Extracellular matrix (ECM) homeostasis is essential for normal tissue function, and its disruption by iatrogenic injury, trauma, or disease results in fibrosis. Skin ECM homeostasis is maintained by a complex process that involves an integration of cytokine and environmental mediators. However, it is unclear, in both normal and disease states, how these multifactorial processes converge to shift ECM homeostasis towards accumulation or degradation. Here we show a consistent downregulation in fatty acid oxidation (FAO) and upregulation of glycolysis in fibrotic skin and in normal skin with abundant ECM. Perturbation of glycolysis and FAO pathway enzymes reveals their reciprocal effects in ECM upregulation and downregulation, respectively. Increasing peroxisome proliferator-activated receptor (PPAR) signalling, an inducer of the FAO pathway, generates a catabolic fibroblast phenotype characterised by inhibition of ECM transcription and enhanced ECM internalization and lysosomal degradation. In contrast, suppression of glycolysis inhibits ECM gene transcription and protein levels, independently of an intact FAO pathway or PPAR signalling. Moreover, we show that CD36, a multifunctional fatty acid transporter, connects the metabolic state of fibroblasts with their capacity for ECM regulation, as internalization and degradation of collagen-1 is abrogated in fibroblasts lacking CD36. Finally, restoring FAO and upregulating CD36 reduces ECM accumulation in murine skin fibrosis. These findings indicate that metabolic perturbation of ECM homeostasis may have broad implications for therapies aimed at ECM regulation, such as fibrosis, regenerative medicine, and ageing.


Skin is the largest organ of the human body, and 70% of its dry weight is comprised of extracellular matrix (ECM). The majority of ECM is found within the dermis and is primarily composed of collagen-1 and protein components such as fibronectin and proteoglycans. Fibroblasts are the predominant mesenchymal cell type and the main effector of ECM homeostasis, mediating its continuous turnover.

Dysregulation of ECM homeostasis leads to diseased states. Excess ECM accumulation is a hallmark of fibrotic diseases such as keloid or hypertrophic scarring, scleroderma, and fibrosis induced by medication, surgery, or radiotherapy. The cascade of events that establish fibrosis is complex and remains incompletely understood. It involves multiple mediators acting through an interactive network of signalling pathways to dysregulate ECM homeostasis1. Given that metabolism is downstream and necessary for all cellular functions, divergent upstream signalling pathways may converge at key metabolic alterations to ultimately regulate phenotype. In tumour biology, pathways that promote tumour growth merge to induce a reliance on glycolysis, termed the Warburg Effect, which provides anabolic metabolites necessary for proliferation2. In inflammation, pro-inflammatory M1 versus anti-inflammatory M2 macrophage polarization is regulated by altering glycolysis and fatty acid oxidation (FAO) metabolism3. Therefore, a similar principle may be applied to ECM regulation.

Here we show that FAO and glycolysis are key metabolic pathways that govern fibroblast behaviour and ECM homeostasis. Regulation of glycolysis and FAO in fibroblasts altered their capacity for ECM production and degradation, respectively. To connect the metabolic state of a fibroblast with its capacity for ECM regulation, CD36, a fatty acid transporter, was identified to be a crucial mediator for collagen-1 internalization and degradation. Metabolic reprogramming using pharmacotherapy or CD36hi fibroblast transplantation reduced murine skin fibrosis.


Metabolic alterations correspond to increased ECM deposition

To uncover metabolic alterations that drive ECM regulation, genome-wide transcriptome profiling was performed on human skin fibrosis post radiotherapy and on age-matched healthy controls. Skin fibrosis is a common long-term sequela for patients with cancer treated with radiotherapy4. Hierarchical clustering of gene expression profiles revealed a significant separation between the two groups (Fig. 1a and Supplementary Fig. 1). Pathway analysis via gene set enrichment analysis (GSEA) showed that the most significantly downregulated genes in fibrosis were in catabolic pathways involved in oxidative phosphorylation, FAO, and the tricarboxylic acid (TCA) cycle (Fig. 1a and Supplementary Fig. 2). These pathways are linked, as fatty acids undergoing oxidation are ultimately catabolised in the TCA cycle through oxidative phosphorylation in the mitochondria. Conversely, the top upregulated pathways in fibrosis were associated with anabolism, such as ribosome biogenesis, spliceosome function, and messenger RNA (mRNA) processing (Fig. 1a and Supplementary Fig. 2). As glycolytic intermediates are necessary for anabolic pathways such as amino acid synthesis, we confirmed that there was a significant increase in glycolysis processes, including an upregulation in glucose transport (normalised enrichment score (NES) 2.45, FDR = 0.002) and a downregulation in gluconeogenesis (NES −2.34, FDR = 0.007) in patients with skin fibrosis.

Fig. 1: ECM accumulation in normal and fibrotic skin corresponds to a shift in metabolism from FAO to glycolysis.
Fig. 1

a, Genome-wide transcriptome analysis of normal and fibrotic human skin reveals a shift from catabolism to anabolism. Top, trichrome sections of normal and fibrotic skin. Middle, unsupervised hierarchical clustering based on the top 1,000 genes with highest variance of expression across samples. Fisher’s exact test, P = 0.00002 (two largest clusters). Bottom, top pathways ranked by NES (gene set enrichment with FDR < 0.05) altered in skin fibrosis compared to healthy controls. Reactome results are available in Supplementary Fig. 2. n = 9 healthy control samples, n = 13 skin fibrosis samples. b, Temporal analysis of genome-wide transcriptomic alterations in murine skin fibrosis. RNA-seq of skin at 0, 42, and 140 d post 40 Gy radiation. Left, temporal expression patterns with corresponding trichrome-stained skin sections above. Middle, heatmap of expression patterns. Right, over-representation pathway analysis (KEGG and Reactome) for each pattern. n = 3 animals per time point. C1, C2, and C3 are the three expression clusters. ce, Pairwise fixed reallocation randomization test performed using REST. c, PPAR signalling and glycolysis gene transcription alterations in skin fibrosis and in dPHFs induced to produce ECM in vitro. qRT–PCR of PPAR and glycolysis genes are relative to control. Left, in vivo profiling of human and murine skin fibrosis. Human, n = 13 biologically independent skin fibrosis samples, n = 9 biologically independent control samples. Murine, n = 7 biologically independent skin fibrosis samples, n = 5 biologically independent control samples. Right, in vitro profiling of dPHFs with hypoxia (2%), TGF-β1 (3 ng ml−1), and PGDF-BB (5 ng ml−1). n = 3 biologically independent samples per group. d, PPAR signalling and glycolysis gene transcription alterations in normal skin are shown for thin and dense ECM. Left, trichrome staining of thin ECM abdominal skin and dense ECM foot pad skin. Top right, qRT–PCR of dense versus thin murine skin. Bottom right, unsupervised hierarchical clustering using PPAR signalling and glycolysis genes of dense versus thin murine skin. n = 5 animals per group. e, PPAR and glycolysis gene transcription alterations in normal mesenchymal-enriched murine skin are shown. Top, mesenchymal enrichment assessed by qRT–PCR. Bottom, qRT–PCR of PPAR signalling and glycolysis genes for mesenchymal cells from dense versus thin skin. epCAM, epithelial cell adhesion molecule. n = 3 animals per group. All data are expressed as means ± s.e.m. *P < 0.05, **P < 0.01. Scale bars, 200 μm.

To further characterise metabolic dysregulation in skin fibrosis and elucidate its temporal evolution, genome-wide transcriptome profiling was performed on a murine model of radiation-induced skin fibrosis at 0, 42, and 140 d post radiation (Fig. 1b and Supplementary Fig. 3). Hierarchical clustering was performed on temporal expression patterns. Pathway analysis of the largest clusters revealed an early and sustained increase in glycolysis after radiation, which was also demonstrated functionally through increased fludeoxyglucose (FDG) uptake (Supplementary Fig. 4). FDG uptake is found to be elevated in a number of fibrotic conditions and correlates with fibrosis severity5,6. Furthermore, temporal transcriptome analysis detected a reciprocal decrease in PPAR signalling (Fig. 1b). PPAR signalling activates pathways involved in FAO and mitochondrial oxidative metabolism. Its downregulation inhibits fatty acid degradation, the TCA cycle, and oxidative phosphorylation; potentially accounting for the similar pattern of catabolic downregulation observed in human skin fibrosis.

