Reversal of proteomic aging with exercise—results from the UK biobank and a 12-week intervention study

Abstract

Biological aging varies between individuals and may be influenced by health behaviors. Using data from 45,438 UK Biobank participants, we found that a higher proteomic aging score (ProtAgeGap) was linked to lower physical activity and increased risk of type 2 diabetes. The UK Biobank cohort included both men and women. In a 12-week supervised exercise study (MyoGlu) in 26 men, ProtAgeGap decreased by the equivalent of 10 months. While most of the 204 proteins in the score remained stable, some, like CLEC14A, changed with exercise and were linked to improved insulin sensitivity. Transcriptomic data from muscle and fat tissue supported these protein-level changes, highlighting pathways, such as PI3K-Akt and MAPk signaling, involved in tissue remodeling and metabolism. Our findings suggest that while proteomic aging is mostly stable, it can be modestly reversed by exercise. Specific proteins within the signature may act as sensitive indicators of metabolic adaptation, supporting the idea that proteomic aging is a modifiable marker linked to lifestyle and disease risk. Clinical trial number: clinicaltrials.gov: NCT01803568 registered 2013-02-26.

Introduction

Chronological age is a strong predictor of chronic disease risk¹, although individuals of the same age can differ markedly in their physiological function and health status². This has led to growing interest in biological aging - the cumulative decline in system integrity more accurately reflecting individual risk of disease, disability, and mortality^3,4. Unlike chronological age, biological age aims to capture the true functional state of the body, making it a promising concept to increase our understanding of disease risk and serve as a potential tool for risk stratification and personalized prevention⁴.

Several molecular biomarkers have been proposed to estimate biological age, including epigenetic clocks⁴, such as one developed by our lab (eABEC)⁵, telomere length⁶, and inflammaging scores⁷. Whereas these tools have shown associations with morbidity and mortality^8–10, they also have limitations, including tissue specificity, limited responsiveness to interventions, and uncertainty regarding what biological processes they capture¹¹. Recently, proteomic-based age predictors have emerged as an alternative approach, offering a systemic, blood-based readout of biological age reflecting diverse physiological pathways^10,12,13. One such marker is ProtAge¹³, a proteomic aging clock derived from circulating protein profiles. The difference between an individual’s ProtAge and chronological age, referred to as ProtAgeGap, represents accelerated or decelerated biological age. ProtAgeGap has been linked to all-cause mortality and cardiometabolic outcomes, but its responsiveness to lifestyle factors such as physical activity remains unclear.

Given the well-established benefits of exercise on metabolic health and longevity¹⁴, we hypothesized that physical activity may reduce biological aging, as reflected by a lower ProtAgeGap. To test this, we first examined observational associations between ProtAgeGap, physical activity, and type 2 diabetes risk in the UK Biobank. We then evaluated whether a structured 12-week exercise intervention could reduce proteomic age in an independent cohort. Finally, we integrated transcriptomic data from skeletal muscle and adipose tissue to explore how tissue responses relate to circulating proteins and biological aging. Together, these analyses aimed to shed light on the interplay between exercise, biological age, and metabolic health.

Results

ProtAgeGap, physical activity, and metabolic markers in the UK biobank

In the UK Biobank cohort participants with available plasma proteomics (n = 45,438) (Supplementary Table S1), higher ProtAgeGap, reflecting older proteomic age relative to chronological age, was significantly associated with a range of physical activity measures, body composition, and metabolic traits (Fig. 1). Across physical activity measures, higher levels of total activity, number of activity days, and MET-minutes per week, were consistently associated with lower ProtAgeGap (standardized β range: −0.02 to −0.06, FDR < 0.01), suggesting a lower rate of proteomic aging in more physically active individuals (Fig. 1). Similarly, meeting physical activity recommendations and belonging to a higher IPAQ activity group were associated with significantly lower ProtAgeGap (Fig. 1). In contrast, ProtAgeGap showed positive associations with all measures of adiposity, including body fat percentage (β ≈ 0.04) and whole-body fat mass (β ≈ 0.06) (Fig. 1), as well as with cardiometabolic biomarkers such as glucose, HbA1c, and triglycerides (β ≈ 0.02 to 0.08, all FDR < 0.0001) (Fig. 1).

