The mitochondrial multi-omic response to exercise training across rat tissues

Summary

Mitochondria have diverse functions critical to whole-body metabolic homeostasis. Endurance training alters mitochondrial activity, but systematic characterization of these adaptations is lacking. Here, the Molecular Transducers of Physical Activity Consortium mapped the temporal, multi-omic changes in mitochondrial analytes across 19 tissues in male and female rats trained for 1, 2, 4 or 8 weeks. Training elicited substantial changes in the adrenal gland, brown adipose, colon, heart and skeletal muscle. The colon showed non-linear response dynamics, while mitochondrial pathways were downregulated in brown adipose and adrenal tissues. Protein acetylation increased in the liver, with a shift in lipid metabolism, while oxidative proteins increased in striated muscles. Exercise upregulated networks were downregulated in human diabetes and cirrhosis. Knockdown of the central network protein HSD17B10 elevated oxygen consumption, indicative of metabolic stress. We provide a multi-omic, multi-tissue, temporal atlas of the mitochondrial response to exercise training and identify candidates linked to mitochondrial dysfunction.

eTOC

Amar et al. delineate the extensive molecular modifications occurring in mitochondria, central metabolic organelles, in response to endurance exercise training across diverse rat tissues. Their translational analysis suggests clinical relevance of the metabolic regulator HSD17B10. Its suppression induces metabolic stress in liver cells, underlining its potential role in disease pathology.

Graphical Abstract

Introduction

Mitochondria generate over 90% of the ATP required for mammalian cellular homeostasis via oxidative phosphorylation. Through ATP production and sensitivity to the cellular redox state and other signaling events, mitochondria support and regulate many cellular processes, including steroid biosynthesis, ketone body generation, gluconeogenesis, ion homeostasis, cellular calcium signaling and programmed cell death¹. Novel functions of mitochondrial proteins were recently discovered², affirming mitochondria as central hubs of metabolism. The circular mitochondrial DNA (mtDNA) encodes only 13 proteins, thus the majority of mitochondrial proteins are nuclear-encoded³.

A critical characteristic of mitochondria is their ability to adapt to subcellular-, cellular- and tissue-specific metabolic demands^4,5. For example, mitochondria of cardiac and skeletal muscle have high capacities for ATP production to support increased energy demand during contraction⁶. In the liver, mitochondria support gluconeogenesis, ketone production, and fatty acid synthesis and oxidation; key functions that allow the liver to contribute to systemic energy substrate supply at rest, fasting or during physiological stress⁷. In brown adipose tissue, uncoupling of the inner mitochondrial membrane by uncoupling proteins results in futile cycling and thermogenesis which are a critical adaptation for maintaining temperature during cold stress⁸. Given their central role in cellular homeostasis, it is perhaps not surprising that mitochondrial diseases are the most common among inherited metabolic disorders⁹, and mitochondrial function has been linked to several diseases, including type 2 diabetes, obesity, cancer, neurodegeneration, fatty liver and cardiovascular diseases^10–13.

Endurance exercise training results in robust increases in mitochondrial volume and ATP-generating capacity in skeletal muscle. Indeed, training-induced increases in skeletal muscle mitochondrial size and number and improvements in substrate oxidation have been known for decades^14,15. Moreover, endurance training imparts a myriad of health benefits through its multipotent effects across tissues and cell types, many of which are likely mediated by adaptations in mitochondrial functions. While endurance training improves mitochondrial quantity and quality in skeletal muscle¹⁶ and liver¹⁷, the effects in other tissues remain largely unknown, and system-wide multi-omic changes specific to the mitochondria have not been investigated.

The overall goal of the Molecular Transducers of Physical Activity Consortium (MoTrPAC) is to map the multi-omic response to exercise and training across tissues¹⁸. We have recently characterized the multi-omic changes across tissues in 6-month old female and male rats that were endurance exercise trained on a treadmill for 1, 2, 4 or 8 weeks¹⁹. Here, we focus on the temporal training response of mitochondrial analytes from the same animals. Importantly, we included many tissues that have not been studied in detail before. Moreover, the majority of efforts in this area have included one sex only, generally studied at a single timepoint. Here, we studied male and female rats across 4 timepoints, enabling insight into time-course and sexually dimorphic mitochondrial responses to training. We generate a molecular map of the multi-omic mitochondrial response to endurance training across 19 different tissues and compare our results to differential protein signatures of human disease, revealing gene networks that are induced by exercise but downregulated in disease. We show that knockdown of one of the central network proteins, HSD17B10, altered mitochondrial oxidative capacity in human liver cells. Our multi-omics data is accessible and searchable through an interactive online resource (motrpac-data.org) and tissues for follow-up experiments are available in the MoTrPAC biorepository.

Results

Endurance training alters biomarkers of mitochondrial volume across tissues

Male and female 6-month old F344 rats were subjected to progressive endurance training on a treadmill for 1, 2, 4 or 8 weeks. Additional sex-matched sedentary animals were collected as controls. Functional physiological assessments revealed increased aerobic capacity (VO₂max) after 8 weeks of training compared to controls in both male and female animals (by 12 and 10 mL/kg/min respectively, p<0.01). Body fat percentage decreased with training in males (8wk: −5 % p<0.001, 4wk −1.6 % ns) with no change in the untrained animals, whereas trained females retained their body fat percentage while untrained females increased body fat (2.8 % p<0.001) over 8 weeks. Within-sex response heterogeneity was relatively low. The detailed training protocol and accompanying phenotypic adaptations have been described elsewhere¹⁹. Blood, plasma and 18 solid tissues were collected 48 hours after the last exercise bout. Samples were profiled for epigenomics using transposase-accessible chromatin using sequencing (ATAC-seq) and reduced representation bisulfite sequencing (RRBS), transcriptomics using RNA sequencing (RNA-seq), proteomics and post-translational modifications (PTMs; phosphoproteome, acetylome and ubiquitylome) using LC-MS/MS, and metabolomics/lipidomics using several targeted and untargeted platforms (Fig. 1A). Transcriptomics and metabolomics/lipidomics were conducted on all 19 tissues, while proteomics and epigenomics were performed on selected tissues only (Fig. S1A). These same datasets are also included in a related study¹⁹.

Figure 1. Training-induced changes in biomarkers of mitochondrial volume.

A) Experimental design. Nineteen tissues were collected from male and female rats that remained sedentary (SED) or completed 1, 2, 4, or 8 weeks of progressive treadmill exercise training. Collected tissues were assayed for epigenomics (8 tissues), transcriptomics (19 tissues), proteomics (7 tissues), PTMs (phosphoproteome 7 tissues, acetylome and ubiquitylome 2 tissues), and metabolomics (19 tissues). Mitochondria-associated transcripts and proteins were selected using MitoCarta 3.0 and mitochondrial metabolites from a previously published dataset (ref. 27). HIPPOC = Hippocampus, HYPOTH = Hypothalamus, SMLINT = Small Intestine, SKM-GN = Gastrocnemius Skeletal Muscle, SKM-VL = Vastus Lateralis Skeletal Muscle, WAT-SC = Subcutaneous White Adipose Tissue, BAT = Brown Adipose Tissue, VENACV = Vena Cava. Created using BioRender.com. B) Correlation between mtDNA quantification and the percentage of mitochondrial RNA-seq reads. Dashed line represents rho=0.5. C) Training response of biomarkers of mitochondrial volume after 8 weeks of training. Cell marks: X=not significant (p>0.05), ?= tissues in which the biomarker was not assessed. Color scale is proportional to the ANOVA-test z-score. D) Comparison of the number of significant training responses of the mitochondrial biomarkers (p<0.05). E-F) Visualization of biomarker data in SKM-GN (E), and liver (F). Each boxplot represents abundance per sex and time group. ANOVA statistics are provided for each tissue and sex combination. The whiskers extend from the hinge to the largest and lowest values, but max 1.5 x(the interquartile range). *indicates timepoint-specific significance at *p<0.05, **p<0.01 and ***p<0.001 (displayed if ANOVA p-value <0.05). The y-axis range differs for visualization purposes.

To estimate changes in mitochondrial volume with training, we quantified biomarkers of mitochondrial quantity, including cardiolipin content and mtDNA copy number normalized to nuclear DNA²⁰. Mitochondrial DNA quantification, a commonly used marker of mitochondrial volume^21–23, was limited to 15 tissues in animals trained for 0, 1, or 8 weeks. Cardiolipins are mitochondrial membrane-specific lipids that correlate well with mitochondrial volume in human skeletal muscle²⁴. The cardiolipin data was available in kidney, liver, lung, gastrocnemius muscle (SKM-GN), and subcutaneous white adipose tissue (WAT-SC).

Striated muscle, brain and BAT were rich in mitochondria, in concordance with previous studies^25,26. In contrast, biomarker analysis indicates a lower relative mitochondrial abundance in spleen, WAT-SC and lung (Fig. S1B–C). MtDNA quantification was not available for all timepoints. However, we observed that the percent of RNA-seq reads mapped to the mitochondrial genome can serve as a proxy for mtDNA in four tissues: SKM-GN, vastus lateralis muscle (SKM-VL), adrenal and BAT (rho>0.5, bootstrap z-score>3, Fig. 1B). We also observed that different cardiolipins were highly correlated (Fig. S1D–F).

We observed a significant response to training in at least one biomarker in 15 of the 19 tissues (Table S1). Focusing on week 8, the percent of mitochondrial reads changed in eight tissues, with a concordant direction of response with mtDNA (Fig. 1C–D). Moreover, the top responding sex-tissue combinations manifested a concordant response of the mitochondria-encoded transcripts, suggesting changes in both mtDNA quantity and transcriptional activity (Fig. 1E). In contrast, the liver manifested a mild response that was limited to upregulation in males (Fig. 1F, see Fig. S2 and Table S1 for all biomarker data). In summary, we describe substantial cross-tissue differential changes of mitochondrial biomarkers in response to training.

The mitochondria-associated multi-omic response to exercise training is largely tissue-specific

We utilized MitoCarta 3.0³ to identify mitochondria-associated genes and proteins, and data from Heden et al.²⁷ to select mitochondrial metabolites. Principal component analysis of the baseline mitochondria-associated transcriptome showed separation by tissue, with additional separation by sex for multiple tissues (Fig. S3A). The baseline mitochondrial proteome was more similar across analyzed tissues, with distinct sex differences observed in WAT-SC and SKM-GN (Fig. S3B).

Training regulated mitochondria-associated analytes across all -omes and tissues. In total, 719 genes, corresponding to 63% of all mitochondria-associated transcripts, and 513 proteins (38% of all mitochondria-associated proteins) significantly changed with training in at least one sex and tissue. The most responsive tissues (>10% differential transcripts and/or proteins) were the adrenal gland, BAT, blood, colon, heart, liver, SKM-GN, SKM-VL and WAT-SC (Fig. 2A). SKM-GN, liver and heart showed the greatest proteomic response, while the metabolome/lipidome changed the most in blood, heart and liver. In contrast, there were few changes in the three investigated brain regions (cortex, hippocampus and hypothalamus), kidney, lung, testes and vena cava (Fig. 2A). Very few changes in phosphorylation and ubiquitination of mitochondrial proteins were observed, likely due to their transient nature.

Figure 2. The multi-omic mitochondrial response to training across tissues.

A) Heatmap of the number of mitochondria-associated analytes that significantly changed over the training time-course in at least one sex (5% FDR). Each cell represents results for a single tissue and data type. Numbers indicate training-differential mitochondrial analytes and colors indicate the proportion of differential to measured MitoCarta analytes. B) UpSet plot of the training-differential MitoCarta transcripts across tissues, identifying tissue-specific changes and differential features shared by many tissues. Numbers above vertical bars indicate transcripts in the connected tissues. Horizontal bars indicate the total number of differential transcripts per tissue. Only interaction sizes of 6 genes or more are shown. Enrichment results (MitoCarta pathways) are shown for two selected gene sets. C) MitoCarta pathway enrichments for the 8-week timepoint in the top 9 responding tissues. The 8-week male and female differential transcripts were identified using our graphical analysis. The plot shows the top pathway from each MitoCarta subcategory with the greatest number of enrichments, taken from the sum-of-log combined p-value per tissue and pathway. Each point represents a significant enrichment in a given node, where the direction of the triangle indicates the direction of effect (△= up, ▽ = down), the fill color indicates the corresponding sex (blue=male, pink=female), while a black triangle indicates sex-consistent enrichment. Because enrichment analyses are performed separately, significant enrichment can be found in multiple sets. For example, increased amino acid metabolism proteins in the heart for both the male-specific and sex consistent differential gene sets.