Quantitative real-time PCR (qRT–PCR) confirmed the downregulation of PPAR signalling in human and murine skin fibrosis through profiling of transcription factors (PPARA and PPARG), their co-activators (PGC1A (also known as PPARGC1A) and PGC1B), and downstream targets acyl-CoA oxidase 1 (ACOX1) and carnitine palmitoyltransferase 1A (CPT1A), which catalyse rate limiting steps of FAO (Fig. 1c). Conversely, activation of glycolysis was observed at multiple steps, including upregulation of glucose transporter 1 (GLUT1) and GLUT3, hexokinase 1 (HK1) and HK2, and phosphofructokinase 1 (PFK1) and PFK2. In vitro studies further demonstrated that multiple mediators known to increase ECM production altered PPAR signalling and glycolysis gene expression with a pattern similar to that detected in skin fibrosis. Dermal primary human fibroblasts (dPHFs) were exposed to hypoxia, transforming growth factor β 1 (TGF-β1), or paltelet-derived growth factor-BB (PDGF-BB), which are purported to increase ECM production primarily via hypoxia-inducible factor-1α (HIF-1α), SMAD, and MEK–ERK signalling, respectively7,8. All three mediators increased the expression of pro-fibrotic proteins: collagen-1, fibronectin, and/or plasminogen activator inhibitor 1 (PAI1), an inhibitor of fibrinolysis (Supplementary Fig. 5). All three mediators also downregulated genes in PPAR signalling and upregulated genes in glycolysis, a pattern that is consistent with human and murine skin fibrosis (Fig. 1c). Functionally, TGF-β1-treated dPHFs increased glycolysis at baseline and with glucose (Supplementary Fig. 6a). TGF-β1 also lowered oxygen consumption rate with the addition of the fatty acid palmitate, and was less sensitive to etomoxir, a CPT1A inhibitor, indicating a lower level of FAO (Supplementary Fig. 6b).

Next, we asked whether these metabolic alterations in skin fibrosis may be observed in normal skin with variations in ECM accumulation. Compared with thin ECM from abdominal skin (Fig. 1d, left), murine skin obtained from the foot pad, which has thick and dense ECM, displayed a similar downregulation of PPAR signalling and increased glycolysis expression, mirroring skin fibrosis (Fig. 1d, top right). Hierarchical clustering using these metabolic genes revealed a clear segregation between skin types (Fig. 1d, bottom right). Similarly, hierarchical clustering using these metabolic genes segregated normal human skin with thin ECM from dense ECM (Supplementary Fig. 7). To confirm that this metabolic signature corresponding to ECM accumulation was observed in fibroblasts, mesenchymal cells isolated from murine foot pad displayed a downregulation of PPAR genes and an increase in glycolysis genes compared to abdominal skin mesenchymal cells (Fig. 1e). Therefore, even in normal states, there appeared to be a consistent downregulation in PPAR signalling and upregulation in glycolysis gene expression with ECM accumulation.

Metabolic perturbation affects ECM levels

To demonstrate that these metabolic alterations have a direct effect on the level of ECM proteins, we altered the metabolic state of dPHFs by inhibiting FAO and glycolysis at multiple steps of their respective pathway (Fig. 2a and Supplementary Fig. 8). Inhibition of FAO in dPHFs through CRISPR–Cas9 (clustered regularly interspaced short palindromic repeats (CRISPR)–CRISPR associated protein 9 (Cas9)) knockdown of PPARG or ACOX1, or by inhibition of CPT1A using etomoxir, resulted in an elevation in extracellular levels of fibronectin, collagen-1, and PAI1. Conversely, suppression of glycolysis through inhibition of glucose transporters using WZB117 and BAY876, or through inhibition of hexokinase by 2-deoxy-d-glucose (2-DG) or CRISPR−Cas9 knockdown of HK2, downregulated ECM protein levels.

Fig. 2: Enhancing FAO induces a catabolic fibroblast that downregulates ECM transcription and enhances ECM degradation.
Fig. 2

a, Inhibiting glycolysis or FAO downregulates or upregulates ECM proteins of dPHFs, respectively. Western blot of extracellular fibronectin, collagen-1, and PAI1. CRISPR–Cas9 knockdown (KD) of HK2, ACOX1, and PPARG. Replicates can be found in Supplementary Fig. 8. b, Pharmacogenomics identifies caffeic acid as a candidate drug to reverse metabolic alterations in skin fibrosis. Left, drugs shifting the fibrotic transcriptome towards normal (x axis); drugs shifting the transcriptome from glycolysis towards PPAR signalling (y axis). Right, top candidate drugs (top-right quadrant) ranked by LD50. c, Caffeic acid reverses TGF-β1-mediated metabolic gene expression alterations. qRT–PCR of PPAR and glycolysis genes relative to untreated control. n = 3 biologically independent samples per group. Pairwise fixed reallocation randomization test performed using the relative expression software tool (REST). d,e, Extracellular and intracellular fibronectin, collagen-1, and PAI1 western blot for dPHFs. Replicates can be found in Supplementary Fig. 11. f, SMAD3-dependent transcription. n = 3 biologically independent samples per group. Values represent Renilla luciferase activity. One-way analysis of variance (ANOVA). g, Fibrotic gene expression for TGF-β1-stimulated fibroblasts treated with caffeic acid. n = 3 biologically independent samples per group. Pairwise fixed reallocation randomization test performed using REST. h, Catabolic potential of dPHFs is regulated by FAO. Top, transcription factor EB (TFEB) expression for dPHFs relative to control (x axis); lysosome signal per cell area using LysoTracker (y axis). n = 3 biologically independent samples per group. Bottom, representative LysoTracker confocal images demonstrating lysosomal signal alterations. i, Collagen-1 co-localization with lysosomes. Top, quantification of co-localization between LysoTracker and DQ collagen-1. Bottom, representative images demonstrating co-localization alterations. n = 8 biologically independent samples per group. One-way ANOVA (left); two-tailed Student’s t-test (right). j,k, NH4Cl and chloroquine regulation of fibrotic protein levels for caffeic acid treated dPHFs. Replicates available in Supplementary Fig. 14. Ponceau and β-actin for extracellular and intracellular western blot loading controls, respectively. Drug doses: TGF-β1 (3 ng ml−1), caffeic acid (40 µM), WZB-117 (10 µM), BAY-876 (2.5 µM), 2-DG (5 mM), Etomoxir (10 µM). All data are expressed as means ± s.e.m. *P < 0.05, **P < 0.01. Scale bars, 20 μm.

Given this direct effect on ECM regulation, we asked whether pharmacological perturbation of fibroblast metabolism may reduce ECM accumulation in fibrosis. A pharmacogenomics screening approach was undertaken to screen for candidate compounds that may suppress glycolysis and upregulate FAO and for those that may restore human skin fibrosis transcriptome back to normal (Fig. 2b). Connectivity map (CMAP) is a database of transcriptome profiles from human cell lines before and after drug treatment9. Drugs in CMAP were assessed in relation to these two conditions and were further filtered on the basis of their toxicity (LD50; the dose lethal to 50% of animals tested) and known mechanisms of action. Caffeic acid, a purported PPAR agonist, emerged as a top candidate10. Caffeic acid upregulated PPAR genes and downregulated glycolysis genes in TGF-β1-treated dPHFs, and it functionally enhanced FAO and suppressed glycolysis (Fig. 2c and Supplementary Fig. 9). Caffeic acid also markedly downregulated extracellular fibronectin and collagen-1, while both intracellular and extracellular PAI1 were reduced (Fig. 2d). This effect of caffeic acid was inhibited by etomoxir, a FAO inhibitor, and was abrogated in dPHFs lacking PPARG or ACOX1, demonstrating that enhancement of FAO was necessary to reduce pro-fibrotic protein levels (Fig. 2e, Supplementary Fig. 10 and Supplementary Fig. 11). To assess whether primary glycolysis suppression may also reduce ECM accumulation, GLUT inhibition with WZB117 was performed. WZB117 reduced extracellular fibronectin, collagen-1, and PAI1 protein levels in the presence of TGF-β1, even in dPHFs lacking PPARG and ACOX1 (Supplementary Fig. 12).

Glycolysis and FAO affect ECM production and degradation

To determine the mechanism by which downregulation of glycolysis or upregulation of FAO reduces ECM accumulation, alterations to both ECM transcription and degradation were interrogated. To assess ECM transcription, a SMAD3-driven luciferase reporter was constructed. SMAD3 is a signal transducer and transcription factor activated by multiple pro-fibrogenic cytokines, and it directly induces transcription of collagen-1, fibronectin, and PAI1 (ref. 11). Caffeic acid reduced SMAD3-dependent transcription in response to TGF-β1 and significantly downregulated the expression of PAI1, whereas FN1 and COL1A1 were moderately decreased (Fig. 2f,g). This inhibitory effect on SMAD3-dependent transcription was rescued by etomoxir, which confirms that enhancing FAO reduces the transcription of ECM genes (Fig. 2f). Similarly, glycolysis inhibition with WZB117 or 2-DG also reduced SMAD3-dependent transcription, but this effect was not rescued by etomoxir (Supplementary Fig. 13), which suggests that primary suppression of glycolysis is also important and may, at least in part, act independently of FAO to downregulate ECM levels.

