Metabolism‐related biomarkers, molecular classification, and immune infiltration in diabetic ulcers with validation

Abstract

Diabetes mellitus (DM) can lead to diabetic ulcers (DUs), which are the most severe complications. Due to the need for more accurate patient classifications and diagnostic models, treatment and management strategies for DU patients still need improvement. The difficulty of diabetic wound healing is caused closely related to biological metabolism and immune chemotaxis reaction dysfunction. Therefore, the purpose of our study is to identify metabolic biomarkers in patients with DU and construct a molecular subtype‐specific prognostic model that is highly accurate and robust. RNA‐sequencing data for DU samples were obtained from the Gene Expression Omnibus (GEO) database. DU patients and normal individuals were compared regarding the expression of metabolism‐related genes (MRGs). Then, a novel diagnostic model based on MRGs was constructed with the random forest algorithm, and classification performance was evaluated utilizing receiver operating characteristic (ROC) analysis. The biological functions of MRGs‐based subtypes were investigated using consensus clustering analysis. A principal component analysis (PCA) was conducted to determine whether MRGs could distinguish between subtypes. We also examined the correlation between MRGs and immune infiltration. Lastly, qRT‐PCR was utilized to validate the expression of the hub MRGs with clinical validations and animal experimentations. Firstly, 8 metabolism‐related hub genes were obtained by random forest algorithm, which could distinguish the DUs from normal samples validated by the ROC curves. Secondly, DU samples could be consensus clustered into three molecular classifications by MRGs, verified by PCA analysis. Thirdly, associations between MRGs and immune infiltration were confirmed, with LYN and Type 1 helper cell significantly positively correlated; RHOH and TGF‐β family remarkably negatively correlated. Finally, clinical validations and animal experiments of DU skin tissue samples showed that the expressions of metabolic hub genes in the DU groups were considerably upregulated, including GLDC, GALNT6, RHOH, XDH, MMP12, KLK6, LYN, and CFB. The current study proposed an auxiliary MRGs‐based DUs model while proposing MRGs‐based molecular clustering and confirmed the association with immune infiltration, facilitating the diagnosis and management of DU patients and designing individualized treatment plans.

Abbreviations

BP

biological pathways

CC

cellular components

CFB

complement factor B

DU

diabetic ulcer

DM

diabetes mellitus

DE

differentially expressed

DFUs

diabetic foot ulcers

EMT

epithelial‐mesenchymal transformation

FDR

false discovery rate

GALNT6

polypeptide N‐acetylgalactosaminyltransferase 6

GEO

gene expression omnibus

GLDC

glycine decarboxylase

GO

gene ontology

GTP

guanosine triphosphate

KEGG

Kyoto encyclopedia of genes and genomes

KLK6

Kallikrein‐related peptidase 6

LC–MS/MS

liquid chromatography–tandem mass spectrometry

LYN

LYN proto‐oncogene, Src family tyrosine kinase

MF

molecular function

MMP

member of the metalloproteinase

MMP12

matrix metallopeptidase 12

MRGs

metabolomic‐regulated genes

NC

negative control

PCA

principal component analysis

qRT‐PCR

quantitative real‐time polymerase chain reaction

RHOH

Ras homologue family member H

ROC

receiver operating characteristic

SD

standard deviation

SPF

specific pathogen free

STZ

streptozotocin

TGF‐β

the transforming growth factor beta

TH1

type 1 helper

XDH

Xanthine dehydrogenase

1. INTRODUCTION

Diabetes mellitus (DM), a severe global illness  is associated with substantial morbidity and mortality. ¹ Diabetic ulcers (DUs) are one of the most grievous complications of DM, amputation is generally necessary within 6 to 18 months for 10% of diabetic lesions, but 15% remain active, and 5% to 24% require recurrence within a year, ² leading to a serious impact on the quality of life of the patients.