Comparing the differential transcripts of the top nine responsive tissues demonstrated a largely tissue-specific response, with mostly low cross-tissue correlations (Fig. S3C) and moderate overlap (Fig. 2B). Similar moderate-low overlap was observed for proteins (Fig S3D). The adrenal gland showed the greatest transcriptional response, followed by BAT and colon, where the commonly regulated genes (m=38) encoded tricarboxylic acid cycle (TCA) enzymes and subunits of Complex I (Fig. 2B).

The differential analysis above mainly focused on an overall training response across sex and time. Subsequently, we used the same models for contrasting each training timepoint with its sex-matched controls to produce time- and sex-specific summary statistics including fold changes and z/t-scores. We utilized an empirical Bayes graphical clustering²⁸ of these timewise z-scores to identify the main trajectories. These are used for visualization, identifying sets of analytes that change in a specific time- and sex combination, and mitochondria-specific pathway enrichments¹⁹ (Methods).

Using the graphical analysis results, we next focused on the adaptation in week 8. MitoCarta³ pathway enrichment analysis identified a consistent multi-omic upregulation of numerous mitochondrial pathways in the heart, skeletal muscles, WAT-SC and liver (Fig. 2C, Table S2). In the striated muscle tissues, the majority of the mitochondrial adaptation was sex-consistent. The liver response was also similar between the sexes, specifically in the proteome and acetylome, with few transcriptional changes. Many mitochondrial pathways also increased in trained WAT-SC, but showed mostly sex-specific regulation. The sexually dimorphic response to training in WAT-SC is analyzed in detail elsewhere²⁹. Mitochondria-associated metabolites primarily changed in liver, heart and blood. Due to the central role in metabolic homeostasis of these tissues, we correlated the metabolomic response between plasma and the other responding tissues (Fig. S3E). The liver response significantly correlated with plasma, supportive of its central role in whole body energy homeostasis.

Epidemiological evidence of reduced risk for colon cancer in trained individuals³⁰ supports an effect of training on the colon. Interestingly, we observed a delayed upregulation after 8 weeks (Fig. S3F). Importantly, many exercise interventions in animals are terminated at 4 weeks, where few transcriptional changes were observed. The overall response was substantially greater in females compared to males. Pathway enrichment analysis showed induction of oxidative phosphorylation (OXPHOS) and mitochondrial chaperone genes (Fig. 2C). However, despite a greater response in females, many differential transcripts converged to similar levels as males by week 8 (Fig. S3G). Sex-consistent upregulation was observed for the presequence import TIM23 pathway. TIM23 is stabilized by cardiolipins in the mitochondrial membrane, thus the combined increase in TIM23 and cardiolipin synthesis genes (i.e. Agpat4, Prelid1 and Ptpmt1) suggests an altered mitochondrial membrane dynamics in the colon. Further studies are needed to establish functional effects on the colon as it is limited to transcriptome and metabolome data in this study.

Adrenal and brown adipose tissue demonstrate differential transcriptional dynamics in response to endurance training

We considered two analyses of our differential trajectories to gain insights into their dynamics. First, we assessed whether our differential analysis results would change when adding markers of mitochondrial volume as covariates, and observed that adjustment primarily affected transcripts in BAT and the adrenal glands (Fig. S4A). Clustering analysis of the timewise results of this analysis identified four clusters that were enriched for different mitochondrial functions (Fig. S4B–C, Table S2–3, Methods). These results illustrate (1) a decreased BAT transcriptomic response in males, week 8 (Cluster 1), and (2) separation of the adrenal response into two clusters (2 and 4). Cluster 1 is consistent with our graphical analysis and its enrichment results (Fig. 3A, S4D–E), and with the BAT transcriptomic pattern of Ucp1, the key thermogenic uncoupler. In contrast, the chromatin accessibility and associated gene expression of Ucp2 are inversely correlated with Ucp1 (Fig. 3B–C). Ucp2 does not have the uncoupling potential of Ucp1 but has been shown to limit glucose oxidation and instead enhance glutaminolysis³¹. Together, these results suggest a potential mechanism for energy preservation in response to training in male BAT. The two adrenal clusters show a consistent downregulation in females across time, with a marked upregulation in males in week 1 (Fig. S4B, S5A). This response covers genes involved in beta-oxidation, the TCA cycle and the ETC (Fig. S5B–C). Post-adjustment, cluster 2 manifests downregulated patterns, with more similar male and female trajectories, while cluster 4 shows a nullified effect (Fig. S4B, Table S3). These two clusters differ in their enrichments, where the changes associated with a reduction in mitochondrial markers in females (cluster 4) were involved in amino acid metabolism (Fig. S5D). The female-specific decreases in fatty acid metabolism, TCA cycle and ETC were largely robust to adjustment.

Figure 3. Adrenal and brown adipose tissue demonstrate differential transcriptional dynamics in response to endurance training.

A) Graphical representation of the mitochondria-associated training-differential analytes in brown adipose tissue (BAT). Each node represents one of nine possible states (row labels, with F for females and M for males, seven states shown) at each of the four sampled training timepoints (column labels). Edges are drawn between nodes to represent the path of differential analytes over the training time-course. This graph includes the five largest paths. Both node and edge size are proportional to the number of analytes represented by the node or edge. B) Expression patterns of Ucp1 and Ucp2 in females (left) and males (right). *indicates FDR<0.05. C) Correlation between changes in Ucp2 expression and chromatin accessibility (intronic region of UCP2, chr1:165508254–165509507). Each point represents the average for n=5 animals per timepoint and sex. D) Expression changes of examples of known PPARGC1A/PGC1ɑ interactors. All are upregulated in males after 1 week and downregulated in females after 8 weeks, with exception of Jund, which is regulated in opposite directions (FDR<0.05). Data is shown as mean +/−SE for each gene per timepoint and sex. E) DREM analysis results of the MitoCarta genes in female adrenal gland predict several transcription factors to be involved in the 1-week responses.

Our second analysis was based on DREM³² to identify potential upstream regulators of the mitochondrial response. DREM combines clustering with transcription factor (TF) enrichment to identify regulation programs that may explain the identified trajectories. In BAT, known regulators of mitochondrial metabolic genes, including PPARA, PPARD and their coregulator PGC1ɑ (Ppargc1a), were associated with the late onset of downregulation of mitochondria-associated genes in males (Fig. S5E). In the adrenal data, we observed that Ppargc1a followed the same main clustering trajectory, increasing in males at week 1 with a decrease in females at week 8. Interestingly, several of the known interactors of PGC1ɑ showed the same pattern, including the PPARs¹⁹, FOSB, FOXO1 and THRB (Fig. 3D). Glucocorticoids, produced by the adrenal cortex, are known activators of Ppargc1a expression³³, as are changes in energy balance, temperature and calcium concentration³⁴. DREM predicted several TFs, including CNBP, SAP18, PHF5A and CHURC1, as upstream regulators in the adrenal in both sexes (Fig. 3E, S5F). In females, additional factors were predicted, including PPARA, PPARG and ESRRA (Fig. 3E), which are all interactors of PGC1ɑ. Thus, the DREM results collectively suggest potential upstream regulators of PGC1ɑ and mitochondrial biogenesis in the adrenal gland in response to endurance training.

Training induces sex-consistent upregulation of mitochondrial metabolism in skeletal muscle

The SKM-GN analytes showed sex-consistent trajectories, with an overall upregulation across timepoints (Fig. 4A). The main trajectory was driven by alterations in protein abundance, with changes in OXPHOS seen for both transcripts and proteins. In contrast, lipid metabolism and TCA cycle enzymes were almost exclusively regulated at the protein level. These findings were concordant with pathway enrichment analyses of the sex-consistent upregulated genes in both SKM-GN and SKM-VL (Fig. 4B, Fig. S6A–C). Overall, there was a significant overlap between the differential transcripts and proteins in the investigated muscle tissues (Fig. S6D–E). Moreover, females specifically increased fatty acid oxidation proteins in SKM-GN, e.g. ACSS1, ECHS1, ECI1 and HADHA (Fig. 4C). Similarly, the complex III pathway was only upregulated in females (Fig. S6F).

Figure 4. Endurance training induces largely sex-consistent increases in metabolic protein abundance in skeletal muscle.

A) The dynamics of the molecular training response visualized by constructing a summary graph in which rows represent nine combined states (row labels, with F for females and M for males, seven states shown) and columns represent the four training timepoints. Nodes correspond to a combination of time, sex, and state. An edge connects two nodes from adjacent timepoints. The differential abundance trajectory of any given training-regulated analyte is represented by drawing a path through the nodes in this graph. This graph represents the five largest trajectories for mitochondria-associated training-differential analytes in SKM-GN. Both node and edge size are proportional to the number of analytes represented by the node or edge. B) Network view of pathway enrichment results corresponding to the analytes of the week 8, sex-consistent upregulation nodes in SKM-GN and SKM-VL. Nodes indicate enriched pathways (10% FDR), and an edge represents a pair of nodes with a similarity score of at least 0.3 between the gene sets driving each enrichment. Node fill color indicates for which –ome or –omes a pathway is significant, while border color indicates if the pathway is significant in one or both tissues. Node size is proportional to the number of differential analyte sets for which the pathway is significantly enriched. Clusters of enriched pathways were defined using Louvain community detection, and are annotated with high-level biological themes. C) Fatty acid oxidation pathway enrichment for the SKM-GN proteome. Only significant genes are shown. Rows are clustered using hierarchical clustering. D) Log2 fold changes of significant differential protein phosphorylation sites in Complex I proteins. All phosphorylation changes are significant in females, whereas all except Ndufs5_T93 are significant in males at 8 weeks. Data is shown as mean +/−SE for each gene per timepoint and sex.

Mitochondrial DNA and biogenesis pathways were upregulated in both SKM-GN and SKM-VL, as expected³⁵. The overall mitochondrial volume and quality is also regulated by mitophagy; controlled degradation of damaged mitochondria, a critical quality control mechanism induced by exercise³⁶. Accordingly, we observed transcriptional induction of mitophagy pathways in SKM-GN and increases in DRP1, an important regulator of mitochondrial dynamics in response to exercise³⁷. Another quality control mechanism is the mitochondrial unfolded protein response; an adaptive stress response to counteract mitochondrial stress and dysfunction³⁸. Key TFs that are expressed to promote the response are ATF4, ATF5 and DDIT3, of which Atf5 was upregulated primarily in female SKM-GN and SKM-VL in the early training timepoints (Fig. S6G).

Phosphoproteome changes have been described in skeletal muscle in response to acute exercise³⁹, and increased phosphorylation of Complex I has been associated with increased activity in cardiac muscle⁴⁰. We observed an induction of Complex I phosphorylation in skeletal muscle with training (Fig. 4D). The abundance of several of the Complex I proteins also changed, but to a lesser extent compared to phosphorylation. Thus, we observe several levels of regulation associated with increasing oxidative capacity in skeletal muscle, which is critical for preventing age-related decline in mitochondrial functions⁴¹.

Endurance training alters the cardiac mitochondrial acetylome

Differential analysis of cardiac mitochondrial analytes revealed a sex-consistent upregulated trajectory of transcripts and proteins across timepoints (Fig. 5A). The second largest trajectory showed delayed downregulation at week 8 (Fig. 5A–B). Training increased transcript abundance of OXPHOS genes, similar to skeletal muscle, while TCA, BCAA (branched chain amino acid) and fatty acid metabolism pathways were regulated across multiple -omes (Fig. 5C). These findings were corroborated by significant enrichment of cardiac tissue TCA acids and amino acids. Sex-consistent downregulated proteins were associated with Coenzyme Q metabolism, whereas a reduction in acetylation occurred for BCAA and lipid metabolism proteins (Fig. 5D).

Figure 5. Endurance training alters the cardiac mitochondrial acetylome.

A) Graphical representation of the mitochondria-associated training-differential analytes in cardiac muscle. Each node represents one of nine possible states (row labels, with F for females and M for males, seven states shown) at each of the four sampled training timepoints (column labels). Edges are drawn between nodes to represent the path of differential analytes over the training time-course. This graph includes the five largest paths. Both node and edge size are proportional to the number of analytes represented by the node or edge. B) Number of significantly up- and downregulated mitochondria-associated transcripts and proteins at each timepoint, with color representation based on the main MitoCarta pathway association of each analyte. C-D) Network view of pathway enrichment results corresponding to the analytes C) downregulated in both sexes after 8 weeks (the 8w_F-1_M-1 node in (A)) and D) upregulated in both sexes after 8 weeks (the 8w_F1_M1 node in (A)). Nodes indicate enriched pathways (10% FDR), and an edge represents a pair of nodes with a similarity score of at least 0.3 between the gene sets driving each pathway enrichment. Node fill color indicates for which –omes a pathway is significant, while a black border color indicates if the pathway is significant in both the down- and upregulated nodes. Node size is proportional to the number of differential analyte sets for which the pathway is enriched. Clusters of enriched pathways were defined using Louvain community detection, and are annotated with high-level biological themes. E-F) Correlation between changes in protein levels and acetylation levels in males (E) and females (F). Orange=MitoCarta proteins, gray=other proteins. G) Acetylation and phosphorylation changes (FDR<0.05) of metabolic proteins in males and females after 8 weeks (sites changing in only one sex are not illustrated). Each lollipop represents a specific acetylation (rounded top) or phosphorylation (diamond top) site, red=increase, blue=decrease. Multiple lollipops on the same protein indicates several sites changed. Hadha had 8 differentially acetylated sites, with only 6 illustrated due to space constraints. H) Site-specific acetylation changes in ACAT1 in males (top panel) and females (bottom panel), and in I) ACO2 in males (left panel) and females (right panel). All displayed sites were differentially acetylated overall (taking all timepoints and sexes into account, FDR<0.05), and sites that reach timewise significance (FDR<0.05) are highlighted with black frames.