Interestingly, upregulation of PPAR signalling reduced extracellular, but not intracellular, collagen-1 and fibronectin (Fig. 2d). As fibroblasts are capable of ECM phagocytosis and lysosomal degradation, we proposed that enhancing PPAR signalling promotes ECM degradation12. Consistent with its role in ECM production, TGF-β1 suppressed the catabolic potential of dPHFs by downregulating the expression of transcription factor EB (TFEB), a master regulator of lysosomal biogenesis, and significantly decreased lysosome signal in dPHFs (Fig. 2h, left). This effect of TGF-β1 may be secondary to a downregulation in PPAR signalling, as fibroblasts lacking PPARG mirrored the effect of TGF-β1 (Fig. 2h, right) and caffeic acid restored TFEB gene expression and lysosome signal in TGF-β1-stimulated dPHFs (Fig. 2h, left).

To determine whether perturbation of dPHF catabolic potential altered its capacity for lysosomal ECM degradation, co-localization between DQ collagen-1, which activates fluorescence after its degradation, and lysotracker was performed (Fig. 2i). Under basal conditions, collagen-1 fluorescence was primarily detected intracellularly and co-localised with lysosomes. In the presence of TGF-β1 and with PPARG knockdown, DQ collagen-1 was found lining the cell membrane, with almost no signal detected intracellularly, leading to a significant reduction in co-localization with lysosomes. Enhancing PPAR signalling in dPHFs by caffeic acid reactivated internalization of collagen-1 in dPHFs, leading to a significant increase in co-localization between DQ collagen-1 and lysosomes. Furthermore, NH4Cl and chloroquine (two lysosomal inhibitors) were shown to inhibit the downregulation of extracellular fibronectin and collagen-1 by caffeic acid in a dose-dependent manner but with minimal effect on intracellular levels, which confirms that reactivation of co-localization upregulates lysosomal ECM degradation (Fig. 2j,k and Supplementary Fig. 14). Interestingly, PAI1 levels did not increase substantially with lysosomal inhibition, which combined with the large transcriptional changes to PAI1 expression with TGF-β1 and caffeic acid (Fig. 2g), suggests that PAI1 is more transcriptionally regulated. Taken together, the results show that enhancement of PPAR signalling in dPHFs generates a catabolic fibroblast phenotype that reduces ECM gene transcription and enhances ECM degradation.

CD36 regulates ECM degradation in fibroblasts

We next asked whether there exists a surface receptor that serves as a mediator to connect the metabolic state of a fibroblast with its capacity for ECM regulation. CD36 is a transmembrane glycoprotein that imports long-chain fatty acids intracellularly for FAO and is the only fatty acid transporter well known to bind components in the ECM13. CD36 expression was inversely correlated with ECM abundance in normal skin and was downregulated in human and murine radiation-induced skin fibrosis (Fig. 3a), all conditions that demonstrated a downregulation in FAO (Fig. 1c,d). Consistent with this, FAO suppression in dPHFs in vitro using etomoxir or with knockdown of PPARG or ACOX1 downregulated CD36 protein expression (Supplementary Figs. 15, 16). CD36 gene, total protein, and surface expression were also downregulated by TGF-β1 in dPHFs and were rescued by caffeic acid (Fig. 3b–d and Supplementary Fig. 17).

Fig. 3: CD36 is metabolically regulated in fibroblasts and crucial for collagen-1 degradation.
Fig. 3

a, CD36 gene expression in fibrotic skin and normal skin with ECM accumulation. Log2 fold change relative to control or normal thin ECM skin, for skin fibrosis and normal dense skin, respectively. All samples are biologically independent. Human, n = 9 control and thin ECM samples, n = 13 skin fibrosis samples, n = 12 dense ECM skin samples. Murine, n = 5 control animals, n = 5 skin fibrosis animals, n = 5 thin ECM samples, n = 5 dense ECM samples. Pairwise fixed reallocation randomization test performed using REST. All data are expressed as means ± s.e.m. *P < 0.05, **P < 0.01. bf, dPHFs treated with TGF-β1 (3 ng ml–1) ± caffeic acid (40 µM). All data are expressed as means ± s.e.m. *P < 0.05, **P < 0.01. b, CD36 gene expression by qRT-PCR; n = 3 biologically independent samples per group, one-way ANOVA. c, CD36 surface expression by flow cytometry; n = 3 biologically independent samples per group, Student’s t-test. d, CD36 total protein detected by ELISA; n = 3 biologically independent samples per group, one-way ANOVA. e, FAO of CD36 wild-type (WT) and knockdown (KD) dPHFs; palmitate (170 µM), etomoxir (40 µM), n = 18 biologically independent samples per group, Student’s t-test. f, CD36 knockdown abrogates the effect of caffeic acid on fibrogenic protein levels; ponceau used for loading control. Replicates are found in Supplementary Fig. 18. g, CD36 knockdown abrogates collagen-1 degradation in dPHFs. DQ collagen-1 is fluorescence quenched until its degradation. Cells were profiled by flow cytometry at 1 h (red) and 8 h (blue) after the addition of DQ collagen-1. Top, scatter plot of collagen-1 degradation (x axis) and CD36 expression (y axis) for wild-type and CD36 knockdown dPHFs. Bottom, DQ collagen-1 fluorescence intensity for wild-type and CD36 knockdown dPHFs. h, Live cell imaging of DQ collagen-1 degradation by dPHFs. In the top graph, the y axis represents fluorescence relative to cell surface area. The bottom panel shows representative images of DQ collagen-1 fluorescence over time. n = 5 biologically independent samples per group. Two-way ANOVA, P < 0.001. i, Live cell imaging of FITC collagen-1 degradation by PHFs. FITC collagen-1 degradation results in a decrease in fluorescence. The top graph represents quantification of fluorescence relative to time, starting at 60 min after the addition of FITC. Quantification was performed by measuring fluorescence intensity normalised to cell area. The bottom panel shows representative images of FITC collagen-1 binding to a cell (white arrow). n = 3 biologically independent samples per group. Two-way ANOVA, P < 0.001. ELISA, enzyme-linked immunosorbent assay. Scale bars, 20 μm.

As CD36 expression correlated with the FAO–PPAR signalling status of fibroblasts, we asked whether CD36 may directly promote ECM degradation. CD36 knockdown using CRISPR–Cas9 significantly reduced the capacity of dPHFs to upregulate FAO in response to palmitate (Fig. 3e) and inhibited the capacity of caffeic acid to suppress TGF-β1-mediated upregulation of extracellular collagen-1 and fibronectin (Fig. 3f and Supplementary Fig. 18). To determine whether the degradation of extracellular collagen-1 was impaired by CD36 knockdown, DQ collagen-1 was used. As assessed by flow cytometry, the majority of wild-type dPHFs were CD36+ and demonstrated a cumulative increase in DQ collagen-1 degradation over time (Fig. 3g). CD36 knockdown enriched for a population of CD36 dPHFs with no capacity to degrade collagen-1, even after 8 h in the presence of DQ collagen-1 (Fig. 3g). This effect was recapitulated with a second guide RNA targeting CD36 (Supplementary Fig. 19). Live cell imaging revealed that DQ collagen-1 degradation occurred intracellularly in dPHFs, but was almost entirely abrogated by CD36 knockdown (Fig. 3h). To confirm the crucial role of CD36 in mediating collagen-1 degradation, FITC collagen-1, which decreases in florescence with degradation, was used. Live cell imaging revealed that FITC collagen-1 bound to the cell surface and over time decreased in signal intensity for wild-type dPHFs (Fig. 3i). CD36 knockdown did not affect FITC collagen-1 binding to the cell surface, but significantly reduced the rate of collagen-1 degradation (Fig. 3i).