Patients with diabetes suffer from wounds that are laborious to heal due to high glucose levels, various biological factors, and changes in microenvironments and immune chemotaxis, ³ which is a complex biological process closely related to biological metabolism. Metabolomics is one of the important components of systems biology, undergoing rapid technological development (Christodoulou et al., 2020); ⁴ , ⁵ these advances have facilitated the application of metabolomics in determining predictive biomarkers of metabolically dysfunctional disease occurrence and pathophysiology. ⁶ , ⁷ Gradually, metabolomics marker has been used to study diabetic foot ulcers (DFUs). The healing results of DFUs have revealed appreciable differences in serum amino acid metabolism levels in the healing group compared with the non‐healing group. ⁸ Therefore, understanding DUs through a metabolic perspective may lead to breakthroughs. However, there is still a lack of systematic research on the significance of metabolic genes in DUs.

Recent studies have shown that metabolism is a regulatory pathway for immunity (Figure 1). Activated immune cells are present in many metabolically relevant pathways in cancer cells. ⁹ For instance, targeting specific aspects of intratumor metabolisms, like the hexosamine biosynthesis pathway or glutamine metabolism, can boost the immune response and make tumors sensitive to checkpoint blockade. ¹⁰ Likewise, in DU wounds, the immune infiltrating microenvironment sustains a persistent inflammatory response. ¹¹ It has been proven that the transition of M1 to M2 macrophages appears impaired in diabetic wounds. ¹² Immunological research at DUs has led to many immunotherapies. ¹³ Thus, exploring the impact and role of metabolic trends on immune infiltration in the DU microenvironment may lead to novel heterogeneous wound healing strategies.

FIGURE 1.

Cancer cell metabolism and derangements in the TME. Reviews of metabolism of immune cells in cancer. MDSC, myeloid‐derived suppressor cell; R‐2‐HG, (R)‐2‐hydroxyglutarate; ROS, reactive oxygen species; T_eff, effector T; T_reg, regulatory T.

It is, therefore, urgent to develop a predictive screening model that can identify high‐predictive‐accuracy metabolism‐related genes (MRGs) in DUs. The random forest algorithm comprised of many decision trees is iteratively constructed using random predictor and dependent variable sets. ¹⁴ , ¹⁵ It is based on the principle of ensemble learning that previously has shown high predictive accuracy (62%‐71%) in modeling and provides more important variable estimates than classifiers. ¹⁶

Here, the random forest algorithm constructed a novel DUs diagnostic model associated with MRGs. Further, the molecular classification based on MRGs of DUs was proposed using unsupervised topic clustering. Then, we investigated the relationship between MRGs and immune infiltration in DUs. Finally, the expression of MRGs in DUs was verified by clinical validations and animal experiments.

2. MATERIALS AND METHODS

2.1. MRGs diagnosis model construction using random forest algorithm

2.1.1. mRNA expression profile collection

Data on the expression profiles of mRNA in DU patients and controls were downloaded from the GEO database (https://www.ncbi.nlm.nih.gov/geo/, GSE80178, GSE68183, and GSE37265). The median of multiple probes for a gene was taken as the gene expression. The k‐nearest neighbor method of the DMwR2 package (version: 0.0.2) of R software was used to compensate for the missing values. Furthermore, the batcom method of the sva package (version: 3.44.0) in R was manipulated to handle the batch effect of the three sets.

2.1.2. Metabolomic regulators obtained

From the Kyoto Encyclopedia of Genes and Genomes (KEGG), Reactome, Human‐GEM, and BRENDA databases, ¹⁷ 4120 metabolomics families were collected, of which 40 were associated with DUs (log2FC > 1, P < 0.05). The location of these genes was obtained using the R packet, and the location relationship of the 40 genes was characterised by the circlize package (version: 0.4.14) of R software.