Endurance training induced multiple acetylation changes of mitochondrial proteins (220 significant sites after 8 weeks, Table S4); a known regulatory mechanism of cardiac bioenergetics that has been associated with cardiac disease pathogenicity⁴² and aging⁴³. The majority of acetylation changes occurred independently of changes in protein abundance (Fig. 5E–F). Protein-specific acetylation changes occurred in key bioenergetic pathways (Fig. 5G). Multiple β-oxidation, BCAA and TCA cycle enzymes were significantly regulated at several acetylation sites. Reduced acetylation was primarily observed in HADHA (Fig. S7A), NDUFA12 and ACAT1. Reduced acetylation of ACAT1, involved in both β-oxidation and BCAA catabolism, occurred at lysines 260 and 265 (Fig. 5H), two sites where deacetylation increases protein activity due to increased affinity for Coenzyme A⁴⁴. Conversely, training induced acetylation of multiple lysines in the cardiac TCA cycle enzyme aconitase (ACO2, Fig. 5I), which increases its activity⁴⁵. Thus, our study demonstrates major acetylation changes at specific cardiac mitochondrial proteins in response to training.

Training induces remodeling of the metabolic protein acetylome in the liver

Endurance training leads to functional improvements in hepatic mitochondria⁴⁶, but the molecular mechanisms driving these changes are unknown. Eight weeks of training altered mitochondrial analytes across all -omes, with the majority of changes occurring at the protein and PTM levels. Very few transcriptional changes occurred (Fig. 6A), which could be tied to higher rates of mRNA turnover in the liver⁴⁷. The main trajectory showed male-specific increases across the first 4 weeks of training and an increase in both sexes after 8 weeks (Fig. 6A). Pathway enrichments for this trajectory were dominated by changes in acetylation of Complex V proteins, mitochondrial biogenesis and BCAA catabolism, as well as lipid metabolism proteins (Fig. S7B). The second largest trajectory (>200 analytes) remained unchanged during the first 4 weeks of training but subsequently increased in both sexes by week 8 (Fig. 6A). Here, pathway enrichment was dominated by changes in mitochondria-encoded Complex I transcripts (Fig. S7C).

Figure 6. Training-induced mitochondrial adaptation in the liver through protein acetylation.

A) Graphical representation of the mitochondria-associated training-differential analytes in the liver. Each node represents one of nine possible states (row labels, with F for females and M for males, seven states shown) at each of the training timepoints (column labels). Edges are drawn between nodes to represent the path of differential analytes over the training time-course. This graph includes the five largest paths. Both node and edge sizes are proportional to the number of analytes represented by the node or edge. B-C) Correlation between changes in protein abundance and acetylation in B) male and C) female liver. Pink=MitoCarta proteins, gray=other proteins. D) Acetylation and phosphorylation changes (FDR<0.05) of mitochondrial metabolic proteins in male and female liver after 8 weeks of training (sites changing in only one sex are not illustrated). Each lollipop represents a specific acetylation (rounded top) or phosphorylation (diamond top) site, red=increase, blue=decrease. Multiple lollipops on the same protein indicates several sites changed. Proteins with more significant differential sites than could be fitted into the illustration due to space were; Atp5c 9 sites, Atp5a1 9 sites, Atp5h 13 sites, and Idh2 15 sites in total. E) Protein expression changes in Sirt3 and Sirt4. Females are represented by circles and males by triangles. Data is shown as mean +/−SE for each gene per timepoint and sex. *indicates significance at FDR<0.05. F) Site-specific acetylation changes in HMGCS2 in males (left) and females (right). All displayed sites were differentially acetylated overall (taking all timepoints and sexes into account, FDR<0.05), and sites that reach timewise significance (FDR<0.05) are highlighted with black frames.

Mitochondrial protein acetylation accounted for over 60% of the significant changes observed in liver following training (Fig. 2A). The majority of the differentially acetylated proteins remained unchanged in abundance (Fig. 6B–C), demonstrating a specific post-translational regulatory mechanism in response to training. We observed changes in acetylation of multiple sites across mitochondrial proteins involved in all major bioenergetic processes (Fig. 6D). In particular, there was an increased acetylation of enzymes involved in lipid transport, lipid catabolism, OXPHOS and the BCAA degradation pathway. We also observed an increase in the SLC25A44 protein, a mitochondrial BCAA transporter that increases with training also in human skeletal muscle⁴⁸. Increased acetylation was especially notable in Complex V (Fig. 6D), similar to human skeletal muscle⁴⁹. Acetylation of mitochondrial proteins is an important mechanism for regulating fatty acid metabolism⁵⁰ and Complex I activity⁵¹, although the effect of site-specific changes on enzymatic activity remains largely unknown. Acetylation occurs primarily through non-enzymatic mechanisms^52,53, while deacetylation occurs through deacetylating enzymes. There was a small increase in the NAD+ dependent mitochondrial sirtuin SIRT3 across all training timepoints in males, and after 8 weeks in females (Fig. 6E), but no change in the NAD+/NADH ratio (Fig. S7D) or the mitochondrial acetyltransferase GCN5. These data suggest that greater acetylation of hepatic mitochondrial proteins occurred despite concomitant elevation in SIRT3 activity, similar to the effect observed following training in cardiac tissue.

Next, we delineated training-induced, site-specific acetylation changes of critical hepatic metabolic enzymes. HMGSC2 is the rate-limiting enzyme in the synthesis of ketones. Exercise increases ketogenic flux particularly in the postabsorptive state⁵⁴; however, the chronic effects of exercise on ketogenic capacity are less known. Liver ketone production provides an alternative fuel source during exercise when glucose levels are challenged, and attenuates skeletal muscle proteolysis during intense exercise⁵⁵. We observed deacetylation of lysines 310, 447 and 473 in HMGCS2 (Fig. 6F), which are known to enhance HMGCS2 enzyme activity⁵⁶. Deacetylation of pyruvate dehydrogenase complex component PDHA1 was also observed (Fig. S7E), which enhances enzyme activity⁵⁷, thus promoting oxidative phosphorylation. These findings demonstrate the plasticity of the hepatic mitochondrial proteome and acetylome in response to endurance exercise training, and provide novel targets for future mechanistic studies.

Training upregulates protein networks that are downregulated in type 2 diabetes and cirrhosis in humans

Mitochondrial dysfunction is a hallmark of chronic diseases, including obesity, type 2 diabetes (T2D), nonalcoholic fatty liver disease (NAFLD) and neurodegenerative diseases¹¹. We therefore examined the potential relevance of our identified mitochondrial gene sets to disease data. First, we performed disease ontology enrichment analysis of our identified analyte sets, limited to features that can be mapped to MitoCarta genes. The most significant enrichments were of NAFLD and other metabolic disease terms in our liver up- and downregulated sets (Table S5).

Second, we compared our sets to nine proteomics case-control datasets, which covered skeletal muscle of T2D patients; liver of NAFLD, NASH (nonalcoholic steatohepatitis) and cirrhosis patients; cardiac muscle of hypertrophic cardiomyopathy (HCM) patients; and rodent models of myocardial infarction and heart failure. We focused on proteomics as the majority of the observed training-induced changes in these tissues occurred at the protein level. We found significant overlaps in: T2D (skeletal muscle), obesity and cirrhosis (liver), and HCM and heart failure (cardiac muscle, Fig. 7A). The overlap with protein changes in response to myocardial infarction or NASH were not significant.

Figure 7. Training results in opposite regulation of mitochondrial proteins compared to type II diabetes and cirrhosis.

A) Significance of the overlap between the exercise-regulated differential proteins compared to identified proteins in case-control proteomics disease cohorts. The horizontal line represents p=0.05. MI=Myocardial Infarction, HCM=Hypertrophic Cardiomyopathy, NASH=Non-alcoholic Hepatosteatosis, Cirr=Cirrhosis, T2D=Type 2 Diabetes, HF=Heart Failure. N=8 studies with both differential proteins and the background set of all quantified proteins. Each presented dataset is named by disease and first author. B) Significance of the opposite directionality (Fisher’s exact test) when comparing the fold change sign of the overlapping proteins from (A) plus one study on NAFLD=Non-alcoholic Fatty Liver Disease. C-D) GeneMANIA protein-protein networks of proteins in opposite directions. C) Sex-consistent, week-8, skeletal muscle differential proteins that had opposite direction of effect in two separate T2D cohorts. D) Liver cirrhosis network of the 8-week female differential proteins that were both significant and had opposite direction of effect in the liver cirrhosis cohort. E) Protein abundance of HSD17B10 in HepG2 cells after siRNA-based knockdown (n=3) compared to negative controls (scramble siRNA, n=3). Number represents p-value. Representative western blot shown above bars. F-G) Oxygen consumption rate (OCR) in knockdown (n=10) versus control cells (n=9) assayed using Seahorse. Data is shown for all collected timepoints in F, with the non-mitochondrial respiration shown in the yellow section. Each calculated variable from F is shown in G. Numbers represent FDR-corrected p-values. H) Protein abundance of electron transport chain components measured using Western blot (n=3 knockdown, n=3 controls).

Next we interrogated the directionality of the proteomic signatures in response to training compared to disease. We found a robust, opposite regulation in response to training in our rats compared to protein signatures of human patients with T2D in two separate datasets (Fig. 7B). Focusing on our sex-consistent differential proteins that were regulated in the opposite direction in T2D, we identified a dense protein-interaction network (Fig 7C) that includes mitochondrial matrix proteins and proteins involved in ATP production through the TCA cycle and ETC (q-value <5.4×10⁻⁷), with HSD17B10 (17-beta-hydroxysteroid dehydrogenase 10) as a main hub.

In the liver, there was a striking opposite regulation in cirrhosis patients, uniquely in comparison to the training response in females. Next, we overlaid the common differential proteins that had an opposite fold change in exercise compared to disease on known protein-interaction and pathway networks. Exercise-upregulated proteins clustered mostly separately from the exercise-downregulated proteins (Fig. 7D). This dichotomy was also observed when performing enrichment analysis: mitochondrial matrix and metabolic pathways were upregulated in response to training (downregulated in cirrhosis), and RNA processing pathways were instead downregulated in response to training (q-value <3.7×10⁻⁷). HSD17B10, which was also observed in the T2D network, is the main hub of the exercise-upregulated proteins. It is a multi-functional mitochondrial dehydrogenase involved in fatty acid, amino acid and steroid metabolism, as well as tRNA maturation⁵⁸. Mutations in Hsd17b10 cause HSD10 mitochondrial disease, leading to neurodegeneration, liver dysfunction and cardiomyopathy⁵⁹. In human liver cells, silencing of Hsd17b10 transcripts (Fig. 7E), to simulate the disease condition, resulted in elevated oxygen consumption rate and proton leak (Fig. 7F–G), an indicator of mitochondrial metabolic stress that also occurs in cancer cells⁶⁰. No difference was observed in levels of key OXPHOS proteins (Fig. 7H). In line with this finding, increased activity of Complex II has been observed in the liver of an HSD10 patient⁶¹.

Discussion

Exercise is a powerful modulator of mitochondrial function and health, an effect that is conserved across mouse, rat and human. While concerted efforts have been made to understand the mitochondrial response in one or two tissues^62,63, this has not been undertaken at the scale or breadth of this work. We show that exercise training activates a concerted mitochondrial response in multiple tissues, including the scarcely investigated colon and adrenal glands. We observe the greatest multi-omic mitochondrial response in skeletal muscle, heart, liver, colon, adrenal gland, BAT, WAT-SC and blood. Importantly, identifying induction of Complex I phosphorylation as a potential mechanism for training-induced increases in ETC activity in skeletal muscle demonstrates how the breadth and scale of our data allows inspection of where regulation checkpoints may occur. We detected minimal mitochondrial changes in several tissues, including the brain, small intestine and spleen. While brain activity increases during exercise⁶⁴, the brain regions assayed here are already rich in mitochondria with a high rate of glucose consumption at rest⁶⁵ and the increase in metabolic demand during exercise is comparatively low⁶⁶.