Restoration of FAO and CD36 reduces ECM accumulation in skin

To determine whether enhancing PPAR signalling may upregulate CD36 expression in skin fibrosis, mice with radiation-induced skin fibrosis were treated with caffeic acid phenethyl ester (CAPE), a more bioactive derivative of caffeic acid14. CAPE was confirmed to downregulate fibronectin, collagen-1 and PAI1 in TGF-β1-treated dPHFs and, similar to caffeic acid, relied on an intact FAO pathway to exert its anti-fibrotic effects (Supplementary Fig. 20). CAPE upregulated PPAR signalling and the expression of CD36 in mice with skin fibrosis (Fig. 4a,b). Co-immunofluorescence demonstrated that mesenchymal (vimentin+) cells exhibited an increase in CD36 expression with CAPE treatment (Fig. 4c). This increase in PPAR signalling and CD36 expression corresponded to a reduction in both functional and histological evidence of fibrosis severity (Fig. 4d and Supplementary Fig. 21).

Fig. 4: Pharmacological and cellular therapy restores FAO in skin and reduces ECM accumulation.
Fig. 4

a,b, Metabolic reprogramming of skin fibrosis using CAPE. CAPE was delivered intraperitoneally, 10 mg kg–1 3 times per week for 10 weeks. Analysis was carried out 16 weeks post treatment. n = 8 CAPE-treated animals, n = 8 vehicle-treated animals. a, qRT–PCR normalised to vehicle control. b, qRT–PCR of CD36. c, Co-immunofluorescence of CD36 and vimentin+ cells. Left, representative confocal images. Right, quantification of co-localization. n = 7 biologically independent animals per group. d, Leg contracture for mice with skin fibrosis treated with CAPE or vehicle. Delivery and analysis were as described above for a,b. Two-way ANOVA, P < 0.001. ej, Metabolic reprogramming of skin fibrosis using CD36hi fibroblast transplantation. Analysis at 16 weeks post treatment. e, Localization of GFP-expressing CD36hi fibroblasts post transplantation. Tissue collected 7 d post transplantation. Left, GFP-expressing CD36hi fibroblasts confirmed using anti-GFP/2°Cy3. Right, overlay of anti-GFP/2°Cy3 with adjacent H&E section. f, Leg contracture at 16 weeks after CD36hi fibroblast transplantation. g, ECM reduction after CD36hi fibroblast transplantation. Top, quantification of trichome sections. Bottom, representative trichome sections. Additional images are available in Supplementary Fig. 23. h, PPAR signalling among the most significant pathways upregulated by CD36hi fibroblasts. Left, volcano plot of all detected genes from RNA-seq. Genes with FDR = 0 have been removed for clarity. Right, over-representation pathway analysis of genes, FDR < 0.05 with log10[fold change] > 1.25 or < −1.25. Fisher’s exact test with FDR correction. n = 3 CD36hi fibroblast-treated, n = 3 vehicle-treated. i, qRT–PCR of metabolic genes after CD36hi fibroblast transplantation, normalised to vehicle control. In f,g,i, n = 5 untreated, n= 8 CD36hi fibroblast-treated, n = 7 vehicle-treated. In f,g, two-tailed Student’s t-test was used. j, Untargeted metabolomics analysis post CD36hi fibroblast transplantation. Over-representation pathway analysis using MouseCyc. Red bars denote FAO pathways. No significant metabolic pathways were downregulated. Fisher’s exact test with FDR correction. n = 7 CD36hi fibroblast-treated, n = 6 vehicle-treated. k, Schema depicting the FAO and glycolysis regulating ECM homeostasis in fibroblasts. In a–j, all biological samples are independent. For a,b,i, we performed the pairwise fixed reallocation randomization test using REST. All data are expressed as means ± s.e.m., *P < 0.05, **P < 0.01. Scale bars, 40 μm (c); 200 μm (e,g).

Given the importance of CD36 as a direct regulator of collagen-1 degradation, we asked whether transplanting fibroblasts with high expression of CD36 could reduce ECM accumulation and improve metabolic dysregulation in skin fibrosis. One of the defining surface markers for mesenchymal cells derived from adipose tissue is their high expression of CD36. Spindle-shaped, plastic-adherent, vimentin+CD45CD31 cells (defined herein as fibroblasts) isolated from abdominal skin and adipose tissue confirmed that CD36 surface and gene expression was highest in adipose-derived fibroblasts (Supplementary Fig. 22a,b). This population was not composed of adipocytes, based on negative oil red staining. CD36hi fibroblasts were transplanted into radiation-induced skin fibrosis and localised to the reticular dermis, a region enriched with ECM producing fibroblasts (Fig. 4e). Sixteen weeks post transplantation, CD36hi fibroblasts reduced ECM deposition and improved tissue elasticity (Fig. 4f,g and Supplementary Fig. 23). To confirm this effect was specific to CD36, we transplanted adipose fibroblasts isolated from CD36 knockout mice, known to have a deficiency in fatty acid uptake15, and from CD36 wild-type mice into the fibrotic skin of CD36 wild-type mice (Supplementary Fig. 24). Only CD36 wild-type fibroblasts were capable of reducing skin fibrosis, and CD36 knockout fibroblast transplantation resulted in a trend towards even greater fibrosis than vehicle control. Genome-wide transcriptome profiling and pathway analysis revealed that the most significant effect of CD36hi fibroblast transplantation was an upregulation of the PPAR pathway and lipid metabolism (Fig. 4h), which was confirmed by qRT–PCR (Fig. 4i). To assess whether these genetic alterations correspond to metabolite changes, untargeted metabolomics analysis was performed to assess relative metabolite levels after CD36hi fibroblast transplantation. Skin fibrosis treated with CD36hi fibroblasts clustered separately from vehicle control, by principal components analysis, and upregulated fatty acid metabolites while downregulating glycolysis metabolites (Supplementary Fig. 25). Pathway analysis of significantly altered metabolites revealed multiple FAO degradation pathways altered by CD36high fibroblast transplantation (Fig. 4j). To elucidate whether the relative downregulation in glucose metabolites by untargeted metabolomics represented pathway activity or inactivity, targeted glycolysis metabolite profiling was performed. This analysis demonstrated that CD36hi fibroblast transplantation significantly downregulated multiple glycolysis pathway and pentose phosphate pathway intermediates, but had no effect on TCA cycle metabolites (Supplementary Fig. 26), which suggests that CD36hi fibroblast transplantation functionally downregulates the glycolysis pathway.


In conclusion, we have demonstrated that perturbation of PPAR signalling and FAO and glycolysis may be an important mechanism in normal and disease states to regulate ECM in skin (Fig. 4k). At a cellular level, metabolic regulation was demonstrated to be a major driver of ECM production and degradation by fibroblasts. We demonstrated that PPAR signalling promoted ECM degradation, through upregulation of lysosomal biogenesis and an increase in surface expression of CD36, a metabolically regulated fatty acid transporter that has been shown to be necessary for collagen-1 internalization and degradation. In support of this model, hypertrophic and keloid scarring is known to upregulate glycolysis enzymes16. Furthermore, caveolin 1, a scaffold protein necessary for CD36 surface expression, is reduced in fibroblasts derived from patients with systemic sclerosis17,18. Concordantly, caveolin 1 (Cav1)−/− mice display a scleroderma phenotype19.

Using this metabolic model of ECM regulation, a multitude of currently available compounds that inhibit glycolysis or upregulate FAO may have a role in the treatment of skin fibrosis, as exemplified by caffeic acid in this study. Interestingly, although caffeic acid was capable of both enhancing FAO and suppressing glycolysis in dermal fibroblasts in vitro, it was not capable of downregulating glycolysis in skin fibrosis in vivo, which suggests that multimodality treatment targeting glycolysis and FAO together may enhance the therapeutic effect of metabolic reprogramming for regulating skin ECM levels. Furthermore, by uncovering CD36 as a functional surface marker for ECM degradation, future cellular therapy may be tailored for conditions that require ECM production or degradation, based on their expression level of CD36. Finally, our metabolic model of ECM regulation may be applied to other organs, as a commonality among fibrosis of different tissues is the activation of ECM-producing mesenchymal cells. Indeed, fibroblasts isolated from patients with idiopathic pulmonary fibrosis display an increase in aerobic glycolysis while renal tubular epithelial cells undergoing epithelial mesenchymal transition in renal fibrosis exhibit a deficiency in FAO20,21,22. Therefore, metabolic regulation of dPHFs may be an important mechanism in controlling ECM homeostasis and may present exciting new therapeutic opportunities for the management of skin ECM dysregulation.