2.1.3. Expression levels of MRGs between DU and NC

Spearman correlation analysis determined the correlation between two continuous variables with the non‐normal distribution. A P‐value of less than 0.05 was considered statistically considerable. ¹⁸ , ¹⁹ , ²⁰ , ²¹ Additionally, spearman's correlation analysis was used to view correlations in metabolomic expression. The Wilcox test was applied to compare differences in expression levels between DU and negative control (NC) tissue samples.

2.1.4. MRGs‐based machine learning model established using random forest

Applied the randomForest (version: 4.7‐1.1) package in R software factorized the predictors, ²² split the data, and split them into training sets and validation sets according to 7:3. Three 10‐fold model pre‐trainings were carried out, and final model training was carried out, the diagnostic contribution of MRGs was sorted, and the top eight genes with the “Gini score” were used as hub genes. Test set predictions were performed, the validation set was performed with ROC, and AUC was used to evaluate the model diagnostic performance. The violin diagram was used to compare the differences in the expression of hub genes in DU and NC tissue samples.

2.2. Molecular classification identification based on MRGs

2.2.1. Consensus clustering analysis

Manipulated the MRGs of the random forest model, the DU samples was subtyped using the R‐package NMF (version: 0.23.0), and the point with the greatest variation in the cophenetic coefficient was selected as the optimal number of subtypes. In addition, the effect of the subtype division was visualized using PCA, and applied the Wilcox test to compare scores. ROCs were used to test the predictive effectiveness of a model.

2.2.2. Characteristics of biological functional among molecular classification

To verify potential targets' underlying functionality, we performed Gene Ontology (GO) and KEGG functional enrichment analyses on subtypes using clusterProfiler (version: 4.2.1) package in R. The GO tool was extensively utilized to annotate genes with functions, including molecular functions (MFs), biological pathways (BPs), and cellular components (CCs). In addition to providing useful high‐level genomic functional information, KEGG enrichment analysis was a valuable analytical tool for gene function analysis. ²³

2.3. Association between metabolism and immune infiltration

The biomarkers for 28 immune cells from various literature and the degree of immune cell infiltration were determined using the ssGSEA method of the GSVA (version: 1.42.0) package in R. From the Immport database, immune‐related pathways were identified. Further, the immune‐associated pathway scores were assessed using the ssGSEA method of the GSVA package in R. Wilcox test was used to compare DUs with healthy immune cell infiltration and expression of immune‐related pathways, as well as different subtypes. These features of MRGs and immune microenvironments were assessed using person‐related tests.

2.4. Experimental validation

2.4.1. Clinic specimens

Patients were informed and consented to the acquisition of their skin tissues. The Ethics Committee of Yueyang Hospital of Integrated Traditional Chinese and Western Medicine, Shanghai University of TCM, approved the procedure (Supplementary File 1). DU skin tissue was collected from the wound edge of DU patients' undergoing surgery, and normal skin samples were obtained from healthy non‐diabetic volunteers who underwent plastic surgery. All skin tissues were immediately stored at ‐80°C following excision.

2.4.2. Animals and wound healing experiments

The specific pathogen‐free (SPF) C57BL/6 male mice were provided by Shanghai Slac Laboratory Animal Co., Ltd. (scxk Shanghai 2017‐0005), and all the experimental procedures were approved by the Ethics Committee of Yueyang Hospital of Integrated Traditional Chinese and Western Medicine at Shanghai University of TCM (Supplementary File 2). Kept aseptic at standard temperature (23 ± 2°C), all mice were accessible to water and had a standard or high‐fat diet (supplied by Pu Lu Tong Biological Technology Co., Ltd.). To establish an animal model of DU, C57BL/6 mice were induced by streptozotocin (STZ) as previously reported. ²⁴ Briefly, following 2 weeks of high‐fat feeding, mice received intraperitoneal injections of 0.2 mL STZ dissolved in 0.1 M sodium citrate buffer twice every other day, and those with blood glucose levels over 16.7 mmol/L were used for the establishment of DU models.

