CD93) (week
   12, log2 FC = 0.44, FDR < 0.05), a receptor expressed in myeloid and
   endothelial cells that contributes to the removal of apoptotic
   cells,[73]^18 and SPARC-related modular calcium-binding protein 2
   (SMOC2) (week 12, log2 FC = 0.47, FDR < 0.05), a matricellular protein
   relevant to bone homeostasis.[74]^19 Furthermore, proteins that
   stimulate hematopoiesis such as erythropoietin (EPO) and
   colony-stimulating factor 3 (CSF3), as well as myokines (i.e., IL-15,
   IL-6, IL-8/CXCL8) implicated in energy metabolism and angiogenesis,
   also presented this trajectory.

Cluster 2

   Proteins in cluster 2 (n = 15, FDR < 0.05 at week 12) presented a
   delayed increase in response to 12-week HIIT compared with proteins in
   cluster 1. A large portion of the molecules in cluster 2 are involved
   in inflammation, including the anti-inflammatory cytokine IL-10.

Cluster 3

   Cluster 3 contained proteins (n = 54, FDR < 0.05 at week 12) that
   decreased in the first 4 weeks of HIIT and continued to decrease to the
   same degree until the end of the intervention. The strongest decreases
   were observed in BID and the enzyme catalyzing muscle glycogen
   synthesis (muscle glycogen synthase [GYS1]) (week 12, log2 FC = −1.32,
   FDR < 0.05). Many other proteins had the same trajectory including
   kynurenine-oxoglutarate transaminase 1 (KYAT1), an enzyme involved in
   tryptophan metabolism,[75]^20 and interleukin-1 beta (IL-1β).

Cluster 4

   Cluster 4 was enriched in proteins (n = 12, FDR < 0.05 at week 4) that
   decreased after the first 4 weeks of HIIT and returned to baseline by
   the end of the intervention. Two appetite-stimulating hormones,
   agouti-related protein (AGRP; week 4, log2 FC = −0.33, FDR < 0.05) and
   ghrelin (GHRL; week 4, log2 FC = −0.34, FDR < 0.05), exhibited the
   strongest decreases in this cluster.

Cluster 5

   Cluster 5 was abundant in proteins (n = 23, FDR < 0.05 at week 12) that
   decreased in the first 4 weeks of HIIT but did not return to baseline
   by the end of the intervention. Well-known hormones such as MSTN (week
   12, log2 FC = −0.46, FDR < 0.05) and leptin (LEP; week 12, log2 FC =
   −0.50, FDR < 0.05) presented this pattern.

Cluster 6

   Cluster 6 contained proteins (n = 43, FDR < 0.05 at week 12) that
   required more time to respond to exercise, exhibiting a significant
   reduction by the end of the intervention. Of these, mesencephalic
   astrocyte-derived neurotrophic factor (MANF) showed the highest
   decrease (week 12, log2 FC = −1.19, FDR < 0.05). In addition, some
   pro-apoptotic markers such as caspase-3 (CASP3) and diablo homolog
   (DIABLO) followed this pattern. Taken together, the above findings
   demonstrated that hundreds of proteins changed in response to 12-week
   HIIT, mainly delineating six different trajectories. To further
   understand the biological significance of these changes, we next
   performed pathway enrichment analysis.

Ingenuity pathway analysis (IPA) in response to 12-week HIIT

   IPA was performed among the significantly altered serum proteins during
   12-week HIIT. In the early phase of HIIT (the first 4 weeks), the “EPO
   signaling pathway” was the only pathway identified as activated
   ([76]Figure S2A). Conversely, seven pathways were significantly
   downregulated including the “hepatic fibrosis signaling pathway.”
   Likewise, analysis of diseases and biological functions revealed an
   increase in “quantity of myeloid cells” ([77]Figure S2B), whereas
   several functions including “activation and migration of leukocytes”
   ([78]Figure S2C) as well as “generation of reactive oxygen species”
   were decreased ([79]Figure S2D).

   After 12 weeks of HIIT, 7 pathways were identified as activated
   ([80]Figure 3A), with most of them playing a key role in metabolism
   (i.e., “peroxisome proliferator-activated receptor (PPAR) signaling,”
   “IL-15 production,” and “liver X receptor (LXR)/retinoid X receptor
   (RXR) pathway”). Conversely, 30 pathways with a high number implicated
   in inflammatory (e.g., “NF-κB,” “Toll-like receptor signaling,” and
   “pyroptosis signaling”) and apoptotic responses (e.g., “death receptor
   signaling,” “apoptosis signaling,” and “tumor necrosis factor receptor
   1 (TNFR1) signaling”) were inhibited. Analysis of diseases and
   biological functions demonstrated increased “angiogenesis”
   ([81]Figure 3B), whereas “activation and recruitment of leukocytes”
   ([82]Figure 3C) as well as “apoptosis” ([83]Figure 3D) showed a
   decrease.

Figure 3.

   [84]Figure 3
   [85]Open in a new tab

   Ingenuity pathways analysis (IPA) for serum proteins altered during
   12-week HIIT

   (A) IPA analysis using proteins that significantly changed in response
   to 12-week HIIT. Z score is the activation score for a pathway, which
   reflects how much the pathway is activated (orange) or deactivated
   (blue) due to the changes in the expression of proteins involved in
   those pathways. -log (B-H p value) refers to Benjamini-Hochberg
   adjusted p value <0.05 .