Human sample collection and processing

This study was approved by the research ethics board of the University Health Network (REB 15-9359-CE) and informed consent was obtained for each patient. Patients 18 years of age and over, and planning to undergo head and neck surgical procedures at the University Health Network between February 2015 to July 2016 were screened for previous radiotherapy. Patients were included if they had received full-dose radiotherapy to the neck (50–70 Gy) for cancer therapy, had completed treatment at least 150 d earlier, and had a medical history and physical examination record consistent with fibrosis. Patients treated with anti-fibrotic agents and those with fibrotic disorders were excluded. Consent was also obtained from normal age-matched control patients who were undergoing head and neck surgical procedures. Convenience sampling was used. Full-thickness dermis from the irradiated or normal neck was taken on the day of surgery and suspended in PBS solution. Samples were snap frozen in liquid nitrogen for storage.

Human RNA transcriptome processing

Human dermal tissue was pulverised with liquid nitrogen using a pestle and mortar. RNA was extracted using Qiagen RNeasy. RNA that passed the quality threshold (RNA integrity number (RIN) > 7), as assessed using the Agilent 2100 Bioanalyzer, were then run on the Human HT-12 V4 BeadChips (Illumina) according to the manufacturer’s instructions. Data were then checked for overall quality using the LUMI Bioconductor package in R (v2.15.3). Data were normalised using quantile normalization followed by a per-probe median-centered normalisation. Probes that were above the 20th percentile of the distribution of intensities in 80% of one of the two groups were included for analysis.

Animal experiments

All animal procedures were approved by the Animal Care Committee at the University Health Network under AUP 3422. A murine skin fibrosis model was created by radiating the hindlimb of C3H/HeJ mice (Jackson Laboratory) or B6.129S1-Cd36tm1Mfe/J mice (Jackson Laboratory) using an XRAD 225Cx (Precision X-Ray) and a 2.5-cm collimator. Prior to irradiation, fluoroscopic imaging was used to confirm and standardise the irradiation site. 40 Gy for C3H mice or 30 Gy for B6 mice was divided equally on the dorsal and ventral side of the leg. Animals were randomly assigned to groups for treatment by a researcher blinded to the treatments. For CD36hi cell treatment, 1 × 106 cells in 100 µl PBS were injected into the hindlimb subdermally at three sites weekly for 3 weeks starting at 4 weeks post irradiation. For caffeic acid phenethyl ester treatment, 10 mg kg−1 was given by intraperitoneal injection by a researcher blinded to treatment. Caffeic acid phenethyl ester or vehicle control was given three times per week for 10 weeks starting at 1 week post irradiation. Leg contracture measurement was performed, as described previously23, after first sedating mice with inhalation anesthesia and ensuring deep sedation by checking that there was no pain reflex and no muscle tone. Mice hips were then fixed on a custom apparatus, and 0.1 N equivalent weight was used to extend the fibrotic and contralateral normal leg, and leg extension was measured using a standardised ruler by two assessors blinded to treatment groups. Percentage leg contracture was calculated as ‘1 – leg extension of fibrotic leg/leg extension of contralateral normal leg’. Leg contracture was measured twice a week and averaged for each week.

Murine RNA-seq processing

For temporal transcriptome profiling, normal and/or fibrotic skin tissue was dissected at 0, 42, and 140 d post radiation from age-matched C3H/HeJ mice. CD36hi cells were collected at 16 weeks post treatment along with vehicle controls. CD36hi vehicle control was also used as the last time point for temporal transcriptome profiling. Samples were snap frozen in liquid nitrogen. For RNA extraction, samples were pulverised with liquid nitrogen using a pestle and mortar. RNeasy (Qiagen) was used for RNA extraction as per the manufacturer’s instructions. RNA samples were quantified using the qubit RNA kit (Life Technologies) on the Qubit 2.0 fluorometer. The RNA integrity was assessed using the Agilent RNA 6000 nano chip on the Agilent Bioanalyzer 2100. Each sample had a RIN score greater than 8.0 and 250 ng of each sample was used for library preparation. Libraries were prepared using the TruSeq Stranded mRNA Library Prep Kit according to the manufacturer’s instructions (Illumina). Final complementary DNA (cDNA) library size distribution was assessed using an Agilent Bioanalyzer 2100 and quantified by qPCR (Kapa Biosystems). All libraries were normalised to 10 nM, pooled (with 5 samples per pool), and diluted to 2 nM. The 2 nM pools were denatured using 0.2 N NaOH and diluted again to a final concentration of 10 pM. The 10 pM pools were loaded onto a cBot (Illumina) for cluster generation. The clusters on the flow cells were sequenced (paired-end sequencing for 100 cycles using v3 reagents on the Illumina HighSeq 2000 at the Princess Margaret Genomics Facility) to achieve a minimum of 30 million reads per sample. Raw RNA-seq data quality control was performed using FASTQC v.0.11.2 (Babraham Bioinformatics). The quality metrics for each sample included per-base sequence quality, per-tile sequence quality, per-base sequence content, per-sequence GC content, sequence length distribution, duplication levels, over-represented sequences and adapter content. All samples were considered acceptable based on FASTQC analysis. TopHat and Cufflinks were used according to a protocol outlined previously24. Raw sequencing data were aligned to the human genome using Bowtie (version 2.2.3) and TopHat (version 2.0.13). Transcript assembly was performed using Cufflinks (version 2.2.1). Post-alignment quality checks were performed using RNA-SeQC (version 1.1.7). Read-count metrics, such as total, unique, duplicate, mapped, mapped unique, and transcript annotated reads, and expressed transcripts and coverage metrics, including mean coverage, mean coefficient of variation, coverage gaps, 5′ to 3′ coverage, GC bias, coverage plots, and cumulative gap length were assessed for each sample. All samples passed post-alignment quality checks.

Untargeted metabolomics processing

Fibrotic and normal murine skin samples were pulverised in liquid nitrogen using a custom pestle and mortar. For metabolite extraction, samples were resuspended in ice-cold acidic (pH < 2) PBS to make a 100 mg ml−1 homogenate. 50 µl was used for metabolite extraction. 13C6-phenylalanine (2 µl at 250 ng µl−1) was added as an internal standard to all samples 1 min prior to sonication in an ice bath (four pulses of 15 s). Protein was then precipitated with 250 µl cold acetonitrile and methanol solution (v:v ratio of 1:1). The supernatant was then divided into aliquots and dried for LC–MS (liquid chromatography–mass spectrometry) analysis. For this, a mass spectrometer (Agilent Technologies 6550 Q-TOF) was coupled with an ultra high pressure liquid chromatograph (1290 Infinity UHPLC Agilent Technologies). Profiling data were acquired under both positive and negative electrospray ionization conditions over a mass range m/z of 100–1,700 at a resolution of 10,000 (separate runs) in scan mode. Metabolite separation was achieved using two columns of differing polarity, a hydrophilic interaction column (HILIC, ethylene-bridged hybrid 2.1 × 150 mm, 1.7 mm; Waters) and a reversed-phase C18 column (high-strength silica 2.1 × 150 mm, 1.8 mm; Waters) with a gradient described previously25. Samples were injected in duplicate, wherever necessary, and a pooled quality control sample, made up of all of the samples from each study, was injected several times during a run. A separate plasma quality control sample was analysed with a pooled quality control sample to account for analytical and instrumental variability. Automated tandem mass spectrometry (MS/MS) data were also acquired with a pooled quality control sample to aid the identification of unknown compounds using fragmentation patterns. Data alignment and filtering was performed using Mass Profiler Professional software (Agilent B.12.05). Default settings were used with the exception of the signal-to-noise ratio threshold (3), mass limit (0.0025 units), and time limit (9 s). Putative identification of each metabolite was carried out based on accurate mass (m/z) against the METLIN database using a detection window of ≤7 ppm. Metabolites detected in at least 80% of one of two groups were selected for differential expression analyses.

Cell culture

dPHFs obtained from ATCC (PCS-201-012) were cultured in DMEM with 5% FBS (fetal bovine serum). Mesenchymal cells were isolated primarily from C3H or B6 mice. Animals were euthanised and intraperitoneal fat or skin was dissected and minced. Collagenase and hyaluronidase (Stemcell Technologies) digestion was performed in a rotating oven at 37 °C for 1 h, passed through a 70 µM filter, and then plated in DMEM + 10% FBS to expand the mesenchymal population. Cells were starved for 12 h in MEM medium prior to treatment with TGF-β1 (Sigma), caffeic acid (Sigma), dichloroacetic acid (Sigma), PDGF-BB (R&D Systems), and/or etomoxir (Tocris). All protein and mRNA profiling was performed 24 h post treatment except for PDGF-BB, which was performed 48 h post treatment.