Fig. 1. Associations between ProtAgeGap and physical activity, body composition, and metabolic biomarkers in the UK Biobank cohort (n = 45,438).

The forest plot shows standardized beta coefficients and 95% confidence intervals for each association, derived from linear regression models adjusted for age, sex, ethnicity, recruitment centre, smoking status, alcohol intake, alcohol intake frequency, household income, and college education. Outcomes are grouped into physical activity metrics (top), body composition (middle), and metabolic health biomarkers (bottom). The size of each point corresponds to the false discovery rate (FDR)-adjusted p-value, with larger and filled circles indicating stronger statistical significance (FDR < 0.0001 to > 0.05). Negative beta values indicate that higher levels of the outcome are associated with a lower ProtAgeGap (i.e., lower proteomic age relative to chronological age), whereas positive values indicate associations with older proteomic age. IPAQ = International Physical Activity Questionnaire. MET = Metabolic equivalent of task.

ProtAgeGap, physical activity and type 2 diabetes in the UK biobank

In a Cox proportional hazards model assessing incident type 2 diabetes, ProtAgeGap was significantly associated with incidence of type 2 diabetes (HR: 1.06; 95% CI: 1.04–1.09, p = 1.76e⁻¹²). However, in a second Cox model adding an interaction term between ProtAgeGap and physical activity (IPAQ activity group), we found that while ProtAgeGap was no longer significantly associated with diabetes risk overall (HR = 1.02, 95% CI: 0.99–1.06, p = 0.11), there was a significant interaction between ProtAgeGap and the high physical activity group (HR = 1.05, 95% CI: 1.01–1.09, p = 0.02), indicating that the effect of biological age acceleration on diabetes risk was modified by activity level (Table 1).

Table 1.

Physical activity modifies the association between biological age (ProtAgeGap) and type 2 diabetes incidence in the UK biobank

	HR[95%CI]	z	p
ProtAgeGap	1.02[0.99,1.06]	1.61	0.11
IPAQ moderate activity group	0.73[0.65,0.81]	−5.78	<0.005
IPAQ high activity group	0.61[0.55,0.69]	−8.62	<0.005
Age	1.06[1.06,1.07]	19.59	<0.005
Sex	1.78[1.63,1.95]	12.98	<0.005
ProtAgeGap x IPAQ moderate activity group	1.02[0.99,1.06]	1.21	0.23
ProtAgeGap x IPAQ high activity group	1.05[1.01,1.09]	2.27	0.02

This table shows results from a Cox proportional hazards model examining the risk of type 2 diabetes diagnosis in relation to biological age acceleration (ProtAgeGap), physical activity level (IPAQ activity group), and their interaction, adjusted for age and sex. Hazard ratios (HR) with 95% confidence intervals (CI) are presented. A significant interaction between ProtAgeGap and physical activity suggests that the effect of biological age acceleration on diabetes risk differs by physical activity level. IPAQ low activity group is the reference.

Physical activity showed strong independent protective associations: participants in the moderate activity IPAQ group had a 27% lower hazard of developing type 2 diabetes (HR = 0.73, 95% CI: 0.65–0.81, p < 0.005), whereas those in the high activity IPAQ group had a 39% lower hazard (HR = 0.61, 95% CI: 0.55–0.69, p < 0.005), compared to the low activity IPAQ group (Table 1). Age and sex were also significant predictors (HR = 1.06 and 1.78, respectively, both p < 0.005) (Table 1).

In another Cox proportional hazards model evaluating risk of ischemic heart disease (IHD), we found significant independent associations of ProtAgeGap and physical activity with IHD, but no evidence of an interaction. Specifically, higher ProtAgeGap was associated with increased risk of IHD (HR = 1.04, 95% CI: 1.02–1.06, p < 0.005). Physical activity was associated with reduced risk of IHD; moderate activity IPAQ group had a 14% lower hazard of IHD (HR = 0.86, 95% CI: 0.80–0.92, p < 0.005), whereas those in the high activity IPAQ group had a 15% lower hazard (HR = 0.85, 95% CI: 0.79–0.91, p < 0.005) as compared to the low activity IPAQ group. We observed no interaction between ProtAgeGap and physical activity (e.g. ProtAgeGap × high activity IPAQ group: HR = 1.00, 95% CI: 0.98–1.02, p = 0.90).