   (B–D) IPA of diseases and biological functions of significant proteins,
   which are displayed as nodes (proteins) and edges (biological
   relationship between nodes). The color intensity of each node
   represents the direction of change in response to 12-week HIIT, with
   red (upregulated), and blue (downregulated). The edges connecting the
   genes to the respective functions represent the predicted
   relationships, with blue representing inhibition, orange representing
   activation, gray representing effect not predicted, and yellow
   indicating finding inconsistent/contradictory with state of downstream
   molecule, respectively.

Serum proteomic changes associated with changes in metabolic outcomes

   To identify potential mediators for the metabolic benefits of exercise,
   we analyzed the relationship between exercise-induced proteomic and
   metabolic changes using regression analyses adjusted for age, Δ weight,
   and Δ percentage of fat mass. We identified changes in 11 and 21
   proteins that were positively associated and 1 and 2 that were
   negatively associated with Δ fasting insulin and Δ homeostatic model
   assessment for insulin resistance (HOMA-IR), respectively (p < 0.05;
   [86]Figures 4A and 4B). Interestingly, proteomic changes positively
   associated with changes in both outcomes included those of trefoil
   factor family member 2 (TFF2), a protein secreted by the gut and
   involved in metabolism;[87]^21 Fas cell surface death receptor (FAS), a
   pro-apoptotic receptor activated in adipocytes from obese mice;[88]^22
   and protein-arginine deiminase type-2 (PADI2), an enzyme promoting the
   secretion of pro-inflammatory cytokines in macrophages.[89]^23 In
   addition, we identified 9 and 194 proteomic changes that were
   positively and 13 and 4 that were negatively associated with Δ fasting
   glucose (FG) and Δ 2h OGTT glucose (2hG), respectively (p < 0.05;
   [90]Figures 4C and 4D). The proteomic changes positively associated
   with Δ FG included those of canopy FGF signaling regulator 2 (CNPY[2]),
   an angiogenic growth factor expressed in the liver and pancreas and
   modulated by FGF21,[91]^24^,[92]^25 and well-known proteins associated
   with cardiometabolic diseases including plasminogen activator
   inhibitor-1 (PAI-1/SERPINE1) and soluble epoxide hydrolase (sEH/EPHX2).
   Likewise, changes in circulating levels of endothelial adhesion
   molecules (e.g., E-selectin [SELE] and intercellular adhesion molecule
   1 [ICAM-1]) were positively associated with changes in 2hG.

Figure 4.

   [93]Figure 4
   [94]Open in a new tab

   Proteomic changes associated with metabolic outcomes in response to
   12-week HIIT

   (A–D) Changes in serum proteins after 12-week HIIT associated with
   changes in (A) fasting insulin, (B) homeostatic model assessment for
   insulin resistance (HOMA-IR), (C) fasting glucose, and (D) 2h oral
   glucose tolerance test (OGTT) glucose in a linear regression model
   adjusted for age and Δ weight and Δ fat mass. Number of biological
   observations for each graph (n = 36). Orange and blue dots indicate
   proteins that are positively or negatively associated with metabolic
   outcomes (p < 0.05), respectively, whereas gray dots indicate not
   significant proteins (p > 0.05).

Serum baseline proteins associated with changes in metabolic outcomes

   We also identified baseline serum protein levels that predicted
   metabolic responses to 12-week HIIT using regression analyses adjusted
   for age, weight, and percentage of fat mass at baseline. We identified
   1 protein that positively associated with Δ fasting insulin and Δ
   HOMA-IR, while 55 and 54 showed negative association, respectively
   (p < 0.05; [95]Figures 5A and 5B). Baseline proteins that associated
   negatively with changes in fasting insulin included molecules
   implicated in cardiovascular disease (CVD) such as the receptor for
   advanced glycation end products (AGER); tyrosine kinase with
   immunoglobulin-like and EGF-like domains 1 (TIE1), an angiopoietin
   receptor expressed in endothelial cells and associated with a
   pro-atherogenic phenotype;[96]^26 and SELE. In addition, we identified
   9 and 5 proteins that associated positively and 44 and 93 that
   associated negatively with Δ FG and Δ 2hG, respectively (p < 0.05;
   [97]Figures 5C and 5D). Notably, the protein with the strongest
   positive association (in terms of p value) with changes in FG and
   glucose tolerance was pancreatic secretory granule membrane major
   glycoprotein GP2 (GP2), a secreted protein expressed in zymogen
   granules of pancreatic acinar cells as well as plasma membrane of M
   cells in the intestinal epithelium.[98]^27

Figure 5.

   [99]Figure 5
   [100]Open in a new tab

   Baseline proteins associated with exercise-induced metabolic outcomes

   (A–D) Baseline serum proteins associated with changes in (A) fasting
   insulin, (B) HOMA-IR, (C) fasting glucose, and (D) 2h OGTT glucose in a
   linear regression model adjusted for age, weight, and fat mass. Number
   of biological observations for each graph (n = 36). Orange and blue
   dots indicate proteins that are positively or negatively associated
   with metabolic outcomes (p < 0.05), respectively, whereas gray dots
   indicate not significant proteins (p > 0.05).