Trichrome quantification

Formalin-fixed, paraffin-embedded 8µm sections treated with Masson’s Trichrome Stain were imaged with a whole slide scanner (Aperio ScanScope XT). Analysis was performed by a researcher blinded to treatment. Using ImageJ, whole slide scanned images were loaded using the Bioanalyzer plugin, and then stacked to RGB. Colour deconvolution was used to separate the full-colour images into their red, green and blue components. A threshold pixel intensity was manually adjusted to provide contrast between background and pixels of interest, and this setting was set and standardised for all images. A ratio of blue (trichrome) to dermis area was derived for each image, representing the extracellular matrix percentage area in skin.

Western blot

For supernatant proteins, supernatant was spun at 6,000g for 10 min at 4 °C. For cell lysate, protein was extracted in RIPA (radioimmunoprecipitation assay) buffer with a protease inhibitor (Roche) and spun at 6,000g for 10 min at 4 °C. Antibodies used were collagen-1 (ab138492, abcam, 1:5000), fibronectin (ab2413, abcam, 1:5000), PAI1 (sc5297, Santa Cruz, 1:1000), beta-actin (ab8229, abcam, 1:1,000). Uncropped scans of representative blots and molecular weight markers for all immunoblots in the main text are available in Supplementary Figs. 2734.


Human and murine skin samples were macrodissected, snap frozen in liquid nitrogen, and stored at −80 °C. For in vitro experiments, cells were trypsinised and then snap frozen in liquid nitrogen or immediately underwent RNA extraction. All RNA was extracted using the RNeasy Mini Kit (Qiagen) according to the manufacturer’s instructions. RNA quality was measured and quantified using the NanoDrop 2000c Spectrophotometer (Thermo Scientific); we used either A260 or A280 for cell samples, or RIN for human and murine samples using a 2100 bioanalyzer (Aligent). cDNA was synthesised from normalised amounts of RNA using random hexamer primers and SuperScript III (Invitrogen) on a PTC-200 Peltier Thermal Cycler. Primers were obtained from peer-reviewed publications whenever possible. See Supplementary Dataset 1 for references and amplicon size. Primers were validated using BLAST to ensure specificity, gel electrophoresis to confirm size, and melting curve assessment for all reactions. PCR efficiency was measured using a standard curve through the dilution of a representative sample to determine the linear range of cDNA amounts for qRT-PCR reactions. qRT-PCR reactions were performed on 384-well plates (Applied Biosystems) using SYBR Green PCR Master Mix (Applied Biosystems). The analysis was run on the 7900HT Fast Real-Time PCR System (Applied Biosystems). Reaction conditions for cDNA synthesis and qRT-PCR are found in Supplementary Tables 1, 2 respectively. A no-template control was used to monitor contamination and false-positive results from primer dimers. Results of no-template controls were confirmed to have higher cycle threshold (CT) values and distinct melting curves compared to template reactions. 18S ribosomal RNA was used as an internal control. A 2−ΔΔCT method was used for relative quantification and for graphical representation in figures. Statistical analysis was performed on biological replicates using the REST26 set to default parameters.

Flow cytometry

For DQ collagen-1 and CD36 experiments, dPHFs were treated as above (Cell culture) for 24 h. DQ collagen-1 (Sigma) at 4 µg ml−1 was added to and incubated with cells for an additional 5 h. Cells were then prepared for flow cytometry using 5 µl of CD36-PE (336206, Biolegend (for human), 102605, Biolegend (for mouse)) added to 200,000 cells in 100 µl FACS buffer (PBS + 1% FBS) and incubated for 30 min on ice. DAPI (1:2,000) was added just prior to flow cytometry for identification of live cells. For cell surface profiling, cells were incubated in separate FACS tubes with CD31-PECy7 (102418, BioLegend) and CD45-PECy7 (102418, BioLegend). Reference gating is available in Supplementary Fig. 35.

CRISPR–Cas9 knockdown

Guide RNAs were created using a guide design web tool (http://crispr.mit.edu/). CD36 guide 1, CGGAACTGTGGGCTCATCGC; CD36 guide 2, TGGGCTGTGACCGGAACTGT; PPARG guide, CTCCGTGGATCTCTCCGTAA; ACOX1 guide, TGACAGCCACCTCATAACGC; HK2 guide, TGACCACATTGCCGAATGCC. pSpCas9(BB)-2A-GFP (Addgene, 48138) was used. The protocol was performed according to methods outlined previously27. dPHFs (5 × 105) were electroporated with 5 µg of construct using an Amaxa mouse neural stem cell nucleofactor kit (Lonza). Cells were grown for 24 h and then sorted for GFP.

SMAD3 luciferase reporter

SMAD3 luciferase reporter was constructed according to previous methods28. Oligonucleotides containing SMAD3 binding sites were designed with BamH1 and Sal1 digestion sites. Forward: 5′-GATCCGAGCCAGACAAAAAGCCAGACATTTAGCCAGACACG-3′, reverse: 5′-TCGACGTGTCTGGCTAAATGTCTGGCTTTTTGTCTGGCTCG-3′. SMAD oligonucleotides were inserted into a p-delta-51 LUC II reporter. dPHFs were electroporated with construct using an Amaxa mouse neural stem cell nucleofactor kit (Lonza). 24 h after electroporation, cells were treated as above (Cell culture) for 24 h and lysed to assess for luciferase and renilla activity. Relative luciferase activity was defined as the level of luciferase bioluminescence divided by renilla bioluminescence for each sample.

GFP CD36high cell localization

pHAGE PGK-GFP-IRES-LUC-W-packaged lentivirus was created by transient transfection of 293 T cells and infected into CD36high fibroblasts. CD36high fibroblasts were sorted for GFP positivity and injected into murine skin fibrosis. For localization of GFP-CD36high fibroblasts, skin fibrosis tissue was dissected 7 days post transplantation and 8 µm formalin-fixed paraffin-embedded (FFPE) sections were created. Adjacent FFPE sections were stained with hematoxylin and eosin (H&E) or anti-GFP (ab6556, Abcam, 1:1000), secondary Cy3 (ab6939, Abcam, 1:2000), and DAPI. H&E sections were imaged using a Leica DC300 and immunofluorescence was imaged with an Olympus Fluoview laser-scanning confocal microscope.

Lysosome quantification

dPHFs were treated as above (Cell culture) and exposed to LysoTracker (ThermoFisher) as per the manufacturer’s instructions on a four-chamber cover glass system (ThermoFisher). Images were obtained using an Olympus Fluoview laser-scanning confocal microscope. Fluorescence intensity was quantitated per cell using the Object Identification module and Measurement function of Volocity software. Cell area was measured using ImageJ software.

Live cell imaging and quantification

dPHFs were treated as above (Cell culture) on an eight-chamber cover glass system (ThermoFisher) for 24 h prior to addition of DQ collagen-1 (4 µg ml−1) (ThermoFisher) or FITC collagen-1 (7 µg ml−1) (Chondrex). Cells were assessed immediately by live cell imaging using a Quorum WaveFX spinning disk confocal system. Per well, 3–5 locations were selected at random for imaging every 6 min at 20× magnification. Fluorescence intensity was quantified using the Object identification and Measurement modules in Volocity. Overall fluorescence intensity was measured as the sum of all identified objects within the selected threshold. Cell area was measured using ImageJ software.

FAO and glycolysis functional assessment for dPHFs

Seahorse experiments were performed on an XFe96 analyzer (Aligent) according to the manufacturer’s instructions. dPHFs were prepared as above (Cell culture) on an XF96 culture plate. For glycolysis profiling, a final concentration of 1 mM or 10 mM glucose, 1 µM oligomycin, and 100 mM 2-DG was used. For FAO, palmitate sodium (Sigma) and BSA-fatty acid free (Sigma) were used to make 0.17 mM palmitate-BSA, as per the manufacturer’s instructions. A 40 µM final concentration of Etomoxir (Tocris) was used. After seahorse profiling, wells were normalised using the CyQUANT Direct Cell Proliferation Assay Kit (ThermoFisher) as per the manufacturer’s instructions.