Followed by wound experiments, using 8‐10‐week‐old mice to generate skin wounds, as previously reported. ²⁵ Anaesthetised mice's dorsal skin was excised of four full‐thickness wounds using sterile 6‐mm‐diameter dermal punches. A collection of wound skin tissues was performed on day 9 for further investigation.

2.4.3. Quantitative real‐time polymerase chain reaction

Quantitative real‐time polymerase chain reaction (qRT‐PCR) was applied to determine mRNA expression levels of MRGs identified by the random forest algorithm. The skin tissues of DU patients and mice with DU were taken, and the total RNA was extracted according to the Trizol reagent (Beyotime, China) protocol. The quality of the extract was also evaluated using agarose gel electrophoresis after its concentration and purity were determined by ultraviolet spectrophotometer. The relative quantification method (2^−ΔΔCT) was used. For cDNA preparation, reverse transcriptase was used. The results were normalized to β‐actin. Primers have been designed and synthesized, and the primers are shown in Table 1 and Table 2.

TABLE 1.

Primers for qRT‐PCR of DU patient skin tissues.

Gene name	Forward primer	Reverse primer
CFB	ACTGTCCAAGACCACACGAC	CACTCCATCGGCCATTCACT
GALNT6	CCTTGGCGCTTACGAGATGA	ATGATTGGCCCTGGATGCAA
GLDC	AAGAAACATCTCGCCCCGTT	GCAGTTTCCGTGGCTTGTTT
KLK6	TGTGGTGACACACGCTGTAG	CTTGAGTCGGGGGAAGGAAC
LYN	CAGCTTGAGTGACGATGGAGT	GAAAGACAAGTCGTCCGGGT
MMP12	CGGGCAACTGGACACATCTA	AGCTTTCCGGATTGCGTAGT
RHOH	TGCCAGCCCAACTGTTGTAT	GTGAAAACAAAAGCGGCCCA
XDH	GGCCAAGGCCTTCATACCAA	GTTGGGCACAGTGTTAGTGC
β‐Actin	GTCATTCCAAATATGAGATGCGT	TGTGGACTTGGGAGAGGACT

TABLE 2.

Primers for qRT‐PCR of STZ‐induced DU mice tissues.

Gene name	Forward primer	Reverse primer
Cfb	TCTACAGGAGATTCCGGGGG	GAAGTCCCGGGCATAAGAGG
Galnt6	CGTCGGGAAACCCACGAG	CTCTCTGGATCCAATCCGGC
Gldc	ACGGAGATCAGCACTTGGTC	GGGTCAGCCCTTTGAACGTA
Klk6	GCATCCAAGTGCCCCATATTT	AGGTTGGAGGGGAATGACAC
Lyn	TAACCGAAGTCACCGTGGAG	TGGACGTTGGATCTCTCACA
Mmp12	AACCGTGTGCAATGCTTGTG	TGAGTTGTCCAGTTGCCCAG
Rhoh	TCGGCACAGGAACTTGCTAAT	AAAAGCTGCAACACCGGACT
Xdh	CGATGACGAGGACAACGGTAG	CTGAAGGCGGTCATACTTGGA
β‐Actin	CAATTCCATCATGAAGTGTGAC	CCACACAGAGTACTTGCGCTC

2.5. Statistical methods

The ggplot2 and pheatmap software packages, along with R version 4.2.1 and R foundation for statistical computing (2022), were used to implement all the above analysis methods. Statistical analysis of wholly experimental data was conducted using SPSS 24.0 (IBM Corp., Armonk International Business Machines, New York, United States). Results were presented as mean standard deviation (SD). Two‐way ANOVA was used to compare the two groups. Remarkable differences between the two groups were indicated by **P < 0.01, *P < 0.05, or ***P < 0.001, two‐sided.