Exercise induced differential changes of serum proteins in responders and
non-responders

   Participants presented a substantial inter-individual variation in
   terms of fasting insulin and HOMA-IR in response to the 12-week HIIT
   intervention.[101]^11 Therefore, the participants were further
   classified into responders (Rs; n = 28) and non-responders (NR; n = 8)
   depending on whether they could or could not demonstrate a decrease in
   HOMA-IR greater than 2-fold technical error. Responders and
   non-responders did not differ in clinical parameters at baseline and
   displayed a similar reduction in weight and adiposity as well as
   similar increase in cardiorespiratory fitness ([102]Table S3). Notably,
   responders showed a striking 41.51% and 47.18% decrease in fasting
   insulin and HOMA-IR index, respectively, whereas non-responders showed
   a lack of improvement in both outcomes. In parallel, we detected
   differential changes in 23 proteins between responders and
   non-responders (p < 0.05; [103]Figure 6A). Of these, TFF2 and FAS
   exhibited an opposite trend of changes between responder and
   non-responder groups. In addition, pancreatic α-amylase (AMY2A), a
   digestive enzyme involved in the metabolism of starch and
   glycogen,[104]^28 and TIE1 were exclusively upregulated in
   non-responders, whereas their levels did not change in responders.

Figure 6.

   [105]Figure 6
   [106]Open in a new tab

   Differential changes of serum proteins in exercise responders and
   non-responders

   (A) Significantly altered serum proteins (p < 0.05, unpaired Wilcoxon
   rank-sum test) between responders and non-responders. Log[2] FC was
   defined as the ratio between levels after exercise to those at
   baseline.

   (B and C) The receiver operating characteristic (ROC) curves and area
   under curve (AUC) of the proteomics-based algorithm for the
   discrimination between responders and non-responders in (B) discovery
   cohort (n = 19) and (C) validation cohort (n = 29).

   (D) Feature importance contributing to this classifier.

   (E) The differentially abundant proteomic biomarkers (p < 0.1, by
   unpaired Wilcoxon rank-sum test) between responders and non-responders
   at baseline with significance calculated as −log10 (p value). Red:
   enriched in non-responders; blue: enriched in responders.

   Finally, we also investigated whether proteomic differences at baseline
   could be integrated into a ML algorithm to predict individualized
   exercise responsiveness. To this end, a random forest algorithm
   integrating baseline serum proteomic data was built using the discovery
   cohort and achieved an area under the receiver operating characteristic
   curve (AUROC) of 0.87 ([107]Figure 6B). The performance of this model
   was further evaluated in the validation cohort and achieved an AUROC
   value of 0.79 for the discrimination between responders and
   non-responders ([108]Figure 6C). We compared the performance of the
   random forest with other ML models. In this regard, the random forest
   performs best, and the logistic regression model performs slightly
   better than the generalized linear model ([109]Figure S3). One of the
   most informative proteins contributing to this classifier was major
   histocompatibility complex class I polypeptide-related sequence A_B
   (MICA/B) ([110]Figure 6D), a stress-induced ligand for the natural
   killer (NK) group 2D receptor found on cell lineages with NK activity
   as well as gamma delta and alpha beta T cells.[111]^29^,[112]^30
   Notably, lipocalin 2 (LCN2), a secretory glycoprotein associated with
   obesity and insulin resistance,[113]^31 was elevated in responders
   compared with non-responders at baseline (p < 0.05). Conversely, MICA/B
   levels were increased in non-responders compared with responders,
   although it did not reach statistical significance ([114]Figure 6E).

Discussion

   The present study provided characterization of inflammatory and
   cardiometabolic molecules responsive to chronic exercise in a cohort of
   36 medication-naive overweight and obese men with prediabetes. Our
   Olink-based quantitative proteomic profiling revealed dynamic changes
   in hundreds of molecules and signaling pathways during the 12-week HIIT
   intervention, including regulators of metabolism, cardiovascular
   system, inflammation, and apoptosis. Proteins involved in
   gastro-intestinal (GI) mucosal immunity were strongly associated with
   metabolic outcomes of exercise intervention. Particularly,
   exercise-induced changes in TFF2 were positively associated with
   changes in insulin resistance and fasting insulin, whereas baseline
   levels of GP2 were positively associated with changes in FG and 2hG. We
   also observed a highly heterogeneous response to chronic exercise with
   respect to insulin sensitivity, which was accompanied by differential
   changes in 23 serum proteins, including TFF2. Finally, an ML algorithm
   integrating baseline proteins was developed and accurately predicted
   personalized metabolic responsiveness.

   This study examines the molecular effects of chronic exercise in
   individuals with prediabetes using a large-scale quantitative analysis
   of serum proteins. Previous studies performed metabolite profiling in
   response to chronic exercise in healthy young and obese
   individuals.[115]^32^,[116]^33 Likewise, a recent study shed light on
   baseline proteins that may predict exercise responsiveness in terms of
   cardiorespiratory fitness.[117]^34 However, in the latter study, there
   was no assessment of proteins after the exercise intervention, and
   metabolic outcomes were not investigated, highlighting the need for a
   human study that examines proteomic changes in relation to metabolic
   adaptations.