Effect of a 12-week exercise intervention on ProtAgeGap

The MyoGlu cohort has been described in detail previously^15–17 but main subject characteristics are presented in Supplementary Table S2. Briefly, following the 12-week intervention, both groups showed reduction in body fat mass (2–4 kg) and an increase in lean mass ( ~ 2 kg), accompanied by significantly improved insulin sensitivity ( ~ 40%), maximum oxygen uptake, and muscle strength (Supplementary Table S2).

ProtAge was significantly correlated with age in the MyoGlu participants both before (untrained state) and after (trained state) the exercise intervention (Fig. 2a, b). The difference between ProtAge and age–representing ProtAgeGap, our marker of biological age acceleration–was similar between overweight and normal weight men before intervention (Fig. 2c) and showed a significant reduction after 12 weeks of exercise (Fig. 2d). Three months of exercise reduced ProtAgeGap by 10.2 months. Since mean ProtAge was 58.5 years in the whole cohort at baseline, this reduction equals to 1.5% reduced proteomic age. The response in ProtAgeGap to exercise was similar for overweight and normal weight men (interaction effect p > 0.05) and remained significant after adjustment for age and BMI (Supplementary Table S3).

Fig. 2. ProtAgeGap in response to 12 weeks of strength and endurance exercise.

a Spearman’s correlation between ProtAge and chronological age before intervention, and b after intervention in overweight (OW) and normal weight (NW) men. c ProtAgeGap in OW (median(interquartile range(IQR)); 0.46(2.70), n = 13) vs NW (median(IQR) = −0.58(3.23), n = 13) men (ns = not significant). d Reduced ProtAgeGap after 12 weeks of exercise across all men (untrained; median(IQR) = 0.44(3.13), n = 26, trained; median(IQR) = −0.39(3.19), n = 26).

Exercise-responses in the 204 proteins used to construct ProtAge and associations with metabolic traits

Of the 204 individual proteins in the ProtAge signature, only 8 proteins showed significant changes following 12 weeks of exercise training after correction for multiple testing (Fig. 3a). Proteins such as CLEC14A and PRND displayed the strongest responses. Hierarchical clustering of Spearman correlations between changes in serum protein levels and metabolic traits after 12 weeks of exercise (Fig. 3b) revealed positive correlations between several proteins and improvements in insulin sensitivity (ΔGIR), ΔVO₂max, and Δfat-free mass, whereas negative correlations were noted with reductions in fat mass, insulin levels, and liver fat content. Notably, changes in CLEC14A levels were positively associated with ΔGIR (Fig. 3c), with similar trends observed in both overweight (red dots) and normal-weight (black dots) participants.

Fig. 3. Individual ProtAge proteins and exercise.

a A volcano plot showing responses to 12-weeks of exercise for each of the 204 proteins used to construct ProtAge. b The heatmap shows hierarchical clustering of exercise responses in metabolic traits and serum proteins (columns). Clustering was performed using Spearman correlation as the distance metric and complete linkage. The dendrograms indicate similarity between metabolic traits (rows) or proteins (columns), where traits/proteins with similar expression profiles are grouped closer together. The color gradient in the heatmap reflects Spearman correlation values, where blue indicates negative correlation, white is near zero, and red indicates positive correlation between traits and proteins. c A scatter plot of change in serum CLEC14A protein levels and the change in insulin sensitivity in response to the 12-week intervention. Red dots = overweight men; black dots = normal weight men GIR: Glucose infusion rate during hyperinsulinaemic glucose clamp; NPX: normalized protein expression.