Despite common metabolic functions of mitochondria across tissues, many differential analytes were tissue-specific, with a greater overlap among the striated muscle tissues. Further work is required to delineate the specific mechanisms underlying this diversity in response. Potential contributing factors include blood flow, exposure to exerkines and energy demand and utilization. Specifically, blood flow to adipose tissue, the contracting muscle and the heart increases during exercise⁶⁷, while flow to the liver is decreased although oxygen uptake increases⁶⁸, with such repeated changes likely driving a mitochondrial response. Several differential pathways were affected in multiple tissues, including oxidative phosphorylation. Improved bioenergetics is a critical exercise adaptation, and while ETC complexes increase in abundance in response to exercise, enhanced electron flux to the ETC could be more important for ATP-generating capacity⁶². Formation of supercomplexes (assemblies of ETC complexes⁶⁹) has been proposed as a mechanism for improved respiratory capacity⁷⁰, also in response to exercise^62,71,72, although its effect is controversial^62,73,74.

Thermogenesis is the principal function of BAT, a tissue rich in mitochondria⁷⁵. In contrast to female rats, males lost WAT-SC mass with endurance training¹⁹. In this context, our observed reduction in the metabolic/thermogenic transcriptional activity of BAT could serve as an overall energy preservation effect in males. Decreased BAT mass and activity is also observed in chronically trained endurance athletes⁷⁶, corroborating this hypothesis. There was also a substantial decrease in mitochondria-associated transcripts in the female adrenal gland. While the physiological effects of these changes remain elusive, the mitochondria are important stress response modulators known to affect sympathetic adrenal-medullary activation, catecholamine and cortisol/corticosterone levels⁷⁷, and this mitochondrial adaptation is potentially reflective of reduced physiological stress after adapting to repeated bouts of exercise. The responses in BAT and the adrenal gland were based on transcriptomics only, and were sensitive to adjustment for markers of mitochondrial volume. Future studies are needed to validate these findings and further analyze the correlation between relevant biomarkers and actual mitochondrial volume.

Mitochondrial biogenesis and oxidative function are strongly impacted by sex⁷⁸. Sexual dimorphism in the skeletal muscle baseline transcriptome indicates higher lipid oxidation in females⁷⁹, but sex-specific responses to training are less clear, although the inclusion of more females in exercise studies indicate sex-specific responses⁸⁰. This work provides a foundational resource to understand the importance of sex on temporal mitochondrial adaptations to exercise across 19 different tissues. To this point, we find large sex differences in the temporal dynamics of mitochondrial analytes in response to training in the adrenal gland and BAT. We observed sex-consistent mitochondrial responses in skeletal muscle, as well as sex-specific metabolic responses, including greater induction of lipid oxidation enzymes in females.

Liver mitochondria play a critical role in oxidizing fat to provide ATP and substrates to fuel TCA cycle flux and gluconeogenesis⁸¹, which in turn maintain circulating glucose levels⁸². Endurance training leads to functional improvements in hepatic mitochondria independent of increases in mitochondrial volume^46,83. Moreover, females show increased hepatic mitochondrial capacity both at baseline and following exercise^84,85, changes likely due to metabolic demands of gestation and lactation. While exercise training increases protein acetylation in skeletal muscle⁸⁶, much less is known in the heart and liver. We find a dramatic remodeling of the liver mitochondrial acetylome with training, likely due to increased substrate flux and associated turnover of acetyl-CoA in the mitochondria. Current dogma suggests that nutrient excess causes increased acetylation of mitochondrial proteins that lead to downregulated oxidation metabolism⁸⁷. However, exercise is known to protect hepatic metabolic health and improve oxidative capacity and our results show dramatic increases in acetylation in response to training. We observed increased acetylation of mitochondrial proteins in both liver and heart despite a concomitant increase in SIRT3, a mitochondria-specific deacetylase. This finding suggests that SIRT3 is not the cause of the changes in acetylation but may be increasing to control and break a PTM mechanism, such as the shift of the ratio between NAD+ and NADH linked to cellular oxidative stress⁸⁸. While we did not observe a change in the NAD+/NADH ratio 48 hours after exercise, the redox state may have been altered in association to each exercise bout. Recent studies have shown significantly modified liver proteome and PTMs in response to acute exercise^89,90, while the role of acetylation of metabolic proteins in the training-induced improvement in liver metabolic health is unknown and warrants further study.

Maintaining mitochondrial function is critical to healthspan, as mitochondrial dysfunction is a hallmark of aging and is associated with cardiovascular, metabolic and neurodegenerative diseases⁹¹. In contrast, endurance training is a promising intervention for prevention or attenuation of mitochondrial decline. The mitochondrial proteome robustly responded to training in cardiac, skeletal muscle, and liver tissues, which motivated our comparison of the effects in these tissues with proteomic datasets from disease cohorts. We found opposite proteomic skeletal muscle response to training compared to changes observed in human skeletal muscle in T2D patients. There was also opposite regulation of the liver proteome in response to exercise in females compared to changes induced by liver cirrhosis in human patients. Interestingly, the same mitochondria-specific protein, HSD17B10, was identified as the central hub from protein interaction network analyses of the oppositely regulated proteins in both T2D and liver cirrhosis, though its first degree (exercise) upregulated neighbors were different. While complete HSD17B10 knockout is embryonically lethal⁹², missense mutations in humans cause severe, but variable mitochondrial dysfunctions that primarily affect the nervous system, liver and heart⁶¹. We showed that silencing of HSD17B10 in human liver cells changes cellular energetics, an indicator of cellular stress, such as in cancer⁶⁰. The training-induced increase in HSD17B10 may help coordinate adaptations to exercise in response to the exercise-mediated perturbations in cellular metabolic stress.

Collectively, our work expands upon previous findings by providing an unprecedented multi-omic resource of the mitochondrial adaptation to endurance exercise training. We have concurrently mapped the time-course of mitochondrial multi-omic responses to endurance training across 19 tissues in males and females, reflecting an integrative effect on whole-body metabolism. Importantly, this MoTrPAC resource provides much-needed insight into sex-specific mitochondrial adaptations and how these changes occur over time. All results can be interrogated in a user-friendly, openly accessible data portal (motrpac-data.org). Altogether, considering the critical role that mitochondria play in maintaining tissue-specific and whole body metabolic health, this work provides an unparalleled resource to stimulate hypothesis-driven, mechanistic studies, as well as work aimed at identifying targets that can be leveraged therapeutically to combat mitochondrial dysfunction and metabolic diseases.

Limitations of study

Our study has some limitations that can be addressed by future studies. Our differential analysis focused on timepoint- and sex-specific effects, and future work can directly compare the sexes or tissue pairs. Our rat model was beneficial for exploring many tissues from the same animals and tissues that are inaccessible in humans, but all findings should be interpreted in the context of the investigated species and strain. The sample size of 5–6 animals per sex and timepoint is limited and the sampling timepoint, 48 hours after the last exercise bout, is important to consider for interpretation of the results. Although we assess functional physiological changes and biomarkers of mitochondrial volume, we lack direct measurements of mitochondrial respiration and volume. The variability in the response of biomarkers in our study suggests that replication with larger sample size is needed. The acute, transient changes that govern mitochondrial dynamics are not captured using our experimental design. Our datasets are limited in not covering all omes across tissues, and by not having direct single cell or cell type composition measures. Future studies would also benefit from inclusion of isolated mitochondria. A general challenge of mitochondrial adaptation studies is that multiple proteins that localize to mitochondria (48%) can also be found in other cellular compartments⁹³. Finally, translational relevance is a concern when extrapolating findings from rodents to humans.

STAR Methods

Resource Availability

Lead Contact

Further information and requests for resources should be directed to and will be fulfilled by the Lead Contact, Malene E. Lindholm (malenel@stanford.edu).

Materials Availability

This study did not generate new unique reagents.

Data and Code Availability

All -omics data from this study are available in the R package https://motrpac.github.io/MotrpacRatTraining6moData/

Timewise differential results adjusted for biomarkers of mitochondrial volume are available in Zenodo: https://zenodo.org/record/7459795#.ZEAw9ezMKLs. The cardiolipin and mtDNA data are also available in this repository. Analysis code for reproducing the results from this study are available here: https://github.com/MoTrPAC/motrpac-rat-training-mitochondria.

Experimental Model and Subject Details

Exercise training protocol

Inbred male and female Fischer 344 rats were obtained from the National Institute of Aging (NIA) rodent colony. Rats were housed in pairs at a reverse dark-light cycle, kept at a temperature of 20–25°C and fed normal chow (Lab Diet 5L79). After familiarization (>10 days for reverse light cycle, 12 days for treadmill), rats were randomized to training or control. The rats were partitioned into three groups; 8-week rats that were randomized to training or control, 4-week rats that were all assigned to training, and 1- and 2- week rats that were randomly assigned to 1- or 2-weeks of training. A total of 50 rats (5 males and 5 females per timepoint) were used for molecular analyses, with the exception of proteomics that was performed on 60 animals (6 males and 6 females per timepoint). These sample sizes were based on a balance between the project’s allocated budget (i.e. limitations due to performing multiple assays per animal across tissues) and power to detect VO₂max changes in response to training. Power calculations were based on statistics from previous exercise studies with the same rat model and guaranteed power >85% in males for n=5 or n=6 per group, and 74% and >85% in females for n=5, and n=6 per group, respectively^94,95. All training groups started training at 6 months of age (corresponding to about 18 years of age in humans⁹⁶) and trained on a Panlab 5-lane rat treadmill (Harvard Instruments, Model LE8710RTS). Rats were exercised 5 days per week using a progressive protocol aimed to maintain an intensity corresponding to approximately 70% of VO₂max (increasing grade and speed, see¹⁹ for details), with a maximal duration of 50min for the last two weeks of training. The starting treadmill speed was based on VO₂max measurements obtained following familiarization. All animal procedures took place during the dark cycle and were approved by the Institutional Animal Care and Use Committee at the University of Iowa, where the training intervention took place.

Phenotyping

Body composition was measured using nuclear magnetic resonance (Minispec LF90II Body Composition Rat and Mice Analyzer) for all rats prior to training, and for the 4-week and 8-week animals five days prior to tissue harvest. Maximal oxygen consumption (VO₂max) was similarly measured in all rats prior to training, and during the last week of training for the rats in the 4-week and 8-week training groups and the sedentary group. The testing protocol consisted of a 15-minute warm up at a treadmill speed of 9 m/min and 0° incline. The incline was subsequently increased to 10° and the speed was increased by 1.8 m/min every 2 minutes⁹⁷ until exhaustion, defined as when the rat sat on the shock area 3 consecutive times without responding to increased shock. Blood for lactate assessment was taken from the tail immediately after the test. The criteria for reaching VO₂max was a leveling off of VO₂ despite increased workload, a respiratory exchange ratio >1.05, and a blood lactate concentration ≥6 mM.

Tissue collection

All tissues were collected 48 hours after the last exercise bout. Food was removed three hours prior to the start of dissections, for which rats were sedated with inhaled isoflurane (1–2%) and kept under anesthesia until death. Blood was obtained through cardiac puncture, then the gastrocnemius muscle, subcutaneous white fat, right lobe of the liver, heart, and lungs were removed in that particular order. Removal of the heart resulted in death. A guillotine was subsequently used for decapitation, after which the brain was removed, and the hypothalamus, right and left hippocampus, right and left cerebral cortex were dissected out. After decapitation, the right kidney, right and left adrenal glands, spleen, brown adipose tissue, small intestine, colon, right testes or ovaries and right vastus lateralis were removed in that order. All tissues were flash-frozen in liquid nitrogen and stored at −80°C.

Cell culture

Human liver (HepG2) cells (Sigma-Aldrich) were cultured in Dulbecco’s Modified Eagle Medium (DMEM, Life Technologies) supplemented with 10% FBS, 100U of penicillin and 0.1mg/ml streptomycin. Cells were cultured in a humidified chamber at 37 °C, 5% CO₂.

Method details

Aliquots of assay- and tissue-specific reference standards were included in all molecular assays to evaluate technical differences across batches. All assays, omic quantification pipelines, and quality assurance processes are described in¹⁹ and an overview is provided here.