PET-CT assessment of FDG uptake

18F-fluorodeoxyglucose (FDG) was purchased from the Centre for Probe Development and Commercialization (CPDC). Mice were fasted overnight. Each mouse was injected with 9.7 ± 1.1MBq and maintained under anesthesia (2% inhaled isoflurane and air mixture) during the tracer-uptake period. Mice were then placed into a triple mouse bed (Minerve) for imaging. Images were acquired 1 h post injection of tracer in a single 10-min frame using a Focus 220 preclinical µPET scanner (Siemens), followed by an 8-min 57Co transmission scan using a rotating source for attenuation and scatter correction. Image reconstruction was performed using a three-dimensional (3D) ordered subset expectation maximization (OSEM3D) (2 iterations) and maximum a posteriori (MAP) (18 iterations), and produced voxel dimensions of 0.146 × 0.146 × 0.796 mm. The bed containing the mice was transferred immediately after the PET acquisition was completed to a Locus Ultra μCT (GE Healthcare) for computed tomography (CT) image acquisition operating at 80 kVp and 50 mA. Images were reconstructed with an isotropic voxel size of 0.154 mm. PET and CT images were co-registered and analysed using the Inveon Research Workplace software v4.2 (Siemens Healthcare).

Glucose consumption assay

Glucose consumption experiments were performed according to methods described previously29,30. Cells were seeded in culture plates for 12 h. The culture medium was then changed and cells were incubated for 24 h. Glucose in medium was measured as per the manufacturer’s instructions using a Glucose Colorimetric Assay Kit (Cayman Chemical).


CD36 ELISA (Sigma-Aldrich) was performed according to the manufacturer’s instructions after total protein isolation. Quantification was performed 24 h after treatment for dPHFs treated using TGF-β1 (3 ng ml−1), caffeic acid (20 µM or 40 µM), or etomoxir (5 µM or 10 µM). CRISPR–Cas9 knockdown cells were compared to Cas9 control cells at the same passage number.

Targeted metabolomics of glycolysis, TCA, and pentose phosphate pathway metabolites

The following targeted metabolite processing was conducted at the analytical facility for bioactive molecules at SickKids Hospital. Metabolism standards and samples were spiked with internal standard mixture. Homogenised tissue samples and standards had 500 µl of 80% aqueous acetonitrile (cooled to −20 °C) added. Tubes were vortexed for 1 min at 4 °C. Extraction proceeded at −80 °C for 4 h. After incubation, tubes were centrifuged at 20,000g for 10 min at 4 °C. The supernatant was transferred to a 2 ml Eppendorf tube and kept at −80 °C. The pellet was re-extracted with 400 µl of 80% aqueous acetonitrile (cooled to −20 °C), vortexed for 1 min at 4 °C, and incubated at −80 °C for 30 min. After 30 min, tubes were centrifuged at 20,000g for 10 min at 4 °C. The supernatant was combined with the previous extract and again centrifuged at 20,000g for 10 min at 4 °C. The resulting supernatant was transferred to a 2-ml Eppendorf tube with a hole punched in the cap and put at −80 °C until frozen. Once frozen, the tubes were lyophilised overnight (approximately 18 h). Lyophilised samples were kept at −80 °C. Samples were analysed with liquid chromatography (LC)–MS/MS using an Agilent 1290 or 1200 HPLC with a Q-Trap 5500 mass spectrometer (AB Sciex). We used Analyst 1.6.3 (AB Sciex) to integrate the peaks, generate area ratios, and quantify the data. Samples were normalised to total protein using the Bradford assay.

Statistical analysis and reproducibility

Differential expression for murine RNA-seq

Cufflinks (version 2.2.1) was used for murine RNA-seq statistical testing, following the standard pipeline described previously24. Files were then loaded into CummeRBund v2.7.2 for final output of significance testing, results, and graphing.

Differential expression for human transcriptome and metabolomic datasets

A moderated Student t-test with Benjamini–Hochberg false discovery rate (FDR) correction for multiple testing was performed on log2-transformed gene expression or metabolite levels for differential expression.

Pathway enrichment analysis

Two methods were used for enrichment analyses with KEGG and Reactome. Gene set enrichment analysis (GSEA) was performed using the Broad Institute GSEA software v3.0. Over-representation analysis was carried out using the Piano R package (version 1.8.2) to perform hypergeometric testing on genes that had log10[fold change]  > 1.25 or < –1.25 and FDR < 0.05.

Computational pharmacogenomics

The Connectivity Map (CMAP) dataset was normalised using Robust Multi-array Average (RMA), as performed in our PharmacoGx package (version 1.1.6)9,31,32. Each gene was fitted using a linear model adjusted for batches and cell line, and the genes were ranked with respect to their associated Student t-test statistics in order to build a ‘perturbation signature’ for each compound tested in CMAP. Genes differentially expressed between human fibrotic and normal skin were ranked based on their associated t statistics, and referred to as the ‘query signature’. A second query signature was created using the curated genes in KEGG for PPAR signalling and glycolysis. These two query signatures were used to query CMAP and thus two ranked lists of drugs were obtained. Drugs were ranked according to their connectivity score, in descending order. The overlap between the high ranked drugs (positive connectivity score) of the two lists were obtained and ranked based on their LD50.

Human Hierarchical clustering and heatmap

For each gene, a probe was selected with the maximum interquartile range. The mapped genes were ranked according to the variance of their expression across the entire dataset. The top 1,000 most variant genes were selected and used for hierarchical clustering (correlation distance and complete linkage) and geneartion of a heatmap (gplots_3.0.1). Finally, to measure significance of the clustering results, a Fisher’s exact test was performed on the two largest clusters and compared with the true labels of the samples.

Temporal clustering

MaSigPro R package (version 1.40.0), updated for clustering temporal RNA-seq data, was used. Genes that were significantly expressed (FDR < 0.05) between time points 0 and 42 days and between 42 and 140 d were aggregated. MaSigPro (v. 1.40.0) pipeline was utilised for hierarchical cluster analysis and visualization of clusters. The optimal number of clusters was defined by the silhouette method described previously33. The most parsimonious solution for k = 3 was selected.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Data availability

Transcriptomic data can be found on the National Center for Biotechnology Information (NCBI) gene expression omnibus (GEO) using ID GSE98157 for murine RNA-seq data and GSE98159 for human HT-12 array data. Software parameters have been uploaded onto Github: https://github.com/bhklab/SkinFibrosis. The data that support the plots within this paper and other findings of this study are available from the corresponding authors upon reasonable request.

Additional information

Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.


  1. 1.

    Watt, F. M. Mammalian skin cell biology: at the interface between laboratory and clinic. Science 346, 937–940 (2014).

  2. 2.

    Vander Heiden, M. G., Cantley, L. C. & Thompson, C. B. Understanding the Warburg Effect: the metabolic requirements of cell proliferation. Science 324, 1029–1033 (2009).

  3. 3.

    Kelly, B. & O’Neill, L. A. Metabolic reprogramming in macrophages and dendritic cells in innate immunity. Cell Res. 25, 771–784 (2015).

  4. 4.

    Stone, H. B., Coleman, C. N., Anscher, M. S. & McBride, W. H. Effects of radiation on normal tissue: consequences and mechanisms. Lancet Oncol. 4, 529–536 (2003).

  5. 5.

    Mehta, A. S. & Blodgett, T. M. Retroperitoneal fibrosis as a cause of positive FDG PET/CT. J. Radiol. Case Rep. 5, 35–41 (2011).

  6. 6.

    Justet, A. et al. [18F]FDG PET/CT predicts progression-free survival in patients with idiopathic pulmonary fibrosis. Respir. Res. 18, 74 (2017).

  7. 7.

    Kok, H. M., Falke, L. L., Goldschmeding, R. & Nguyen, T. Q. Targeting CTGF, EGF and PDGF pathways to prevent progression of kidney disease. Nat. Rev. Nephrol. 10, 700–711 (2014).

  8. 8.

    Higgins, D. F. et al. Hypoxic induction of Ctgf is directly mediated by Hif-1. Am. J. Physiol. Ren. Physiol. 287, F1223–F1232 (2004).

  9. 9.

    Lamb, J. et al. The Connectivity Map: using gene-expression signatures to connect small molecules, genes, and disease. Science 313, 1929–1935 (2006).

  10. 10.

    Vanella, L. et al. Caffeic acid phenethyl ester regulates PPAR’s levels in stem cells-derived Adipocytes. PPAR Res. 2016, 7359521 (2016).

  11. 11.

    Isono, M., Chen, S., Hong, S. W., Iglesias-de la Cruz, M. C. & Ziyadeh, F. N. Smad pathway is activated in the diabetic mouse kidney and Smad3 mediates TGF-beta-induced fibronectin in mesangial cells. Biochem. Biophys. Res. Commun. 296, 1356–1365 (2002).

  12. 12.

    McCulloch, C. A. G. & Knowles, G. C. Deficiencies in collagen phagocytosis by human fibroblasts in vitro: a mechanism for fibrosis? J. Cell. Physiol. 155, 461–471 (1993).