   Numerous proteins changed in response to 12-week HIIT, with the number
   of upregulated proteins being lower than the downregulated ones. This
   may be explained by the functional properties of the included proteins,
   as inflammatory markers are often increased in obese/overweight
   individuals, and ET can counteract these changes.[118]^35^,[119]^36
   Notably, the strongest changes in the upregulated proteins were
   observed in molecules involved in metabolism. GH1, a hormone secreted
   by the anterior lobe of the pituitary gland that stimulates lipolysis
   in adipose tissue and free fatty acid oxidation in the liver,[120]^37
   presented the strongest upregulation, followed by other lipolytic
   factors such as IL-6 and IL-15.[121]^38^,[122]^39 IL-6 is upregulated
   in obese individuals and patients with T2DM.[123]^40 While serum IL-6
   is elevated by acute exercise, both decreased and increased levels of
   circulating IL-6 have been observed in participants with different
   types of chronic ET.[124]^41^,[125]^42^,[126]^43^,[127]^44 Our findings
   are in agreement with a previous report that indicates that vigorous
   chronic exercise, contrary to light and moderate chronic exercise,
   increases circulating levels of IL-6.[128]^44 A potential explanation
   for these findings may be the increased muscle-derived IL-6 with
   high-intensity ET. It should be noted that although IL-6 is considered
   a pro-inflammatory cytokine, muscle-derived IL-6 is associated with
   anti-inflammatory properties.[129]^45^,[130]^46 Furthermore, loss of
   visceral adipose tissue following exercise is dependent on
   IL-6.[131]^38 On the other hand, our results on IL-15 confirm previous
   studies in women indicating that this cytokine increases in response to
   chronic exercise, in particular with resistance
   training.[132]^47^,[133]^48 Notably, lipolytic factors are key
   molecules bridging the anti-inflammatory effects of exercise and
   metabolism since a reduction in visceral adipose tissue decreases the
   secretion of pro-inflammatory cytokines and deleterious
   adipokines.[134]^49 Previous studies have demonstrated a
   pro-inflammatory state in individuals with prediabetes and
   T2DM.[135]^40^,[136]^50^,[137]^51 In this regard, plasma profiling of
   patients with newly diagnosed T2DM using the Olink technology indicated
   increased inflammatory markers (i.e., oncostatin M, IL-16, IL-18, CCL4,
   E-selectin) compared with healthy controls.[138]^40 Conversely, our
   findings demonstrated the decrease of these pro-inflammatory markers as
   well as other cytokines such as IL-1β following 12-week HIIT. Decreased
   levels of IL-1β, which has been associated with the pathogenesis of CVD
   and T2DM,[139]^52^,[140]^53^,[141]^54 were also observed after a
   12-month HIIT program in patients with T2DM.[142]^55 In parallel, we
   observed a marked reduction of pro-apoptotic pathways and molecules
   including BID, CASP3, and DIABLO. Whereas BID is a pro-apoptotic gene
   belonging to the Bcl-2 protein family, DIABLO leads to the activation
   of caspases, which are the effector of apoptosis.[143]^56 Of note,
   apoptosis of adipocytes is a process that links obesity and
   metabolism.[144]^57 In mice fed on a high-fat diet (HFD), the
   inhibition of BID prevented the infiltration of macrophages in adipose
   tissue.[145]^57 Moreover, ET reduced apoptotic and inflammatory markers
   in adipose tissue of obese rats.[146]^58 However, as we cannot
   determine the tissue source of these circulating proteins, further
   research should investigate whether changes in adipose tissue
   remodeling may explain the exercise-induced reduction in circulating
   apoptotic and inflammatory markers.

   Our study also revealed that several appetite stimulating hormones,
   such as GHRL and AGRP, were decreased in the early phase of HIIT but
   returned to baseline by end of the intervention. In contrast,
   appetite-suppressing hormones such as LEP did not return to baseline.
   In plasma samples of patients with T2DM, levels of AGRP and LEP were
   increased, while GHRL was decreased, compared with healthy
   controls.[147]^40 Conflicting results have been reported regarding the
   effect of ET on GHRL, which may depend on the isoform of GHRL analyzed,
   i.e., total, acylated, or non-acylated.[148]^59^,[149]^60 Otherwise,
   accumulating evidence has demonstrated that chronic exercise reduced
   circulating levels of LEP in prediabetic and diabetic
   individuals,[150]^61 which is usually accompanied by improvements in
   LEP sensitivity in the brain.[151]^62 Another well-known hormone that
   exhibits a similar change with LEP in response to 12-week HIIT is MSTN,
   a myokine that inhibits muscle growth and is associated with insulin
   resistance.[152]^63 In agreement with our findings, circulating levels
   of MSTN decreased after resistance training in sedentary men and
   women.[153]^48^,[154]^64 Additionally, a marked reduction of MANF was
   also observed after 12-week exercise intervention. MANF is a secreted
   protein expressed in tissues involved in the neuro-endocrine axis and
   has recently been associated with insulin resistance in patients with
   prediabetes.[155]^65 Our results describe the remarkable changes of
   MANF in response to ET, which warrants further studies to address the
   potential role of this exercise-responsive factor on metabolic
   adaptations.