Exercise-responses in the 204 proteins used to construct ProtAge and associations with changes in skeletal muscle mRNA

Hierarchical clustering indicated some associations between exercise responses in muscle mRNA and serum proteins (Fig. 4). Fig. 4a shows a heatmap of Spearman correlations between exercise-induced changes in serum proteins (columns) and muscle mRNA expression (rows). The clustering algorithm identified four major clusters of mRNA and protein responses, as indicated by the dendrograms and color bars. Notably, a distinct protein cluster (cluster 4, in green) showed strong correlations with two mRNA clusters; cluster 2 (blue) and cluster 3 (black); suggesting shared biological patterns in response to exercise. mRNA cluster 2 was enriched for pathways e.g. PI3K-Akt signaling and regulation of actin cytoskeleton (Fig. 4b), whereas mRNA cluster 3 was enriched for e.g. MAPk signalling and stress responses (Fig. 4c). These pathways are known to play a role in cellular remodeling, and metabolic adaptation during exercise. The top 10 serum proteins from protein cluster 4 that had the highest Spearman correlations with genes in mRNA clusters 2 and 3, are shown in Fig. 4d, e. Proteins such as TNFRSF11B, DLK1, and AHNAK2 show positive correlations with cluster 2 genes (D), whereas others like HMCN2, RET, and CHRD-like proteins, are negatively correlated with genes in mRNA cluster 3 (E).

Fig. 4. Exercise-responses in individual ProtAge proteins and relationships with exercise-responses in muscle mRNAs.

a The heatmap shows hierarchical clustering of muscle mRNA expression data (rows) and serum proteins (columns), with highlighted clusters in different colors. Clustering was performed using Spearman correlation as the distance metric, and complete linkage. The dendrograms indicate similarity between genes (rows) or proteins (columns), where genes/proteins with similar expression profiles are grouped closer together. The color gradient in the heatmap reflects Spearman correlation values, where blue indicates negative correlation, white is near zero, and red indicates positive correlation between genes and proteins. The bubble plots display results from pathway enrichment analyses of genes in clusters 2 (b) and 3 (c). Bubble size represents gene set size, and color indicates statistical significance (e.g. –log₁₀(p-value)). The barplots shows the top 10 proteins from protein cluster 4 with the highest mean Spearman correlation to genes in the mRNA clusters 2 (d) and 3 (e). These proteins may reflect downstream or co-regulated processes associated with the enriched pathways.

Exercise-responses in the 204 proteins used to construct ProtAge and associations with changes in adipose tissue mRNA

Similar to muscle, hierarchical clustering of adipose tissue mRNA expression and serum protein levels also revealed some patterns of exercise-induced responses (Fig. 5a). Protein cluster 2 (blue) showed some associations with two mRNA clusters (clusters 1 (red) and 2 (blue)) based on Spearman correlation (Fig. 5a). mRNA cluster 1 exhibited strong positive correlations with proteins in cluster 2, while mRNA cluster 2 showed inverse correlations with the same protein cluster. Pathway enrichment analysis of mRNA cluster 1 (Fig. 5b) revealed overrepresentation of metabolic processes including branched-chained amino acids catabolism, pyruvate metabolism, and the TCA cycle, whereas mRNA cluster 2 (Fig. 5c) was enriched for pathways involved in MAPK signaling, autophagy and Wnt signaling. The barplots (Fig. 5d, e) highlight the top 10 proteins in protein cluster 2 that were correlated with genes in mRNA cluster 1 and 2.

Fig. 5. Exercise-responses in individual ProtAge proteins and relationships with exercise-responses in adipose tissue mRNAs.

a The heatmap shows hierarchical clustering of adipose tissue mRNA expression data (rows) and serum proteins (columns), with highlighted clusters in different colors. Clustering was performed using Spearman correlation as the distance metric, and complete linkage. The dendrograms indicate similarity between genes (rows) or proteins (columns), where genes/proteins with similar expression profiles are grouped closer together. The color gradient in the heatmap reflects Spearman correlation values, where blue indicates negative correlation, white is near zero, and red indicates positive correlation between genes and proteins. The bubble plots display results from pathway enrichment analysis of genes in clusters 1 (b) and 2 (c). Bubble size represents gene set size, and color indicates statistical significance (e.g. –log₁₀(p-value)). The barplots show the top 10 proteins from protein cluster 2 with the highest mean Spearman correlating to genes in the mRNA clusters 1 (d) and 2 (e). These proteins may reflect downstream or co-regulated processes associated with the enriched pathways.