  13. 13.

    Febbraio, M., Hajjar, D. P. & Silverstein, R. L. CD36: a class B scavenger receptor involved in angiogenesis, atherosclerosis, inflammation, and lipid metabolism. J. Clin. Invest. 108, 785–791 (2001).

  14. 14.

    Zhang, P., Tang, Y., Li, N.-G., Zhu, Y. & Duan, J.-A. Bioactivity and chemical synthesis of caffeic acid phenethyl ester and its derivatives. Molecules 19, 16458–16476 (2014).

  15. 15.

    Coburn, C. T. et al. Defective uptake and utilization of long chain fatty acids in muscle and adipose tissues of CD36 knockout mice. J. Biol. Chem. 275, 32523–32529 (2000).

  16. 16.

    Vincent, A. S. et al. Human skin keloid fibroblasts display bioenergetics of cancer cells. J. Invest. Dermatol. 128, 702–709 (2008).

  17. 17.

    Ring, A., Le Lay, S., Pohl, J., Verkade, P. & Stremmel, W. Caveolin-1 is required for fatty acid translocase (FAT/CD36) localization and function at the plasma membrane of mouse embryonic fibroblasts. Biochim. Biophys. Acta 1761, 416–423 (2006).

  18. 18.

    Del Galdo, F. et al. Decreased expression of caveolin 1 in patients with systemic sclerosis: crucial role in the pathogenesis of tissue fibrosis. Arthritis Rheum. 58, 2854–2865 (2008).

  19. 19.

    Castello-Cros, R. et al. Scleroderma-like properties of skin from caveolin-1-deficient mice: implications for new treatment strategies in patients with fibrosis and systemic sclerosis. Cell Cycle 10, 2140–2150 (2011).

  20. 20.

    Xie, N. et al. Glycolytic reprogramming in myofibroblast differentiation and lung fibrosis. Am. J. Respir. Crit. Care. Med. 192, 1462–1474 (2015).

  21. 21.

    Kang, H. M. et al. Defective fatty acid oxidation in renal tubular epithelial cells has a key role in kidney fibrosis development. Nat. Med. 21, 37–46 (2015).

  22. 22.

    Lovisa, S., Zeisberg, M. & Kalluri, R. Partial epithelial-to-mesenchymal transition and other new mechanisms of kidney fibrosis. Trends Endocrinol. Metab. 27, 681–695 (2016).

  23. 23.

    Stone, H. B. Leg contracture in mice: an assay of normal tissue response. Int. J. Radiat. Oncol. Biol. Phys. 10, 1053–1061 (1984).

  24. 24.

    Trapnell, C. et al. Differential gene and transcript expression analysis of RNA-seq experiments with TopHat and Cufflinks. Nat. Protoc. 7, 562–578 (2012).

  25. 25.

    Dutta, T. et al. Concordance of changes in metabolic pathways based on plasma metabolomics and skeletal muscle transcriptomics in type 1 diabetes. Diabetes 61, 1004–1016 (2012).

  26. 26.

    Pfaffl, M. W. Relative expression software tool (REST(C)) for group-wise comparison and statistical analysis of relative expression results in real-time PCR. Nucleic Acids Res. 30, 36e–36e (2002).

  27. 27.

    Ran, F. A. et al. Genome engineering using the CRISPR-Cas9 system. Nat. Protoc. 8, 2281–2308 (2013).

  28. 28.

    Dennler, S. et al. Direct binding of Smad3 and Smad4 to critical TGF beta-inducible elements in the promoter of human plasminogen activator inhibitor-type 1 gene. EMBO J. 17, 3091–3100 (1998).

  29. 29.

    Khan, M. et al. mTORC2 controls cancer cell survival by modulating gluconeogenesis. Cell Death Discov. 1, 15016 (2015).

  30. 30.

    Jiang, P. et al. P53 regulates biosynthesis through direct inactivation of glucose-6-phosphate dehydrogenase. Nat. Cell Biol. 13, 310–316 (2011).

  31. 31.

    Irizarry, R. A. Exploration, normalization, and summaries of high density oligonucleotide array probe level data. Biostatistics 4, 249–264 (2003).

  32. 32.

    Smirnov, P. et al. PharmacoGx: an R package for analysis of large pharmacogenomic datasets. Bioinformatics 32, 1244–1246 (2016).

  33. 33.

    Rousseeuw, P. J. Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. J. Comput. Appl. Math. 20, 53–65 (1987).

Download references


This research was funded by the Canadian Institutes of Health Research, the Canadian Cancer Society Research Institute, the Physicians Services Incorporated Foundation, the Harry Barbarian Research Grant, the Mariano Elia Chair in Head and Neck Cancer Research, the Campbell Family Institute for Cancer Research, the Ministry of Health and Long-Term Care, and the Princess Margaret Cancer Centre Head and Neck Translational Program, with philanthropic funding from the Wharton family and Joe’s Team.

Author information


  1. Department of Otolaryngology—Head and Neck Surgery, University of Toronto, Toronto, Ontario, Canada

    • Xiao Zhao
    • , David Goldstein
    • , Ralph Gilbert
    •  & Ian Witterick
  2. Princess Margaret Cancer Centre, University Health Network, Toronto, Ontario, Canada

    • Xiao Zhao
    • , Laleh Soltan Ghoraie
    • , Kenneth Yip
    • , Justin Williams
    • , Scott V. Bratman
    • , Laurie Ailles
    • , Benjamin Haibe-Kains
    •  & Fei-Fei Liu
  3. Insitute of Medical Sciences, University of Toronto, Toronto, Ontario, Canada

    • Xiao Zhao
    •  & Benjamin Haibe-Kains
  4. Department of Medical Biophysics, University of Toronto, Toronto, Ontario, Canada

    • Pamela Psarianos
    • , Hilary Pang
    • , Ali Hussain
    • , Ju Hee Lee
    • , Scott V. Bratman
    • , Laurie Ailles
    •  & Fei-Fei Liu
  5. Department of Otolaryngology—Head and Neck Surgery and Surgical Oncology, University Health Network, University of Toronto, Toronto, Ontario, Canada

    • David Goldstein
    • , Ralph Gilbert
    •  & Ian Witterick
  6. Department of Otolaryngology—Head and Neck Surgery, Mount Sinai Hospital, Toronto, Ontario, Canada

    • Ian Witterick
  7. Radiation Medicine Program, Princess Margaret Cancer Centre, Toronto, Ontario, Canada

    • Scott V. Bratman
    •  & Fei-Fei Liu
  8. Department of Radiation Oncology, University of Toronto, Toronto, Ontario, Canada

    • Scott V. Bratman
    •  & Fei-Fei Liu
  9. Department of Computer Science, Princess Margaret Cancer Centre, Toronto, Ontario, Canada

    • Benjamin Haibe-Kains
  10. Ontario Institute for Cancer Research, Toronto, Ontario, Canada

    • Benjamin Haibe-Kains


  1. Search for Xiao Zhao in:

  2. Search for Pamela Psarianos in:

  3. Search for Laleh Soltan Ghoraie in:

  4. Search for Kenneth Yip in:

  5. Search for David Goldstein in:

  6. Search for Ralph Gilbert in:

  7. Search for Ian Witterick in:

  8. Search for Hilary Pang in:

  9. Search for Ali Hussain in:

  10. Search for Ju Hee Lee in:

  11. Search for Justin Williams in:

  12. Search for Scott V. Bratman in:

  13. Search for Laurie Ailles in:

  14. Search for Benjamin Haibe-Kains in:

  15. Search for Fei-Fei Liu in:


X.Z., L.S.G., K.Y., D.G., R.G., I.W., S.V.B., L.A., B.H.-K., and F.-F.L. participated in the research design. X.Z, P.P., .H.P., A.H., J.H.L., and J.W. conducted experiments. X.Z., L.S.G., L.A., B.H.-K. contributed new reagents or analytic tools. G.Z., L.S.G., P.P., H.P., A.H., J.H.L. and J.W. performed data analysis. X.Z., L.S.G., P.P., K.Y., D.G., R.G., S.V.B., L.A., B.H.-K., and F.-F.L. wrote or contributed to the writing of the manuscript.

Competing interests

The authors declare no competing interests.

Corresponding authors

Correspondence to Xiao Zhao or Fei-Fei Liu.

Supplementary information

  1. Supplementary Information

    Supplementary Figures 1–35 and Supplementary Tables 1 and 2

  2. Reporting Summary

  3. Supplementary Dataset 1

    qRT–PCR primers

About this article

Publication history






Further reading