   To a lesser extent, molecules regulating the cardiovascular system were
   also affected by 12-week HIIT. Red blood cell volume (RBCV) and total
   blood volume (BV) expansion is a fundamental adaptation of chronic
   exercise leading to an increased capacity for oxygen (O[2]) delivery
   and aerobic metabolism.[156]^66 In this regard, hypovolemia is a
   prevalent characteristic of patients with diabetes that may contribute
   to the reduced exercise tolerance observed in this population.[157]^67
   The observation of increased circulating levels of EPO in our study
   suggests that 12-week HIIT stimulated erythropoiesis, plausibly
   resulting in RBCV and BV expansion in individuals with prediabetes. In
   addition, we found increased angiogenesis and angiogenic factors,
   including IL-8, vascular endothelial growth factor D (VEFGD), and
   SMOC2, which may contribute to a higher capillary density in skeletal
   muscle,[158]^68 possibly increasing the peripheral capacity to extract
   O[2] from the circulation, another step in the O[2] cascade that can be
   impaired in our study population.

   Both clinical and animal data indicate that ET may beneficially
   modulate the gut microbiome. For instance, elite endurance athletes are
   characterized by a more diverse microbiome, with high abundance of
   health-associated microbial species such as Akkermansia, and increased
   levels of short chain fatty acids (SCFAs) compared with sedentary
   controls.[159]^69^,[160]^70 Likewise, ET modulates gut microbiota
   composition in mice by increasing the numbers of beneficial microbial
   species.[161]^71 In addition, recent publications indicate that the gut
   microbiota play a key role in mediating the beneficial effects of
   exercise on cardiovascular as well as metabolic adaptations. For
   instance, gut microbiome mediated the beneficial effects of exercise on
   cardiac function following myocardial infarction in mice.[162]^72 On
   the other hand, Veillonella, a lactate-metabolizing microbiome, was
   increased after exercise in two independent cohorts of marathoners and
   Olympic trial rowers. In parallel, mice treated with Veillonella
   increased treadmill run time.[163]^73 Veillonella species metabolize
   lactate into SCFAs, which in turn enhance mitochondria oxidative
   capacity and improve aerobic metabolism.[164]^74 Similarly, we recently
   found that the gut microbiota may play a key role in conferring the
   metabolic benefits of exercise. Exercise-induced changes in gut
   microbiota composition were associated with the metabolic improvements
   after the exercise intervention. In addition, transplantation of fecal
   human microbiota from individuals that had improved insulin sensitivity
   after 12-week HIIT ameliorated glucose intolerance and insulin
   resistance in obese recipient mice.[165]^11 In this line, our
   regression analyses showed that exercise-induced changes in TFF2, a
   protein primarily expressed in neck cells of the stomach and duodenum
   Brunner’ gland and involved in innate and adaptive immunity,[166]^75
   positively correlated with changes in insulin resistance and fasting
   insulin. The main function of TFF2 is to protect the GI mucosa from
   internal and exogenous aggressions of the epithelium.[167]^76
   Considering that the mucosal immune response plays a key role in
   modulating the composition and function of gut microbiota,[168]^77 TFF2
   may partly mediate the metabolic effects of exercise by inducing
   changes in the microbiota. In addition, basal levels of GP2, a protein
   expressed in gut immune cells,[169]^78 were positively associated with
   changes in FG and 2hG, independently of body weight and adiposity.
   Notably, GP2 has been shown to be upregulated at the specific inflamed
   region in the gut.[170]^79 It is reasonable to speculate that
   individuals with higher levels of GP2 at baseline may also exhibit
   pro-inflammatory composition of the gut microbiome. Given that subtle
   differences in microbial signatures may lead to different metabolic
   responses to ET,[171]^11 further studies should investigate the role of
   TFF2 in the exercise-gut microbiota-metabolic response axis as well as
   validate the predictive value of GP2 for glycemic responsiveness to
   exercise.