Discussion

Using population-based observational analyses in the UK Biobank, we found that ProtAgeGap, a marker of accelerated biological aging¹³, was negatively associated with vigorous physical activity and positively associated with incidence of type 2 diabetes. Furthermore, we found that physical activity moderated the association between ProtAgeGap and incidence of type 2 diabetes. These findings generated the hypothesis that physical activity may directly reduce proteomic aging, thereby exerting a beneficial effect on glucose metabolism. We tested this hypothesis in a controlled 12-week exercise intervention and observed an approximate 10-month reduction in proteomic age post-intervention¹⁸, along with correlations between specific proteins and improvements in insulin sensitivity, as measured by the hyperinsulinemic-euglycemic clamp^15,16. Since there is an ongoing debate about what different measures of biological aging actually reflect¹¹, we conducted exploratory analyses to identify specific proteins that capture the effects of exercise on proteomic age and to explore what they might reflect biologically using mRNA expression data from both skeletal muscle and adipose tissue.

The associations observed in the UK biobank suggest that both proteomic age and physical activity are important contributors to type 2 diabetes risk¹³, with a notable interaction between the two. Whereas high levels of (in particular vigorous) physical activity were generally protective, individuals with high ProtAgeGap, i.e., those who are biologically older than their chronological age, remained at increased risk for type 2 diabetes, even when physically active. This observation implies that accelerated proteomic aging may attenuate some of the metabolic benefits typically conferred by exercise. This highlights the potential importance of incorporating biological age into personalized prevention strategies and may suggest that individuals with high ProtAgeGap may require more intensive monitoring or intervention, even if they maintain a physically active lifestyle. Furthermore, our results indicated that while walking was associated with an accelerated ProtAgeGap, moderate-to-vigorous physical activity was linked to a reduction in this measure. Given that walking alone is unlikely to substantially reduce body fat or improve cardiorespiratory fitness, moderate-to-vigorous exercise may be required to promote healthy biological ageing. Light activities such as slow walking might instead reflect lower physical fitness or compensatory behavior in less healthy individuals, whereas moderate-to-vigorous exercise seem to support longevity and metabolic resilience.

In our controlled 12-week exercise intervention, we found that while the overall ProtAgeGap score was only modestly reduced ( ~ 10 months ( ~ 1.5%) reduced proteomic age after 3 months of intense exercise) specific proteins within the signature responded strongly to exercise. This suggests that although proteomic aging as a composite marker is relatively stable, as is necessary for a stable marker of aging^4,5,8, it may include more exercise-responsive subcomponents. Proteins such as CLEC14A showed a significant response to exercise and an association with improved insulin sensitivity. Identification of these responsive proteins offers an opportunity to move beyond composite scores and focus on targeted proteomic biomarkers that may be more sensitive to lifestyle interventions¹⁸. This has implications for how we interpret biological age in interventional studies; rather than expecting general reversal of aging markers, we may find greater value in tracking key proteins linked to metabolic plasticity^18,19.

Despite the overall stability of the proteomic aging signature, clustering analyses suggested some shifts in subsets of proteins that aligned with improved body composition and insulin sensitivity after exercise. This suggests that proteomic signatures can reflect individual variability in response to exercise, as exemplified by CLEC14A²⁰. Interestingly, CLEC14A is an endothelial cell-related protein involved in angiogenesis, which is tightly linked to improved insulin sensitivity after exercise¹⁸.