RNA sequencing

Following tissue lysis, total RNA was extracted in a BiomekFx automation workstation. Total RNA from blood was extracted using the Agencourt RNAdvance blood specific kit (Beckman Coulter). The RNA quantity and quality were assessed with NanoDrop (ThermoFisher Scientific, # ND-ONE-W), Qubit assay (ThermoFisher Scientific), and either Bioanalyzer or Fragment Analyzer. Approximately 500 ng of total RNA was used to generate RNA-seq libraries from poly(A)-selected RNA using the Universal Plus mRNA-Seq kit from NuGEN/Tecan (# 9133). All libraries were prepared using a Biomek i7 laboratory automation system (Beckman Coulter). Pooled libraries were subsequently sequenced using 100bp paired-end sequencing on an Illumina NovaSeq 6000 platform (Illumina, San Diego, CA, USA) targeting a depth of 35 million read pairs per sample. Reads were demultiplexed with bcl2fastq2 (v2.20.0), adapters were trimmed with cutadapt (v1.18), pre-alignment QC metrics generated with FastQC (v0.11.8) and reads aligned using STAR (v2.7.0d). Quantification was performed using RSEM (v1.3.1). Normalized sample-level data were generated using filtered raw counts that were TMM-normalized (edgeR::calcNormFactors), and converted to log counts per million with edgeR (edgeR::cpm)⁹⁸. As all samples for a given tissue were processed in a single batch, batch correction was not possible or necessary.

Reduced representation bisulfite sequencing

Tissues were lysed and DNA extracted using a BiomekFx automation workstation (Beckman Coulter). Approximately 100ng of DNA was used for library preparation using the Ovation^® RRBS Methyl-Seq kit (Tecan Genomics), which includes bisulfite conversion of the DNA. The quantity and quality of the libraries were assessed using Qubit High Sensitivity assays (ThermoFisher Scientific), and the Bioanalyzer High Sensitivity DNA Chip (Agilent Technologies). The pooled libraries were sequenced using 100bp paired-end sequencing on an Illumina NovaSeq 6000 platform (Illumina, San Diego, CA, USA) to reach at least 30 million paired-end reads per library using a custom 1-index primer as per Illumina guidelines. Processing included demultiplexing with bcl2fastq (version 2.20), adapter trimming using TrimGalore (v1.18), and indexing and aligning reads to the Rattus norvegicus (rn6) genome using Bismark (v0.20.0). Bismark’s deduplicate_bismark with “-p --barcode” options was used to remove PCR duplicates from the bam files; and Bismark’s “bismark_methylation_extractor --comprehensive --bedgraph” was used to quantify methylated and unmethylated coverages for all the CpG sites. Bowtie 2 (v2.3.4.3) was used to index and align reads to globin, rRNA, and phix sequences in order to quantify the percent of reads that mapped to these contaminants and spike-ins. SAMtools (v1.3.1) was used to compute mapping percentages to different chromosomes. UMIs were used to accurately quantify PCR duplicates with NuGEN’s “nodup.py” script (https://github.com/tecangenomics/nudup). Only CpG sites with methylation coverage of >=10 in all samples were included for downstream analysis, and normalization was performed separately in each tissue. Individual CpG sites were divided into 500 base-pair windows and were clustered using the Markov Clustering algorithm via the MCL R package. To apply MCL, for each 500 base-pair window an undirected graph was constructed, linking individual sites if their correlation was >=0.7. The resulting sites/clusters were used as input for normalization and differential analysis with edgeR⁹⁸. To generate normalized sample-level data, the methylation coverages of filtered sites/clusters were first log2-transformed, and normalization was performed using preprocessCore’s quantile normalization preprocessCore::normalize.quantiles.robust. As all samples for a given tissue were processed in a single batch, batch correction was not possible or necessary.

ATAC-seq

For ATAC-seq (assay for transposase-accessible chromatin using sequencing), nuclei from aliquoted tissue samples were extracted using the Omni-ATAC protocol with modifications⁹⁹. For all tissues, the homogenate passed through a 40 μm cell strainer to collect nuclei, that were then stained with DAPI and counted. 50,000 nuclei (or max. 500 μL nuclei) was mixed with 1 mL ATAC-RSB buffer and spun at 1000×g for 10 minutes. The nuclei pellet was resuspended in 50 μL of transposition mixture and incubated at 37°C for 30 minutes with 1000 rpm shaking. The transposed DNA was subsequently purified using the Qiagen MinElute Purification kit (Qiagen # 28006). The DNA product was amplified using custom indexed primers and the NEBnext High-Fidelity 2x PCR Master Mix (NEB, M0541L), and cleaned with the 1.8x SPRIselect beads prior to sequencing. Pooled libraries were sequenced using 50bp paired-end sequencing on an Illumina NovaSeq 6000 platform (Illumina, San Diego, CA, USA) targeting a depth of 35 million read pairs per sample.Reads were demultiplexed with bcl2fastq2 (v2.20.0) and processed with the ENCODE ATAC-seq pipeline (v1.7.0) (https://github.com/ENCODE-DCC/atac-seq-pipeline)¹⁰⁰. Adapter-trimming was performed using cutadapt (v2.5), and reads were aligned to release 96 of the Ensembl Rattus norvegicus (rn6) genome using Bowtie 2 (v2.3.4.3). Duplicate reads and reads mapping to the mitochondrial chromosome were removed. Signal files and peak calls were generated using MACS2 (v2.2.4). Optimal peaks from all workflows were concatenated, trimmed to 200 base pairs around the summit, and sorted and merged with bedtools v2.29.0 to generate a master peak list. This peak list was intersected with the filtered alignments from each sample using bedtools coverage with options -nonamecheck and -counts to generate a peak by sample matrix of raw counts. The remaining steps were applied separately on raw counts from each tissue. Peaks from non-autosomal chromosomes were removed, as well as peaks that did not have at least 10 read counts in four samples. Filtered raw counts were then quantile-normalized with limma-voom. This version of the normalized data was used for downstream analyses. Batch correction was performed by including the sample processing batch, coded as a factor, as a covariate during differential analysis.

Proteomics

The proteomic assays consisted of global proteomics and several post-translational modifications (phosphoproteome, acetylome and ubiquitylome). While phosphorylation and ubiquitination are known to be largely transient modifications¹⁰¹, the effect on the balance between modified and non-modified residues in response to training is unknown, motivating the investigation of these modifications in combination with overall protein abundance.

The BCA assay (ThermoFisher) was used to determine protein lysate concentrations. After reduction in 5 mM dithiothreitol (DTT, Sigma-Aldrich) for 1 hour at 37 °C under 1000 rpm mixing, the lysate was alkylated with iodoacetamide (IAA, Sigma-Aldrich) in the dark for 45 minutes at 25 °C under 1000 rpm mixing. All samples were diluted 1:4 with 50 mM Tris-HCl, pH 8.0 and LysC endopeptidase (1mAU/μL, Wako Chemicals) was added at a 1:50 enzyme:substrate ratio, followed by 2 hours of digestion at 25°C and 850 rpm mixing. The samples were then trypsinized at an enzyme:substrate ratio of 1:50 (1:10 for white adipose tissue) for 14 hours at 25 °C and 850 rpm mixing. Digestion was terminated by the addition of formic acid (FA) to reach a 1% total concentration, followed by centrifugation for 15 min at 1500×g at 4 °C. The resulting supernatant was first diluted with 0.1% FA to 3 mL total volume and desalted using Sep-Pac C18 SPE cartridges (Waters), and then concentrated using a speedvac prior to a new final concentration measurement through BCA assay.

Protein samples (400 μg) were first dried, and then resuspended in 200 mM HEPES pH 8.5 (final concentration 5 μg/μL) for TMT labeling. Samples were randomized into ten TMT channels, with the last channel (total 11 plexes) containing a common reference aliquot consisting of pooled peptides from each experimental sample. The TMT reagents and peptide aliquots were combined to reach a 1:1 peptide:tag ratio and mixed for 1 hour at 2 5°C, 400 rpm. Samples were then diluted to 2.5 μg/μL (in 20% acetonitrile). After QC checks, reactions were quenched with hydroxylamine and samples from each multiplex were combined, concentrated in a speedvac, and desalted using Sep-Pac C18 SPE cartridges (Waters). Heart and liver tissues were also subjected to phosphotyrosine enrichment as previously described¹⁰². Additional details on the processing have been published elsewhere¹⁹.

After fractionation by high pH reversed phase separation, samples were resuspended in mobile phase A, centrifuged to remove any debris, and then loaded onto a column. The flow rate through the column was set to 1 mL/min (using mobile phase B), and samples were eluted for 96 minutes, resulting in the collection of 96 fractions. Global proteome analysis was run on 5 % of the concatenated fractions, with the remaining 95 % used for phosphopeptide enrichment using immobilized metal affinity chromatography (IMAC). Further detail on the phosphopeptide enrichment process is provided here¹⁹. Flow through from the IMAC was further processed for acetylpeptide enrichment using an acetyl-lysine antibody (Cell Signaling Technologies, #13416). Four to 5 μL of each fraction were used for LC-MS/MS analysis for each proteomic assay.

The LC-MS/MS analysis was conducted in somewhat different ways for heart and liver, compared to gastrocnemius, white adipose, lung, kidney and cortex due to site-specific differences. For heart and liver, the online separation was performed using a nanoflow Proxeon EASY-nLC 1200 UHPLC system (Thermo Fisher Scientific) and an in-house packed 22 cm × 75 μm C18 silica picofrit capillary column used for the LC step. For global proteome, 1 μg was loaded in a 2μL volume, whereas for phosphoproteome, acetylome, and ubiquitylome, a 4-μL volume was used with 50% of each fraction sample. Mass-spectrometry analysis was conducted using a Q-Exactive Plus mass spectrometer (Thermo Fisher Scientific) for the global proteome, and a Q-Exactive HFX mass spectrometer (Thermo Fisher Scientific) for the phosphoproteome, acetylome and ubiquitylome. Specific analysis settings have been described elsewhere¹⁹. For the other tissues, analyzed at the Pacific Northwest National Laboratory, online separation was instead performed using a nanoAcquity M-Class UHPLC system (Waters), and an in-house packed, 25 cm × 75 μm C18 silica picofrit column used for the LC step. The mass-spectrometry analysis was conducted using a Q Exactive HF mass spectrometer (Thermo Fisher Scientific) for the global proteome. For the phosphoproteome, a Dionex Ultimate 3000 UHPLC direct-inject system (Thermo) was used for online separation, a 30 cm × 75 μm C18 silica picofrit column for the LC and Q-Exactive HFX mass spectrometer (Thermo Fisher Scientific) for the mass spectrometry analysis. Additional details on the sample processing is available elsewhere¹⁹. Acetylome and ubiquitylome was only analyzed in the heart and the liver.

Quantification of all proteomic analysis were calculated as log₂ TMT ratios to the common reference, with analytes not fully quantified in at least two plexes within a tissue removed. Proteomics datasets were examined for sample outliers by looking at the top principal components and by examining median protein abundance across samples. Outlier samples were identified for acetylome samples labeled with channel 130C, and these were suspected to originate from contaminated 130C-TMT reagent. All acetylome samples labeled with TMT channel 130C were excluded from downstream analysis. Proteomics features originating from non-rat contaminants were removed. Sample normalization was performed by median-centering and mean absolute deviation scaling.Plex batch effects were removed using linear models implemented by the limma::removeBatchEffect function in R (v 3.48.0). No imputation was used for the proteomic analyses. Correction for protein abundance in ubiquitylome was accomplished by fitting a global linear model between a site-specific PTM and the cognate protein and extracting the residuals. The other PTMs were not normalized to total proteome due to the lack of complete overlap between PTM and total proteome features (80.5%–89.7% for phosphosites, 95.6–96.9% for acetylsites, and 94.8–95.7% for ubiquitylsites).

Non-targeted metabolomics

Non-targeted metabolomics was performed through hydrophilic interaction liquid chromatography (HILIC) analyses at the Broad Institute of MIT and Harvard, and through reverse-phase and ion pairing profiling at the University of Michigan. For HILIC positive analyses, 10 mg of powdered tissue was homogenized in 300 μL of 10/67.4/22.4/0.018 v/v/v/v water/acetonitrile/methanol/formic acid containing stable isotope-labeled internal standards. For plasma, 10 μL mixed with 90 μL of 74.9/24.9/0.2 v/v/v acetonitrile/methanol/formic acid, was used. After centrifugation for 10 minutes at 9,000 × g, 4°C, samples were injected onto a HILIC column and eluted at 250 μL/min with mobile phase A for 30 seconds, and subsequently with a linear gradient to 40 % mobile phase B for 10 minutes, with stable elution for 4.5 minutes after that. MS analyses using electrospray ionization was conducted in the positive ion mode using full scan analysis over 70–800 m/z at 70,000 resolution and 3 Hz data acquisition rate using a Q-Exactive hybrid quadrupole Orbitrap mass spectrometer (Thermo Fisher Scientific). TraceFinder software (Thermo Fisher Scientific) was used to process raw data for targeted peak integration and Progenesis QI (Nonlinear Dynamics, Waters) for peak detection and integration of metabolites of known and unknown identity.