   Different factors may determine the effectiveness of a lifestyle
   interventions in diabetes prevention including genetics, epigenetics,
   and physiological conditions.[172]^17 Specific genetic patterns
   including variations in metabolic genes such as adiponectin receptor 1
   and muscle expression of genes involved in mitochondrial biogenesis may
   lead to differential responses to exercise.[173]^14^,[174]^80 On the
   other hand, physiological factors such as high basal liver fat and
   visceral fat have been associated with metabolic resistance to
   lifestyle interventions.[175]^81 Notably, we previously demonstrated
   that 12-week HIIT induced a heterogeneous metabolic response associated
   with a maladaptation of the gut microbiota in non-responders.[176]^11
   In addition, our current findings showed that 23 proteins, including
   TFF2, differentially changed between responders and non-responders. The
   increase in TFF2 observed in non-responders might modulate the
   microbiome via regulating genes involved in innate host antimicrobial
   defense,[177]^75 while changes in gut microbiota composition may also
   possibly contribute to increased levels of TFF2. Additional research is
   needed to understand whether TFF2 is associated with the deleterious
   changes observed in the microbiota of non-responders. Furthermore,
   12-week HIIT induced increased levels of AMY2A in non-responders.
   Regarding this finding, acarbose, an oral α-amylase inhibitor that acts
   by slowing down the intestinal absorption of glucose, is used in the
   management of T2DM.[178]^82^,[179]^83 In addition, acarbose improves
   insulin sensitivity in subjects with impaired glucose
   tolerance.[180]^84 Therefore, exercise-induced upregulation of AMY2A in
   non-responders may additionally contribute to the lack of improvement
   in insulin sensitivity. Finally, it is worth to mention that the lack
   of response in a metabolic outcome is not necessarily reflected in
   other variables such as cardiorespiratory fitness (i.e., VO[2max])
   since responders and non-responders displayed similar improvements in
   VO[2max] in our study. Consistently, in the HART-D study, individuals
   with diabetes exhibited improved metabolic parameters following an
   exercise intervention independently of the changes in VO[2max].[181]^85
   Of relevance, VO[2max] response to ET is, to a large extent, determined
   by hematological and cardiac adaptations,[182]^86^,[183]^87 whereas
   skeletal and adipose tissue adaptations are the important contributors
   to metabolic improvements.[184]^9^,[185]^88 Circulating proteins are
   particularly attractive as biomarkers to predict the metabolic response
   to ET. In our protein biomarker analyses, baseline levels of
   circulating proteins accurately discriminated between those individuals
   that benefited from the exercise intervention in terms of metabolic
   adaptations from those who did not. MICA/B was one of the most
   informative proteins to this classifier. Interestingly, MICA/B
   contributes to the detection of damaged, infected, or transformed
   intestinal epithelial cells in a process independent of traditional
   major histocompatibility complex (MHC) class I antigen
   processing.[186]^89 Therefore, these results may further support the
   assumption of a pro-inflammatory composition of the gut microbiome in
   non-responders. Future clinical interventions may apply the present ML
   algorithm to identify exercise non-responders, facilitating that
   alternative strategies can be implemented in those individuals. For
   instance, dietary modifications specifically targeting the gut
   microbiota, as recently described,[187]^90 may maximize the metabolic
   benefits of exercise in these individuals.

   In conclusion, the present study provides a detailed proteomic map of
   serum proteins in response to chronic exercise including potential
   transducers of the metabolic benefits of exercise. These findings,
   together with the recognition of a baseline proteomic signature that
   predicts individualized metabolic responsiveness to ET, may facilitate
   the clinical implementation of personalized exercise interventions for
   diabetes prevention.

Limitations of the study

   Our investigation is limited by a small sample size and restrictive
   inclusion criteria for study participants (only men were included),
   which constrains the generalizability of these results to women and
   other populations. Female participants were not included mainly due to
   the large fluctuations in several sex hormones at this age range (20–60
   years old) and the menstrual period of females, which may influence the
   attendance and compliance to our 12-week HIIT and the interpretation of
   data. In the future, we will certainly expand our pilot study to
   include men and women and other types of exercise to test whether our
   findings can be extrapolated to both sexes. Another limitation refers
   to the confined protein panel (688 cardiometabolic and inflammatory
   proteins), which does not provide a complete coverage of the human
   proteome. Moreover, our proteomic analysis does not include a lean or
   obese control group without exercise, which would have provided
   information on proteins that change naturally within a 12-week period.
   Therefore, future studies in different ethnic groups with large sample
   sizes and inclusion of sedentary subjects are needed to confirm the
   present findings and to explore the potential relevance in clinical
   applications.

STAR★Methods

Key resources table

   REAGENT or RESOURCE SOURCE IDENTIFIER
   Biological samples
     __________________________________________________________________

   Human Blood This paper N/A
     __________________________________________________________________

   Critical commercial assays
     __________________________________________________________________

   Human Insulin ELISA Kit Immunodiagnostics Cat# 31380
   Olink Explore 384 Cardiometabolic &cardiometabolic II Olink N/A
   Olink Explore 384 Inflammation & Inflammation II Olink N/A
     __________________________________________________________________

   Deposited data
     __________________________________________________________________

   Proteomics data [188]Table S2 This paper N/A
     __________________________________________________________________

   Software and algorithms
     __________________________________________________________________

   R Software R Core Team [189]https://www.r-project.org/
   Mfuzz N/A
   [190]https://www.bioconductor.org/packages/release/bioc/html/Mfuzz.html
   (Continued on next PAGE) Ce
   Ingenuity pathway analysis QIAGEN N/A
   igraph N/A
   [191]https://cran.r-project.org/web/packages/igraph/index.html
   [192]Open in a new tab

Resource availability

Lead contact

   Further information and requests for data should be directed to and
   will be fulfilled by the lead contact, Aimin Xu ([193]amxu@hku.hk).

Materials availability

   This study did not generate new unique reagents.

Experimental model and subject details

Ethics statement

   This study was approved by the Institutional Review Board of the
   University of Hong Kong/Hospital Authority Hong Kong West Cluster (UW
   15-370) and endorsed the principles of the Declaration of Helsinki.
   Written informed consents were obtained from all participants.

Study participants

   Overweight and obese Chinese men aged 20 to 60 years were recruited
   from the local community via flyers and online advertisements, as
   previously described.[194]^11 A total of 36 subjects (28 responders and
   8 non-responders) who completed the exercise intervention and had
   samples at different time points were included in the longitudinal
   proteomic profiling. In addition, a total of 48 subjects (34 responders
   and 14 non-responders) who had baseline samples were included in the
   random forest model for predicting exercise responsiveness.