Notably, several individual proteins linked to ProtAgeGap—such as RGMB, CD93, FAP and PRND—have recognized roles in oxidative stress, neurodegenerative processes, and tissue regeneration and are plausible contributors to aging-related changes^21–23. In skeletal muscle, gene-expression changes after training related to ProtAgeGap were enriched for oxidative metabolism, the tricarboxylic acid (TCA) cycle, fatty acid degradation, and pyruvate metabolism, consistent with enhanced mitochondrial biogenesis and metabolic flexibility^24,25. In adipose tissue, we observed correlations with transcripts involved in MAPK, PI3K–Akt, Wnt, and autophagy/mitophagy pathways. Taken together, ProtAgeGap and its modulation by exercise appear to converge on several hallmarks of aging²⁶, including mitochondrial dysfunction, impaired cellular turnover, stress resilience, and insulin signaling, suggesting that exercise promotes systemic rejuvenation of cellular homeostasis.

Our findings also support the view that biological aging is not a fixed trajectory but may be modifiable by targeted exercise intervention. The ability to capture such adaptations at the proteomic level may offer potential for tailoring interventions based on biological age rather than chronological age.

A key strength of this study is the replication of our findings between two independent observational and intervention-based cohorts, supporting the generalizability of the associations between ProtAgeGap, physical activity, and type 2 diabetes. However, several limitations should be acknowledged. First, the observational nature of the UK Biobank analysis limits causal inference, and whereas the intervention study provides mechanistic insight, the sample size was relatively small. Second, the interpretation of proteomic aging is still evolving, and it remains unclear how well ProtAgeGap reflects aging processes across different tissues²⁷. Moreover, the transcriptomic analyses were limited to muscle and adipose tissue and may not capture broader systemic effects. Future research should explore these relationships in larger and more diverse populations, include other lifestyle or pharmacological interventions, and assess long-term changes in proteomic aging and metabolic outcomes. Expanding the focus to additional tissues and validating protein biomarkers over time, may help refining the concept of proteomic age as a tool for risk stratification and personalized prevention strategies. Also, the MyoGlu trial was a within-subject pre–post intervention design with sedentary men serving as their own controls. This design was chosen to maximize statistical power for molecular changes within a limited sample size, but the absence of a separate non-training control group limits causal inference. However, UK Biobank analyses, which include physically inactive participants, provide complementary external comparison supporting the interpretation of exercise-related changes. Lastly, the MyoGlu trial was designed to minimize hormonal and sex-specific variability in muscle biopsies by including only men, and for safety reasons participants above 65 years were excluded. We acknowledge that this selection may bias the results toward male, middle-aged responses to exercise, thereby limiting generalizability to women and older adults.

In conclusion, our findings suggest that proteomic aging, measured by ProtAgeGap, is influenced by physical activity and associated with type 2 diabetes risk. Exercise may partially reverse proteomic aging, with specific proteins reflecting metabolic improvements.

Methods

UK Biobank

The UK biobank is a large prospective population-based cohort including ~500,000 individuals ( ~ 273,000 women), with a variety of phenotypic and genetic data available²⁸. The UK Biobank has ethical approval from the North West Multi-Centre Research Ethics Committee (MREC), which covers the UK, and all participants provided written informed consent. We used data from individuals with both Olink Explore 3072 and physical activity data, totalling 45,438 individuals.

Codes and data used to define prevalent and incident type 2 diabetes and ischemic heart disease (IHD) in the UKB are detailed in Supplementary Table 20 from our previous publication¹³. Diagnoses and date of first diagnosis for all diseases in the UKB were ascertained using ICD diagnosis codes and corresponding dates of diagnosis taken from linked hospital inpatient, primary care and death register data. If a participant did not have a diagnosis code in hospital inpatient or primary care records, but the code was listed as a primary or secondary cause of death, then they were coded as a case with the date of diagnosis as the date of death. Primary care read codes were converted to corresponding ICD diagnosis codes using the lookup table provided by the UKB. Linked hospital inpatient, primary care and cancer register data were accessed from the UKB data portal on 22 February 2024, with a censoring date of 31 October 2022; 31 August 2022 or 31 May 2022 for participants recruited in England, Scotland or Wales, respectively (8–16 years of follow-up). Detailed information about the linkage procedure national registries for mortality and cause of death information in the UKB is available online (https://biobank.ctsu.ox.ac.uk/crystal/refer.cgi?id=115559). Mortality data were accessed from the UKB data portal on 22 February 2024, with a censoring date of 30 November 2022 for all participants (12–16 years of follow-up).