For reverse phase and ion-pairing analyses, 50 μL plasma was mixed with 200 μL of extraction solvent (1:1:1 v:v methanol:acetonitrile:acetone containing internal standards and homogenized through vortexing for 10 seconds. Details about internal standard concentrations are provided elsewhere¹⁹. Non-powdered solid tissues were weighed and mixed with 1:1:1:1 methanol:acetonitrile:acetone:water in a ratio of 1 mL per 50 mg tissue and homogenized using a sonicator. Subsequent steps were the same for plasma and tissues and included a 10 min incubation on ice, followed by centrifugation at 15 000 × g. The supernatant (300 μL for tissues and 150 μL for plasma) was dried with a nitrogen blower and reconstituted in water:methanol (8:2 v:v) for the LC-MS analysis. The reverse phase analyses were conducted on an Agilent 1290 Infinity II / 6545 qTOF MS system with a JetStream electrospray ionization source (Agilent Technologies) using a Waters Acquity HSS T3 column (Waters Corporation). Injection volume was 5 μL and flow rate 0.45 mL/min, mobile phase A water with 0.1% formic acid, and mobile phase B methanol with 0,025% formic acid. Each sample was analyzed in both the positive and negative ion mode. The ion pairing analyses were conducted using an Agilent Zorbax Extend C18 1.8 μm RRHD column, 2.1 × 150 mm ID, equipped with a matched guard column. Mobile phase A consisted of 97% water and 3% methanol, mobile phase B 100% methanol and mobile phase C 100% acetonitrile. Phase A and B also contained 15 mM tributylamine and 10 mM acetic acid. Injection volume was 5 μL, with variable flow rate and time (for details, see¹⁹). Mass spectrometry analysis was conducted in the negative ion mode. Profinder v8.0 (Agilent Technologies) was used for targeted compound detection and relative quantitation, while custom scripts were used for non-targeted feature detection. Agilent Mass Profiler Pro (v8.0) and Masshunter Qualitative Analysis were used for feature alignment and recursive feature detection. Prior to normalization, features missing from >50% of all samples in a batch or >30% of QC samples were removed. Data reduction was then performed using Binner¹⁰³ and normalized using the Systematic Error Removal Using Random Forest approach¹⁰⁴. The performance of the normalization was validated using relative SD for QC samples.

Non-targeted lipidomics

Non-targeted lipidomics was performed using 10 mg of powdered tissue or 25 μL of plasma. Tissue was homogenized in 400 μL isopropanol (containing stable isotope-labeled internal standards from Avanti Polar Lipids (Alabaster)), using freeze thawing in liquid nitrogen and sonication, while plasma was only mixed with 75 μL isopropanol and internal standards. Samples were then centrifuged for 5 min at 21,100 × g and supernatants used for LC-MS on a Vanquish chromatograph with an Accucore C30 column (2.1 × 150 mm, 2.6 μm particle size), coupled to a high-resolution accurate mass Q-Exactive HF Orbitrap mass spectrometer (all from ThermoFisher Scientific). Injection volume was 2 μL, mobile phase A was 40:60 water:acetonitrile with 10 mM ammonium formate and 0.1% formic acid, and mobile phase B 10:90 acetonitrile:isopropyl alcohol, with 10 mM ammonium formate and 0.1% formic acid. Details on the gradient program are provided elsewhere¹⁹. Full mass spectrometry data was acquired with 240,000 resolution over the 150–2000 m/z range. Raw LC-MS data was processed using Compound Discoverer V3.0 (ThermoFisher Scientific). Peak area was corrected for QC sample peak areas across the batch, followed by filtering with background and QC filters. Features absent in >50% of the QC pooled injections and a coefficient of variance <30 % were removed. Annotations were based on mass and relative abundance, retention time and MS2 pattern.

Targeted metabolomics and lipidomics

Targeted assays were performed for branched-chain keto acids, acyl CoA’s and nucleotides at Duke University, and for amino acids and amino metabolites, TCA cycle metabolites, ceramides and acylcarnitines at the Mayo Clinic. Targeted lipidomics was performed at Emory University. Ten μL plasma and 200 μL tissue homogenate was used for analysis of branched-chain keto acids, which were extracted using ethyl acetate as described previously¹⁰⁵. Five hundred μl tissue homogenate was used for acyl CoA extraction, as reported previously for liquid phase¹⁰⁶ and solid phase¹⁰⁷ extraction. Nucleotides were extracted as previously described^108,109. Samples were subsequently centrifuged at 14,000 × g for 5 minutes and supernatant used for LC–MS/MS. Extracts for branched-chain keto acids, acyl CoA’s and nucleotides were analyzed on a Xevo TQ-S triple quadrupole mass spectrometer (Waters) and endogenous levels were quantified using calibrators by spiking tissue homogenates (Acyl CoA’s and nucleotides) or fetal bovine serum (keto acids) with authentic analytes (all from Sigma-Aldrich).

For targeted amino acid and amino metabolite profiling, 20 mL of plasma or 5 mg of tissue homogenate was utilized. Extraction and analysis was conducted as previously described^110,111. Targeted profiling of ceramides and sphingolipids, using 25 μL plasma or 5 mg of tissue homogenate, has also been previously reported^112,113, as have the targeted profiling of acylcarnitines^114,115 (specifically C0-C18:1), which was also performed using LC-MS/MS from the same amount of plasma and tissue. TCA metabolites were profiled from 5 mg of tissue or 50 μL plasma using gas chromatography mass-spectrometry, as previously described^115,116 with minor modifications reported elsewhere ¹⁹.

Targeted lipidomics was performed on 10mg of powdered tissue, homogenized in 100 μL PBS, followed by dilution with 100 μL 20 % methanol and spiked with 1 % BHT solution according to previously used methods^117,118. After centrifugation at 14,000 rpm for 10 min, the supernatant was transferred to 96-well plates and loaded onto C18 SPE columns and eluted with 400 μL methyl formate. External standards were all purchased from Cayman Chemical. An ExionLC (SCIEX) chromatograph with an AccucoreTM C18 column (ThermoFisher), and a SCIEX QTRAP 5500 mass spectrometer was used for LC-MS/MS. Mobile phase A was water with 10mM ammonium acetate, and mobile phase B acetonitrile with 10 mM ammonium acetate. Details on the gradient program and subsequent mass spectrometry is provided elsewhere¹⁹. Raw data was processed using Sciex OS (AB SCIEX, Version 1.6.1).

Metabolomics and lipidomics data processing and normalization

All metabolomics datasets were partitioned into named compounds for analytes that were confidently identified, and unnamed for those without a standard chemical name. Only named metabolites were included in this analysis. For both untargeted and targeted platforms, measurements were log2-transformed, negative values were converted to missing values, and features with >20% missing values were removed. For untargeted datasets and targeted datasets with more than 12 features, remaining missing values were imputed using K-nearest neighbors (k=10 samples). Missing values in targeted datasets with 12 or fewer features were not imputed. For outlier identification in each dataset, we calculated each sample’s median correlation value against the other N-1 samples and selected a threshold to designate outliers as those with below-threshold median correlation values. All outliers were reviewed by Metabolomics chemical analysis sites, and only 21 confirmed technical outliers in the untargeted datasets were removed. To normalize untargeted datasets, we median-centered samples if neither sample medians nor upper quartiles were significantly associated with sex or sex-stratified training group (Kruskal-Wallis p-value < 0.01). Targeted datasets were not normalized as they were quantified using absolute concentrations. Metabolites that were identified by two or more platforms (1116 metabolites in total) were integrated using meta-regression (R’s metafor package). This analysis provided an overall estimate of fold changes and measurements of the heterogeneity among the different platforms, see¹⁹ for details.

Reannotation of cardiolipin data

In comparison to the initial dataset¹⁹, cardiolipin data was re-annotated after improvements to the spectral libraries were introduced. We revisited the annotations to capture more compounds in the class across tissues compared to the initial dataset.

mtDNA quantification

The protein/DNA precipitate resulting from the organic extraction for metabolomics was dried under vacuum for one hour in a GeneVac EZ-2 evaporator. Dried samples were stored at −20°C until DNA extraction. DNA was extracted from dried pellets using a Qiagen DNeasy Blood and Tissue kit (Qiagen, Germany, #69506). The DNA concentration in the eluate was determined and 50 ng of sample DNA was used for multiplex qPCR Mitochondrial levels were quantified following the protocol of Nicklas et al¹¹⁹ using a 5’VIC reporter and a 3’TAMRA quencher dye and D-loop expression with a 5 ‘6-FAM reporter and 3’ TAMRA-labeled quencher. Amplification was carried in a 25μL reaction consisting of 1x TaqMan Universal Master Mix II (ThermoFisher, MA), using 200 nM each β-actin forward (GGGATGTTTGCTCCAACCAA) and reverse primers (GCGCTTTTGACTCAAGGATTTAA) to estimate nuclear DNA and 50 nM each mitochondrial D-loop forward (GGTTCTTACTTCAGGGCCATCA) and reverse (GATTAGACCCGTTACCATCGAGAT) primer and 100 nM each B-actin probe (VIC-CGGTCGCCTTCACCGTTCCAGTT-TAMRA) and D-loop probe (6FAM-TTGGTTCATCGTCCATACGTTCCCCTTA-TAMRA). Forty cycles of amplification were performed on duplicate samples and relative mitochondrial levels calculated as C_T(mito) - C_T(nuclear) using the 2^−ΔΔCT method¹²⁰.

HSD17B10 knockdown experiments

HepG2 cells were transfected using Lipofectamine RNAiMAX (ThermoFisher Scientific) with either 20nM siRNA targeting HSD17B10 (ThermoFisher Scientific #107427) or 20nM Silencer Select negative control siRNA (ThermoFisher Scientific) when replated. All experiments were performed 4 days after transfection.

Total RNA was extracted using the Trizol method (Invitrogen) and reverse transcribed to cDNA using the High Capacity cDNA Reverse Transcription kit (ThermoFisher Scientific). Quantitative real-time PCR (Applied Biosystems ViiA 7 system) was used to quantify HSD17B10, GAPDH and TBP expression (primers from ThermoFisher Scientific, HSD17B10 Hs00189576_m1, GAPDH Hs02786624_g1, TBP Hs00427620_m1). For protein extraction, cells were washed once in PBS and incubated for 10 minutes with ice-cold Pierce RIPA buffer (ThermoFisher Scientific) supplemented with 1% SDS and 2x Halt^™ protease and phosphatase inhibitor cocktail (ThermoFisher Scientific). Protein concentration was assessed using the Pierce BCA protein assay kit (ThermoFisher Scientific). For Western blot, protein homogenates were separated on gradient TGX gels (Bio-Rad, #4568084) and blotted onto PVDF membranes. After blocking with Odyssey TBS blocking buffer (Li-Cor, #927–50000), the membranes were incubated with primary antibody (ERAB/HSD17B10: ab167410, ACTN2: ab68167, Abcam; OXPHOS human antibody cocktail: ab110411, Abcam; TBP: 1TBP18, ThermoFisher Scientific). After washing in TBST, membranes were incubated with the corresponding IRDye fluorescent secondary antibody (Li-Cor). Immuno-complexes were imaged and quantified using the Li-Cor Odyssey Fc imaging system (Li-Cor).

Mitochondrial oxygen consumption rate was assessed using the Agilent Seahorse according to the manufacturer’s specifications. In brief, cells were incubated in Seahorse Media (Agilent) supplemented with 10 mM glucose, 2 mM glutamine, and 1 mM pyruvate one hour prior to the experiment. Oligomycin (1.5 μM), Carbonyl cyanide-4 (trifluoromethoxy) phenylhydrazone (FCCP 1μM) and Rotenone/antimycin were diluted in the supplemented Seahorse media were added sequentially. Respiratory capacity variables were calculated from the final measurement prior to addition of the next supplement.

Quantification and statistical analyses

Differential analysis

Differential analysis was conducted in each tissue, -ome and sex separately. The full details on the processing has been described elsewhere¹⁹. Briefly, DESeq2¹²¹ was used for RNA-Seq, edgeR¹²² for RRBS, and limma¹²³ for proteomics, metabolomics, and ATAC-seq data. As input, normalized data were used for proteomics, and ATAC-seq, and filtered raw counts were used for RNA-seq, and RRBS. For targeted metabolomics, the KNN-imputed (if it included > 12 analytes) log₂-transformed data were used; otherwise, log₂-transformed data were used. For untargeted metabolomics, log₂ KNN-imputed data were used. F-tests (limma, edgeR::glmQLFTest) or likelihood ratio tests (DESeq2::nbinomLRT, lrtest) were used to identify analytes that changed over the 8-week training time-course. Male- and female-specific p-values were combined using Fisher’s sum of logs meta-analysis into a single p-value (training p-value) and p-value adjustment was performed using Independent Hypothesis Weighting (IHW)¹²⁴, with IHW FDR ≤ 5% for each ome across tissues. Time- and sex-specific effects were calculated by comparing each training timepoint with its sex-matched sedentary control animals (timewise summary statistics) using the following functions: limma::contrasts.fit with limma::eBayes, DESeq2::DESeq and edgeR::glmQLFTest. Assay-specific covariates were included based on technical metrics (RNA integrity number, median 5’−3’ bias, percent of reads mapping to globin, and percent of PCR duplicates as quantified with Unique Molecular Identifiers (UMIs) for RNA-Seq; fraction of reads in peaks and library preparation batch for ATAC-seq). For metabolomics data, meta-regression of the 1116 metabolites was performed using R’s metafor package¹²⁵.