Method details

Participants selection

   The inclusion criteria comprised: (i) non-smoking and Chinese men aged
   20 to 60 years; (ii) overweight or obese as defined by Asian criteria
   (body mass index (BMI) > 23 kg/m^2) and with stable body weight (<5%
   weight change over the last 3 months); (iii) prediabetes, defined by
   impaired glucose tolerance (2-h blood glucose ≥7.8 mmol/L and
   ≤11.0 mmol/L after a 75-gram (g) oral glucose challenge) and/or
   impaired fasting glucose (fasting blood glucose ≥5.6 mmol/L and
   ≤6.9 mmol/L) following the American Diabetes Association
   guidelines;[195]^91 (iv) absence of any systemic, metabolic and
   cardiovascular diseases (CVD), as well as any infection within one
   month prior to recruitment; and (v) not on any medication or specific
   diet. Exclusion criteria comprised: (i) acute illness or current
   evidence of acute or chronic inflammatory or infectious diseases; (ii)
   any neurological, musculoskeletal or cardio-respiratory condition that
   would entail risk during exercise or preclude the participant from
   adapting to an exercise program; (iii) specific diet program and/or
   participation in regular exercise for more than 2 times per week in the
   latest 3 months prior to recruitment; and (iv) mental illness rendering
   them unable to carry out exercise activities or understand the nature,
   scope, and possible consequences of the study.

HIIT protocol

   The 12-week HIIT consisted of three sessions per week on
   non-consecutive days at the Active Health Clinic, Center for Sports and
   Exercise, The University of Hong Kong, supervised by certified exercise
   specialists in a one-to-one manner, as previously described.[196]^11
   Compliance in the exercise session was highly encouraged and
   participants were required to take part in at least 85% of all the
   exercise sessions for inclusion into the study. The 70-min (min)
   high-intensity aerobic and resistance interval training sessions
   consisted of a 10 min warm-up, followed by the participants being
   rotated through three high-intensity interval stations involving
   treadmill, resistance/calisthenics, and stationary bike exercises, each
   lasting 10 min with 3–4 min recovery between stations. Each training
   session concluded with 10–15 min of cool-down and stretching exercises.
   Participants wore a wireless heart rate telemetry sensor (Polar H7
   heart rate sensor, Polar Electro Oy, Kempele, Finland) throughout the
   exercise sessions to monitor their heart rate (HR) and to ensure that
   they were working at the appropriate intensity level. The intensity was
   adjusted according to individual participant’s real-time HR telemetry
   and they were encouraged to work at 80-95% of maximal heart rate
   (HR[max]). The treadmill interval station consisted of 3–4 bouts of
   2 min running at 85-95% maximal aerobic capacity (VO[2max]) separated
   by 30–45 s (s) periods of active recovery at 50% VO[2max], during which
   speed and incline were adjusted according to the participant’s fitness
   level. The stationary bike station comprised 4-5 bouts of 45–60 s
   cycling efforts at 90-95% peak power output (W[peak]) interspersed with
   60–75 s active recovery at 30% W[peak], during which resistance and
   cadence were adjusted according to the participant’s fitness level. The
   station of resistance and calisthenics exercises comprised 2-3 sets of
   several types of high-intensity exercises such as squats, kettlebell
   swings, planks and burpees, and resistance with 30 s rest between each
   set. The intensity and work/rest ratio of the resistance and
   calisthenics station progressed weekly over the 12-week training period
   in order to keep the exercises challenging and maintain appropriate
   high intensity interval training stimuli. Aside from ET, all recruited
   participants were instructed to continue their normal routine with
   regard to their physical activity and diet.

Collection of clinical data

   Fasting serum samples were collected at baseline, 4 and 12 weeks after
   the exercise program and were conducted at 48-72 h after the final
   exercise session to prevent the influence of the acute effects of
   exercise. 75-g oral glucose tolerance test (OGTT) was conducted after
   10-12 h overnight fasting. Blood samples were collected for the
   determination of plasma glucose and insulin levels at 0, 60 and 120 min
   after taking the 75 g glucose solution (TRELAN-G75, Ajinomoto
   Pharmaceutical Co. Ltd., Tokyo, Japan). Glucose, serum lipid profiles,
   including triglycerides (TG), total cholesterol (TC), high-density
   lipoprotein (HDL) cholesterol, and low-density lipoprotein (LDL)
   cholesterol, were determined with standard laboratory techniques on a
   Hitachi 717 Analyzer (Roche Diagnostics, Germany). Serum insulin levels
   were measured with ELISA kits (Immunodiagnostics, Hong Kong, China).
   Insulin resistance index (HOMA-IR) (Wallace et al., 2004) was
   calculated using homeostasis model assessment methods, defined as
   fasting insulin (μU/mL) x fasting glucose (mmol/L)/22.5. Body
   composition was assessed via whole-body dual-energy X-ray
   absorptiometry (DXA) scans (Explorer S/N 91,075, Hologic Inc., Waltham,
   MA, USA). Strength and cardiovascular physical assessments were
   conducted at baseline and within one week after completion of the
   12-week intervention. Maximal voluntary muscle strength (chest press
   and leg press) was assessed using the 1-RM method (Keiser A-300, Keiser
   Corp., Fresno, CA, USA) following the manufacturer’s instructions. An
   integrative cardiorespiratory fitness test using the Balke treadmill
   protocol was performed on a motor driven, electronically controlled
   treadmill (TrackMaster, Full Vision Inc., Newton, KS, USA) to assess
   VO[2max] before and after the exercise intervention.