MyoGlu

The MyoGlu study (Clinical trial number: clinicaltrials.gov: NCT01803568 registered 2013-02-26) was a controlled clinical trial conducted in accordance with the Declaration of Helsinki and approved by the Regional Committee for Medical and Health Research Ethics North (ref. 2011/882, Tromsø, Norway). All participants provided written informed consent. The study included sedentary men aged 40–65 years, engaging in fewer than one exercise session per week over the past year. Participants were divided into two groups based on body mass index (BMI) and glucose tolerance: overweight men (BMI 29.5 ± 2.3 kg/m²) with reduced glucose tolerance and/or insulin sensitivity, and normal-weight controls (BMI 23.6 ± 2.0 kg/m²). Both groups (n = 13 per group) completed a 12-week combined strength and endurance training program, including two 60-minute cycling sessions and two 60-minute whole-body strength training sessions per week, as previously described in detail^15–17. The MyoGlu trial was originally designed to discover cytokines (“myokines”) important for glucose metabolism. The use of the novel ProtAgeGap biomarker is an additional analysis based on the large-scale serum proteomics part of MyoGlu.

Serum protein profiling was performed using antibody-based proximity extension assay (PEA) technology (Olink Proteomics AB, Uppsala, Sweden) with the Olink Explore 3072 panel. Pairs of oligonucleotide-labeled antibodies bind to target proteins; upon dual binding, the oligonucleotides hybridize and are extended by DNA polymerase to generate a unique DNA barcode. This barcode is then quantified via next-generation sequencing. The PEA method offers high specificity and sensitivity, as signal generation requires precise dual antibody binding. Protein expression (NPX) values were normalized using median scaling and log₂-transformed prior to analyses, as recommended by the manufacturer (https://olink.com/knowledge/documents).

An euglycemic-hyperinsulinemic clamp was performed after an overnight fast using a fixed insulin infusion rate of 40 mU/m²·min for 150 min. Glucose (200 mg/mL) was infused to maintain blood glucose at 5.0 mmol/L. Insulin sensitivity was calculated as the glucose infusion rate (GIR) during the final 30 min, normalized to body weight and fat-free mass. Whole blood glucose was measured using a glucose oxidase method (YSI 2300, Yellow Springs, OH), and plasma glucose was estimated as whole blood glucose × 1.119.

Body composition was assessed using MRI/MRS with a 3D DIXON protocol (ankle-to-neck coverage), enabling quantification of fat and lean mass. Data were analyzed using nordicICE (NordicNeuroLab, Bergen, Norway) and jMRUI software. Maximum oxygen uptake (VO₂max) was determined following a standardized warm-up, using a ramp protocol with 15 W increments every 30 seconds until exhaustion. Test validity was confirmed by achieving a respiratory exchange ratio >1.10, blood lactate >7.0 mmol/L, and a plateau in VO₂ ( < 0.5 mL·kg⁻¹·min⁻¹ increase over a 30 W load).

Blood samples, as well as biopsies from skeletal muscle (m. vastus lateralis) and subcutaneous white adipose tissue, were collected at baseline and again three days after the final exercise session of the 12-week intervention. Biopsies were taken from the periumbilical region and vastus lateralis, following local anesthesia with lidocaine after skin sterilization. Samples were dissected on a cold aluminum plate to remove excess blood before snap freezing. Standard serum parameters were analyzed using either in-house protocols or by a commercial laboratory (Fürst Laboratories, Oslo, Norway).

Biopsies were snap-frozen in liquid nitrogen, pulverized, and homogenized in QIAzol Lysis Reagent (Qiagen) using a TissueRuptor. Total RNA was extracted with the miRNeasy Mini Kit (Qiagen), and RNA integrity and concentration were assessed using Agilent Bioanalyzer 2100 with RNA 6000 Nano Chips. cDNA was synthesized using the High-Capacity cDNA Reverse Transcription Kit (Applied Biosystems), and 25 ng RNA-equivalent cDNA per sample was used for sequencing.