MitoCarta 3.0³ was used to select mitochondria-associated genes (RNAseq, RRBS and ATAC-seq) and proteins (global proteomics, phosphoproteomics, acetylome and ubiquitylome). The MitoCarta list included 1136 human and 1140 mouse genes identified through mass-spectrometry on isolated mitochondria from 14 different tissues, in combination with GFP-tagging for localization, integration with other datasets and literature curation. This resource also assigns genes to a mitochondria-specific ontology, the MitoCarta 3.0 mitochondrial pathways (149 in total)³. The rat ortholog mapping of MitoCarta that was used to select mitochondrial IDs is provided in Table S7. Metabolite selection was based on Heden et al.²⁷, see Table S6 for the full list of metabolites used.

Graphical clustering analysis

Graphical clustering analysis of the timewise summary statistics was described in¹⁹. In this work the graphical representation results were filtered down to represent the MitoCarta analytes (e.g., for Fig. 3C, Table S3). We now briefly describe this graphical method. Z-scores of the IHW-selected analytes were modeled using a mixture distribution to separate null from non-null cases and identify clusters while accounting for correlation over time and between the sexes. Let $Z\in R^{nxtx2}$ represent the input, where ${\mathrm{z}}_{\mathrm{i},\mathrm{j},\mathrm{k}}$ represents the z-score of analyte $i\in \{1,\dots ,\mathrm{n}\}$ at the training timepoint $j\in \{1,\dots ,\mathrm{t}\}$ of sex $k\in \{\mathrm{m},\mathrm{f}\}$, with m=males and f=females. Under the assumption that ${\mathrm{z}}_{\mathrm{i},\mathrm{j},\mathrm{k}}$ follows a mixture distribution of null and non-null z-scores, each ${\mathrm{z}}_{\mathrm{i},\mathrm{j},\mathrm{k}}$ has a latent configuration ${\mathrm{h}}_{\mathrm{i},\mathrm{j},\mathrm{k}}\in \{-\mathrm{1,0},1\}$, where −1 denotes downregulation, 0 denotes null (no change), and 1 denotes upregulation. A full configuration matrix (e.g., specifying if a z-score is null, up, or down for each timepoint in each sex) is denoted $\mathbf{h}\in {\{-\mathrm{1,0},1\}}^{tx2}$, and ${\mathbf{z}}_{\mathrm{i}}\in R^{tx2}$ is the matrix of all z-scores of analyte $i$. The expectation-maximization (EM) process of the repfdr algorithm^28,126 was used to estimate for each possible $\mathbf{h}$ both its prior probability $\mathrm{\Pi }\left(\mathbf{h}\right)$ and its posterior $\mathrm{P}\mathrm{r}\left(\mathbf{h}\mid {\mathbf{z}}_{\mathrm{i}}\right)$, for every analyte $i$. The locfdr $\mathrm{R}$ package¹²⁷, is used in this process to infer the marginal mixture distribution of each timepoint $j$ and sex $k$. That is, all z-scores (i.e., not limited to mitochondrial analytes) ${\mathbf{z}}_{{\mathrm{*}}_{\mathrm{j}\mathrm{k}}}$ are used to estimate the densities: $f_{j,k}\left(z\mid H_{i,j,k}=-1\right)=f_{-1,j,k}\left(z\right),f_{j,k}\left(z\mid H_{i,j,k}=0\right)=N\left(\mathrm{0,1}\right)$, and $f_{j,k}\left(z\mid H_{i,j,k}=-1\right)=f_{1,j,k}\left(z\right)$. We excluded configurations $\mathbf{h}$ with $\mathrm{\pi }\left(\mathbf{h}\right)<0.001$ and normalized $\mathrm{P}\mathrm{r}\left(\mathbf{h}\mid {\mathbf{z}}_{\mathrm{i}}\right)$ to sum to 1. The new posteriors can be interpreted as a soft clustering solution, where the greater the value is, the more likely it is for analyte $i$ to participate in cluster $\mathbf{h}$.

We use these posteriors to assign analytes to “states”, where a state is a tuple ${(\mathrm{s}}_{\mathrm{m},\mathrm{j}},{\mathrm{S}}_{\mathrm{f},\mathrm{j}})$, where ${\mathrm{S}}_{\mathrm{m},\mathrm{j}}$ is the differential abundance state null, up, or down (0,1, and −1 in the notation above, respectively) in males at timepoint $j({\mathrm{s}}_{\mathrm{f},\mathrm{j}}$ corresponds to females at timepoint $j$), resulting in nine possible states in each timepoint. For example, assume we inspect analyte $i$ in timepoint $j$, asking if the abundance is upregulated in males while null in females, then we sum over all posteriors $\mathrm{P}\mathrm{r}\left(\mathrm{h}\mid {\mathrm{z}}_{\mathrm{i}}\right)$ such that ${\mathrm{h}}_{\mathrm{m},\mathrm{j}}=1$ and ${\mathrm{h}}_{\mathrm{f},\mathrm{j}}=0$. If the result is greater than 0.5, then we assign analyte $i$ to the node set $\mathrm{S}\left({\mathrm{s}}_{\mathrm{m},\mathrm{j}},{\mathrm{s}}_{\mathrm{f},\mathrm{j}}\right)$. We use $\mathrm{S}\left({\mathrm{s}}_{\mathrm{m},\mathrm{j}},{\mathrm{S}}_{\mathrm{f},\mathrm{j}}\right)$ to denote all analytes that belong to a state $\left({\mathrm{s}}_{\mathrm{m},\mathrm{j}},{\mathrm{s}}_{\mathrm{f},\mathrm{j}}\right)$ and for every pair of states from adjacent time points $j$ and $j+1$ we define their edge set $\mathrm{E}\left({\mathrm{s}}_{\mathrm{m},\mathrm{j}},{\mathrm{S}}_{\mathrm{f},\mathrm{j}},{\mathrm{S}}_{\mathrm{m},\mathrm{j}+1},{\mathrm{S}}_{\mathrm{f},\mathrm{j}+1}\right)$ as the intersection of $\mathrm{S}\left({\mathrm{s}}_{\mathrm{m},\mathrm{j}},{\mathrm{S}}_{\mathrm{f},\mathrm{j}}\right)$ and $\mathrm{S}\left({\mathrm{s}}_{\mathrm{m},\mathrm{j}+1},{\mathrm{S}}_{\mathrm{f},\mathrm{j}+1}\right)$. Note that these can be defined using a similar marginalization as was done to define the node sets, but in practice we found that these two approaches resulted in almost identical results. The sets $\mathrm{S}$ and $\mathrm{E}$ together define a tree structure that represent the differential patterns over sex and time.

Enrichment analyses and pathway annotation

Analytes were mapped to Ensembl gene IDs. For each identified analyte set (e.g., a node or an edge set from the graphical clustering above), we performed pathway enrichment analysis using the full MitoCarta 3.0 gene list as background. Enrichment analysis was also performed using the KEGG and REACTOME rat pathways (organism “rnorvegicus”) using the gprofiler2::gost function in R¹²⁸. Nominal p-values were calculated using a one-tailed hypergeometric test, and were then adjusted across all results using IHW with tissue as a covariate. Pathways with a q-value < 0.1 were considered significant. Fig. 2C displays significantly enriched pathways (TRNSCRPT, PROT, ACETYL) for the 8-week node for all tissues with at least one enriched pathway (<5% BH FDR), the top pathway from each MitoCarta subcategory with the greatest number of enrichments in shown. All enrichment results are available in Table S2.

To identify which mitochondrial pathways were most differentially regulated within each tissue across timepoints and within each sex, we used our identified repfdr sets (see previous section) together with the MitoCarta 3.0 human database. Here, each significant analyte was annotated to one of 8 major MitoCarta mitochondrial pathway groups including metabolism, mitochondrial central dogma, mitochondrial dynamics and surveillance, OXPHOS, protein import, sorting and homeostasis, signaling, small molecule transport and “other”. Analytes (proteins or genes) within each pathway were then separated by timepoint and differential regulation by sex according to state score. For example, for a given timepoint we can assign analytes to a pattern that represents upregulation in both males and females (denoted as F1_M1), or upregulation in females and downregulation in males (denoted as F1_M-1). This representation was used to display the number of time-dependent and sex-specific differentially regulated features in each pathway.

Differential analysis adjusted for biomarkers

To determine which of the differential analysis results could be explained by changes in markers of mitochondrial volume we repeated the differential analysis in eight tissues with adjustment for: (1) cardiolipin measured in six tissues (heart, kidney, liver, lung, SKM-GN, and WAT-SC), and (2) percent of mitochondrial RNA-seq reads, as an mtDNA proxy in adrenal, BAT, SKM-GN, and SKM-VL. The reason for using mitochondrial RNA-seq reads instead of mtDNA was that it covered all timepoints, while mtDNA was only measured at 3 out of 5 timepoints.

For timewise results, we used the pre- and post-adjustment models for each analyte and extracted the timewise z-scores, comparing each timepoint to its sex-matched controls. Most differential analysis results did not change substantially after adjustment, with only 799 out of the 2,167 analyzed analytes (37%) producing a difference of three or greater in at least one of the timepoints. Clustering analysis of these 799 z-score trajectories identified four clusters using the k-means algorithm (Fig. S4B). Composition of the clusters by ome and tissue is available in Fig. S4C.

Principal Component Analysis

Principal Component analysis with scaling was performed on the sedentary 8-week control animals from the RNA-seq and proteomics data separately. The first two components contributed to the majority of the variance in both the datasets with the first component revealing clear tissue-specific differences.

Biological network analyses

Pathway and protein interaction networks of a selected set of genes were created using GeneMANIA¹²⁹ and visualized using the Cytoscape software¹³⁰.

Disease ontology enrichment analysis

We first filtered the disease ontology database before applying the enrichment analyses. Our rationale here was that many disease terms may be enriched with general biological processes that are relevant for many tissues both in health and disease states (e.g., cell proliferation in cancer disease terms), and are thus not likely to reflect a true association between our exercise-specific results and diseases. We therefore generated tissue-specific disease ontology terms by utilizing gene expression data from GTEx v8¹³¹. For each disease ontology term and a tissue (covered by GTEx) we computed the p-value for the overlap between the term’s gene set and the tissue’s gene set. If the p-value was greater than 0.01 then we omitted the term from the tissue’s analyses. Disease ontology enrichment analysis was then performed using the DOSE R package⁹⁸ for each of our tissue- and ome-specific gene sets that had at least 10 genes. For our mitochondrial-focused analysis we limited the disease genes to mitocarta genes only. The results are available in Table S5.

Gene and PTM set enrichment analyses

Gene set enrichment analysis (GSEA) and post-translational modification set enrichment analysis (PTM-SEA) was performed using ssGSEA2.0¹³². The input for GSEA was the t-scores from the timewise comparisons for all analytes (not just mitochondrial). Here, analytes were integrated into gene ids by taking the most significant t-score (i.e., the one with the maximum absolute value). Phosphosite-level t-scores and the human PTMSigDB¹³² were used as input for PTM-SEA. We used the MitoPathways database from MitoCarta 3.0³ to identify enriched mitochondrial pathways. Human gene symbols were mapped to rat orthologs before running the analysis. We used the NCBI Reference Protein Sequence database (RefSeq) to annotate protein IDs, and mapped PTM sites from rats to humans using BLASTp to align rat sequences to the human UniProt fasta sequence database, and used alignments with >60% sequence identity for mapping. For both GSEA and PTM-SEA, we ran the ssGSEA2 function with parameters that avoid normalization, required at least 5 overlapping features with the gene set, and used the area under the curve as the enrichment metric (sample.norm.type = “none”, weight=0.75, correl.type = “rank”, statistic = “area.under.RES”, output.score.type = “NES”, min.overlap=5).