Definition of responders and non-responders

   Non-response to the exercise intervention was defined as a failure to
   demonstrate a decrease in HOMA- IR greater than the 2-fold technical
   error.[197]^11

Olink Proteomics

   Antibody-based technology (Olink Proteomics AB, Uppsala, Sweden) was
   used for proteomics profiling in serum samples. Specifically, protein
   biomarkers were detected via the Olink Explore 384 cardiometabolic and
   inflammation panels. Briefly, the proximity extension assay (PEA)
   technology uses DNA oligonucleotide-labelled antibody pairs to bind
   target proteins. After two matched antibodies bind the target protein,
   the oligonucleotide pairs hybridize and are extended by DNA polymerase
   to create a unique DNA barcode that is subsequently readout using
   next-generation sequencing. Since only correctly matched DNA string
   pairs could generate detectable and quantifiable signals, the PEA
   technology exhibited high specificity and excellent
   sensitivity.[198]^92 After excluding 46 proteins that were not detected
   in >50% of the samples and 3 proteins that were duplicated, the final
   analysis included 688 proteins (352 from the cardiometabolic panel, 333
   from the inflammatory panel and 3 included in both panels). The median
   intra-assay coefficient of variability (CV) was 12.5%, as assessed by
   multiple replicates of a pooled sample included in the experiment.
   Values were presented as normalized protein expression (NPX) units on a
   log2 scale.

Mesoscale Discovery (MSD)

   We subsequently performed additional proteomics profiling using an MSD
   technology (MesoScale Discovery, Rockville, MD, USA) in 19 individuals
   from the study cohort to determine the reproducibility of our
   olink-based results. MSD procedure follows that of a sandwich ELISA,
   with any analytes of interest captured on the electrode being detected
   with an analyte specific ruthenium-conjugated secondary antibody.
   Values were presented as pg/ml on a log2 scale.

Quantification and statistical analysis

   Main statistical analyses were performed with R software. Participants’
   baseline characteristics and outcomes were expressed as mean ± SEM.
   Data were tested for normal distribution with the Kolmogorov-Smirnov
   test and for homogeneity of variances with the Levene’s test. Change
   from baseline was evaluated by paired Student’s t test. Difference
   between groups at baseline and after intervention was evaluated by
   independent Student’s t test and ANCOVA model controlling for the
   baseline measurements, respectively. The proteomics data were not
   normally distributed. The paired Wilcoxon rank-sum test was used to
   compare protein levels at 4- and 12-week vs baseline. The
   Benjamini-Hochberg false discovery rate (FDR) method, with a 5% FDR
   significance threshold, was applied to control for multiple testing.
   Fuzzy c-mean clustering was performed using the R package ‘Mfuzz’
   (v2.20.0) after log2-transformation and Z score scaling of the data.
   Ingenuity pathway analysis (IPA, QIAGEN, Hilden, Germany) platform was
   performed to search for enriched pathways using significant proteins. p
   values were corrected for multiple hypothesis using the
   Benjamini-Hochberg method and pathways with FDR below 0.05 were
   considered significant. Linear regression analyses were performed to
   determine the relationship between changes in metabolic parameters and
   (i) baseline levels of proteins and (ii) intervention-induced changes
   in proteins (p < 0.05). Covariates in regression models included age,
   baseline or changes in body weight and body fat percentage. Differences
   in exercise-induced proteins changes between metabolic responder and
   non-responders were evaluated with unpaired Wilcoxon rank-sum test
   (p < 0.05).

ML algorithm

   The proteomic profiles at baseline of 48 subjects were used to build an
   ML algorithm for predicting exercise responsiveness, with R caret
   package. Three ML models including (1) generalized linear regression,
   (2) logistic regression and (3) random forest were constructed using
   the baseline proteomics data. Samples were divided in two cohorts as
   performed in our previous manuscript.[199]^11 The discovery cohort
   included 14 responders and 5 non-responders recruited between September
   2016 and November 2016, whereas the validation cohort included 20
   responders and 9 non-responders that shared similar characteristics
   with the discovery cohort, but were recruited between October 2018 and
   December 2018. A total of 32 differentially abundant proteomics
   biomarkers between responders and non-responders at baseline within the
   discovery cohort (n = 19, p < 0.1, Wilcoxon rank-sum test) were used
   for model construction, with 5-repeated 10-fold cross validation and
   ROSE sampling strategy to account for the imbalance in the two classes.
   The generalization of the random forest model was further tested in the
   validation cohort (n = 29). The area under the receiver operating
   characteristic curve (AUROC) was used as the main indicator of model
   performance.

Additional resources

   Complete clinical trial registration was deposited at
   [200]ClinicalTrials.gov ([201]NCT03240978), as previously
   described.[202]^11

Acknowledgments