All muscle and subcutaneous adipose tissue samples were deep sequenced at the Norwegian Sequencing Centre (University of Oslo) using the Illumina HiSeq 2000 system with RTA v1.17.21.3. Reads were demultiplexed with CASAVA v1.8.2, and quality assessed using FastQC v0.10.1. The average library size was 44–52 million 51 bp single end reads, with no group or time point differences and no batch effects. Alignment to the UCSC hg19 genome was performed using Tophat v2.0.8. Post-alignment quality control was done using IGV v2.3 and BEDTools v2.19.1. Gene-level counts were obtained using HTSeq v0.6.1 in intersection-strict mode. Data were normalized as reads per kilobase per million mapped reads (RPKM).

Statistics and bioinformatics

Olink proteomic data in both the UK biobank and MyoGlu cohorts were pre-processed and annotated as recommended by the manufacturer and described in detail previously^13,18. Biological age was estimated using the ProtAge model, a proteomic age predictor trained to estimate chronological age based on Olink NPX values across 204 proteins, as described previously¹³. Briefly, ProtAge was developed with a LightGBM model that uses plasma proteins to predict chronological age, with the 204 proteins identified via the Boruta algorithm and interpretation based on SHAP (SHapley Additive exPlanations) values¹³. The ProtAge score was calculated in both cohorts, and the difference between proteomic and chronological age, referred to as ProtAgeGap, was used as a measure of accelerated or decelerated biological age. To clarify, ProtAgeGap = protein score–chronological age.

In the UK Biobank, associations between ProtAgeGap and physical activity, body composition, and metabolic biomarkers were assessed in n = 45,438 participants using linear regression models. All models were adjusted for age, sex, ethnicity, recruitment center, and smoking. Standardized beta coefficients with 95% confidence intervals (CIs) were reported. To examine the association between ProtAgeGap and incident type 2 diabetes or IHD, Cox proportional hazards models were used. Incident cases were defined as those occurring after the date of recruitment for each participant, and all prevalent cases (with a diagnosis on or before the date of recruitment) were excluded from analysis. Models included ProtAgeGap, physical activity level (categorized using the International Physical Activity Questionnaire [IPAQ]), and an interaction term between the two. All interaction Cox models were adjusted for age and sex. Hazard ratios (HRs) and 95% CIs are reported. A significant interaction term indicated that the relationship between ProtAgeGap and diabetes/heart disease risk differed by physical activity level.

In MyoGlu, in addition to analyses using ProtAge and ProtAgeGap, a curated subset of the 204 proteins constituting the ProtAge score was selected for more in depth analyses. Transcriptomic data from skeletal muscle and adipose tissue were filtered to remove low-variance genes using the caret R package. Protein and gene expression changes (Δ = post minus pre intervention data) were calculated and used for downstream correlation analyses. Spearman’s correlation was used to compute pairwise associations between Δprotein, and ΔmRNA and Δclinical traits. Correlation matrices were visualized using hierarchical clustering and heatmaps (pheatmap R package), and clusters were defined by dendrogram cutree functions. For each gene cluster, top positively and negatively correlated proteins were identified by averaging Spearman correlation coefficients. Pathway enrichment analyses for gene clusters were performed using KEGG gene sets with the clusterProfiler R package and visualized using dot plots.

All p-values were adjusted for multiple testing using the Benjamini-Hochberg method with an FDR threshold of 5%.

Author contributions

Sindre Lee-Ødegård analyzed the data and wrote the first draft of the manuscript. M. Austin Argentieri analyzed UK Biobank data, read, revised and approved the manuscript. Frode Norheim and Christian André Drevon contributed to data acquisition, read, revised and approved the manuscript. Kåre Inge Birkeland was project leader and contributed to data acquisition and manuscript writing.

Data availability

The datasets generated and/or analyzed during the current study are available at Gene Expression Omnibus ID GSE227419, in the manuscript and its supplementary materials, and from the corresponding author on reasonable request. The UK Biobank data are available to bona fide researchers through the UK Biobank Access Management System (https://www.ukbiobank.ac.uk).

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41514-025-00318-w.