DREM

We used DREM³² for network inference of transcription factors (TFs) driving the transcriptional changes in specific sex and tissue combinations across the 8-week training time-course. Selecting for MitoCarta genes, we used the z-score per gene in each timepoint as input. For TF-target data we used the network inferred by NicheNet¹³³. To use the NicheNet network, we mapped the rat gene symbols to human symbols using data from RGD (v39). We used a low penalty for adding nodes (40) and a convergence likelihood of 0.01%. We also ran the models using all the transcriptome results to confirm that the predicted TFs were identified using the entire transcriptome as well, with the appropriate background.

Comparison to disease datasets

Our differential proteomic results from the endurance training intervention were compared to case-control proteomic results from disease cohorts. The skeletal muscle results were compared to two human skeletal muscle T2D cohorts^134,135, the liver results were compared to human liver datasets for NASH, cirrhosis¹³⁶ and NAFLD¹³⁷, as well as a mouse dataset on obesity¹³⁸. The cardiac results were compared to a human cardiac proteomic study on hypertrophic cardiomyopathy¹³⁹, and the effects of myocardial infarction¹⁴⁰ and heart failure¹⁴¹ in rat cardiac muscle. For each comparison of rat vs. human results we first subsetted the data to the protein ids shared by both platforms. That is, we set the background for the comparison to the set of proteins that were quantified (i.e., not necessarily significantly differential) in our platforms and that of the compared human study (when available). Then, we computed the significance of the overlap between the human study reported significant protein ids and our IHW-selected protein ids via Fisher’s exact test. Among the proteins that were in this overlap, we again tested for directionality of the effects using Fisher’s exact test, but with the alternative being of overlap lower than expected. For example, if 10 proteins were identified as significant in both the human study and our rat study, then we annotated each one by their sign of differential abundance as up/down (week 8 results from the rat, performed for each sex separately). Then, we applied Fisher’s exact test for the null hypothesis that the sign concordance between the two resources is random, and the alternative that the discordance is greater than expected by chance. The full output statistics and the overlapping sets are available through our GitHub repository, which also contains the data from the analyzed studies above (see the dea/compare_to_external_disease_datasets R notebook and the output tables in the data directory).

Supplementary Material

3

Table S1. Training effect on biomarkers of mitochondrial volume, related to Figure 1.

4

Table S2. MitoCarta 8-week enrichment results, related to Figure 2.

5

Table S3. Clustering results for pre- and post-adjustment, related to STAR Methods.

6

Table S4. Repfdr results selected for mitochondrial analytes, related to Figures 3–6.

7

Table S5. Disease ontology results, related to Figure 7.

8

Table S6. List of metabolites used, related to STAR Methods.

9

Table S7. Mitocarta rat ortholog list, related to STAR Methods.

Key resources table.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Acetyl-lysine	Cell Signaling Technologies	#13416
ERAB/HSD17B10	Abcam	ab167410
ACTN2	Abcam	ab68167
OXPHOS human antibody cocktail	Abcam	ab110411
TBP	ThermoFisher Scientific	1TBP18
Chemicals, peptides, and recombinant proteins
Dithiothreitol	Sigma-Aldrich	3483–12-3
lodoacetamide	Sigma-Aldrich	144–48-9
LysC endopeptidase	Wako Chemicals	NC9223464
PTMScan Phospho-Tyrosine Rabbit mAb beads	Cell Signaling Technologies	#8803
LC-MS grade water	Fisher Scientific	W64; CAS:7732–18-5
LC-MS grade 2-propanol	Fisher Scientific	A461–4; CAS:67–63-0
LC-MS grade acetonitrile	Fisher Scientific	A955–4; CAS:75–05-8
Formic acid	Fisher Scientific	A117–50; CAS:64–18-6
Ammonium formate	Fisher Scientific	A11550; CAS:540–69-2
Lysophosphatidyl choline LPC (18:1(d7))	Avanti Polar Lipids	791643; CAS:2097561–13-0
Lysophosphatidyl ethanolamine LPE(18:1(d7)	Avanti Polar Lipids	791644; CAS:2260669–47-2
Phosphatidyl choline PC (15:0/18:1(d7))	Avanti Polar Lipids	791637; CAS:2097561–16-3
Phosphatidyl ethanolamine PE (15:0/18:1(d7))	Avanti Polar Lipids	791638; CAS:2097561–15-2
Phosphatidyl serine PS (15:0/18:1(d7))	Avanti Polar Lipids	791639; CAS:2260669–40-5
Phosphatidyl glycerol PG (15:0/18:1(d7))	Avanti Polar Lipids	791640; CAS:2260669–42-7
Phosphatidyl inositol PI (15:0/18:1(d7))	Avanti Polar Lipids	791641; CAS:2260669–44-9
Cholesterol ester CE (18:1 (d7))	Avanti Polar Lipids	791645; CAS:1416275–35-8
Diacylglyceride DG (15:0/18:1(d7))	Avanti Polar Lipids	791647; CAS:2097561–14-1
Triacylglyceride TG (15:0/18:1(d7)/15:0)	Avanti Polar Lipids	791648; CAS:2097561–17-4
Sphingomyelin SM (18:1(d9))	Avanti Polar Lipids	791649; CAS:2260669–50-7
Cholesterol-d7	Avanti Polar Lipids	700041; CAS:83199–47-7
Lipofectamine RNAiMAX	ThermoFisher Scientific	13778075
HSD17B10 siRNA	ThermoFisher Scientific	#107427
Silencer select negative control siRNA	ThermoFisher Scientific	4404021
Odyssey TBS blocking buffer	Li-Cor	#927–50000
Critical commercial assays
Agencourt RNAdvance blood specific kit	Beckman Coulter	A35604
Universal Plus mRNA-Seq kit	NuGEN/Tecan	# 9133
Ovation® RRBS Methyl-Seq kit	Tecan Genomics	9522-A01
Qiagen MinElute Purification kit	Qiagen	# 28006
Qiagen DNeasy Blood and Tissue kit	Qiagen	#69506
TaqMan Universal Master Mix II	Applied Biosystems	#4440043
High Capacity cDNA Reverse Transcription kit	Applied Biosystems	#4368814
Seahorse XF Cell Mito Stress Test Kit	Agilent	103015–100
Deposited data
MoTrPAC Data Hub	This paper	https://motrpac-data.org/
RNAseq, ATAC-seq and RRBS raw data	This paper	PRJNA908279
RNAseq, ATAC-seq and RRBS processed data	This paper	GSE242354
Metabolomics data	This paper	10.21228/M8V97D
Protein abundance data	This paper	MSV000092911 MSV000092922
Protein phosphorylation data	This paper	MSV000092925 MSV000092923
Protein acetylation data	This paper	MSV000092924
Protein Ubiquitination data	This paper	MSV000092931
Experimental models: Organisms/strains
Rat: Fischer 344	National Institute of Aging colony	N/A
HepG2 cells	Sigma Aldrich	85011430
Oligonucleotides
β-actin forward primer: GGGATGTTTGCTCCAACCAA	This paper	N/A
β-actin reverse primer: GCGCTTTTGACTCAAGGATTTAA	This paper	N/A
B-actin probe: VIC-CGGTCGCCTTCACCGTTCCAGTT-TAMRA	This paper	N/A
D-loop forward primer: GGTTCTTACTTCAGGGCCATCA	This paper	N/A
D-loop reverse primer: GATTAGACCCGTTACCATCGAGAT	This paper	N/A
D-loop probe: 6FAM-TTGGTTCATCGTCCATACGTTCCCCTTA-TAMRA	This paper	N/A
HSD17B10 primers	ThermoFisher Scientific	Hs00189576_m1
GAPDH primers	ThermoFisher Scientific	Hs02786624_g1
TBP primers	ThermoFisher Scientific	Hs00427620_m1
Software and algorithms
R	R Development Core Team	https://www.r-project.org/
Profinder v8.0	Agilent Technologies	N/A
Mass Profiler Pro v8.0	Agilent Technologies	N/A
Sciex OS Version 1.6.1	AB SCIEX	N/A
Compound Discoverer v3.0.0.294	ThermoFisher Scientific	N/A
XCalibur v4.4.16.14	ThermoFisher Scientific	N/A
MoTrPAC RNAseq pipeline	This paper	https://github.com/MoTrPAC/motrpac-rna-seq-pipeline
MoTrPAC BIC QC	This paper	https://github.com/MoTrPAC/MotrpacBicQC
EdgeR	Bioconductor	https://bioconductor.org/packages/release/bioc/html/edgeR.html
Limma	Bioconductor	https://bioconductor.org/packages/release/bioc/html/limma.html
GeneMANIA	Montojo et al.137	https://genemania.org/
Cytoscape	Shannon et al.138	https://cytoscape.org/
ssGSEA2.0	Krug et al.140	https://github.com/broadinstitute/ssGSEA2.0
Prism v9	GraphPad software	https://www.graphpad.com/
Graphics	This paper, Biorender	https://www.biorender.com/
DREM	Schulz et al.38	http://sb.cs.cmu.edu/drem/
Other
Resource website with all omics data from this study	This paper and a related study19	https://motrpac-data.org/
All omics data from this study (R package)	This paper and a related study19	https://motrpac.github.io/MotrpacRatTraining6moData/
Associated data	This paper	https://zenodo.org/record/7459795#.ZEAw9ezMKLs
Code repository	This paper	https://github.com/MoTrPAC/motrpac-rat-training-mitochondria
Normal chow	LabDiet	5L79
Treadmill	Harvard Instruments	LE8710RTS
Body Composition Rat and Mice Analyzer	Bruker	Minispec LF90II

Highlights.

Multi-omic atlas of the mitochondrial response to exercise training in 19 rat tissues

Robust temporal differences in mitochondrial response by -omes, tissues and sex

Most dynamic responses in adrenal, brown adipose, colon, heart, liver and muscle

Protein networks upregulated by exercise are downregulated in human T2D and cirrhosis

Abbreviations

BAT

Brown Adipose Tissue

BCAA

Branched Chain Amino Acid

DREM

Dynamic Regulatory Events Miner

ETC

Electron Transport Chain

GSEA

Gene Set Enrichment Analysis

HCM

Hypertrophic Cardiomyopathy

IHW

Independent Hypothesis Weighting

mtDNA

Mitochondrial DNA

NAFLD

Nonalcoholic Fatty Liver Disease

NASH

Nonalcoholic Steatohepatitis

OXPHOS

Oxidative Phosphorylation

PTM

Post-translational Modification

RRBS

Reduced Representation Bisulfite Sequencing

SED

Sedentary control animals

SKM-GN

Gastrocnemius Skeletal Muscle

SKM-VL

Vastus Lateralis Skeletal Muscle

T2D

Type 2 Diabetes

TCA

Tricarboxylic Acid Cycle

TF

Transcription Factor

WAT-SC

White Adipose Tissue - Subcutaneous

Secondary author list

Joshua N. Adkins, Jose Juan Almagro Armenteros, Mary Anne S. Amper, Dam Bae, Marcas Bamman, Nasim Bararpour, Jerry Barnes, Bryan C. Bergman, Daniel H. Bessesen, Nicholas T. Broskey, Thomas W. Buford, Steven Carr, Toby L. Chambers, Clarisa Chavez, Roxanne Chiu, Natalie Clark, Gary Cutter, Charles R. Evans, Edziu Franczak, Nicole Gagne, Yongchao Ge, Krista M. Hennig, Joseph A. Houmard, Kim M. Huffman, Chia-Jui Hung, Chelsea Hutchinson-Bunch, Olga Ilkayeva, Bailey E. Jackson, Catherine M. Jankowski, Christopher A. Jin, Neil M. Johannsen, Daniel H Katz, Hasmik Keshishian, Wendy M Kohrt, Kyle S. Kramer, William E. Kraus, Bridget Lester, Jun Z. Li, Ana K. Lira, Adam Lowe, DR Mani, Gina M. Many, Sandy May, Edward L. Melanson, Samuel G. Moore, Kerrie L. Moreau, Nicolas Musi, Daniel Nachun, Venugopalan D. Nair, Christopher Newgard, German Nudelman, Paul D. Piehowski, Hanna Pincas, Wei-Jun Qian, Tuomo Rankinen, Blake B. Rasmussen, Eric Ravussin, Jessica L. Rooney, Scott Rushing, Mihir Samdarshi, James A. Sanford, Irene E. Schauer, Stuart C. Sealfon, Kevin S. Smith, Gregory R. Smith, Michael Snyder, Cynthia L. Stowe, Jennifer W. Talton, Christopher Teng, Anna Thalacker-Mercer, Russell Tracy, Scott Trappe, Todd A. Trappe, Mital Vasoya, Nikolai G. Vetr, Elena Volpi, Michael P. Walkup, Laurens Wiel, Si Wu, Zhen Yan, Jiye Yu, Elena Zaslavsky, Navid Zebarjadi, Jimmy Zhen