3. RESULTS

3.1. Study design

Figure 2 shows a summary representing this work. Firstly, we constructed an MRGs‐based DUs diagnosis model using machine learning (random forest), which screened 8 potential biomarkers related to metabolomics. Secondly, consensus clustering analysis was used for molecular classification, dividing the samples into 3 subtypes to determine whether the clusters differed in biological characteristics and immune infiltration levels. Thirdly, the relationship between immune cells and the screened MRGs was confirmed. Finally, qRT‐PCR was utilized to analyze the expressions of MRGs in DU patients and STZ‐induced DU mice tissue samples for clinical and experimental validation.

FIGURE 2.

Literature summary diagram and research flow of metabolomics in DUs. Article framework and workflow. MRGs, metabolism‐related genes. DE, differentially expressed.

3.1.1. MRGS diagnosis model construction using random forest algorithm

Data cleaning

To assess the role of metabolomics regulator expression in DUs, we merged three datasets (GSE80178, GSE68183, and GSE37265) for metabolomic expression level screening. The position of MRGs on chromosomes is displayed in Figure 3A. Correlation analysis showed that the remarkable positive correlations are UBE2L6 and PLAAT4 (R = 0.97, P = 2.2e−16) (Figure 3B). Among the metabolomic differential genes, 32 were upregulated, while 7 were downregulated (Figure 3C‐E). The results suggested a difference in the expression of metabolomics regulators between DU samples and normal samples.

FIGURE 3.

Metabolomic expression levels. (A) Location of 40 metabolomic genomes. (B) Correlation of metabolomic expression. (C) MRGs expression heatmap. (D) Metabolomics volcano plot, blue dots are up‐regulated genes, green are down‐regulated, and grey are insignificant. (E) MRGs expression in DU and normal samples. * < 0.05, ** < 0.01, *** < 0.001, ns not significant.

Construction of machine learning model

We examined the clinical significance of metabolomic analysis of DUs, and a machine learning model (random forest) was used to construct a metabolomic regulator‐based diagnostic model of DUs (Figure 4A). After the machine learning model scored the gene contribution (Gini score), a total of 8 metabolism‐related hub genes were obtained: glycine decarboxylase (GLDC), polypeptide n‐acetylgalactosaminyltransferase 6 (GALNT6), ras homologue family member h (RHOH), xanthine dehydrogenase (XDH), matrix metallopeptidase 12 (MMP12), kallikrein‐related peptidase 6 (KLK6), lyn proto‐oncogene, src family tyrosine kinase (LYN), complement factor B (CFB) (Figure 4B,C). The ROC curves in the validation set showed that the 8 MRGs could distinguish the DUs from normal tissue samples (Figure 4D). Their AUC area was almost equal to 1, suggesting the model had an excellent predictive performance. The hub genes were considerably overexpressed in DU vs NC tissue samples (Figure 4E).

FIGURE 4.

Model building and validation. (A) Machine learning model of MRGs. (B) Evaluate gene contribution according to Gini score. (C) Cross‐validation of the predictors of the random forest model. (D) ROC curve to evaluate the predictive performance of the model. (E) Differential expression of hub genes in DU and NC samples.

3.1.2. Molecular classification

Consensus clustering construction

To investigate the association between the molecular classification of MRGs in DUs, a consensus clustering algorithm was adopted. According to consistent clustering results, when the division point was 3, the cophenetic coefficient changed the most, where the DU samples were best divided into three subtypes for further analysis (Figure 5A‐C). The PCA plot indicated that the DU samples could be distinguished from the three groups of features (Figure 5D). The MRGs expression of the three subtypes revealed that 35 genes different notably (Wilcox test, P < 0.05, Figure 5E,F). The findings showed that DU samples could be consensus clustered into three subtypes by MRGs, and there were appreciable differences in MRGs expression among the subtypes.

FIGURE 5.

Consensus clustering. (A) Samples were classified by NMF clustering by MRGs in the model, ranging from 2 to 10 classes. (B) The distribution of cophenetic coefficients at different points of NMF, the point with the largest change is the optimal threshold point. (C) Consistent clustering plot of the three subtypes. (D) PCA plots of the three subtypes. (E) The expression difference of MRGs among the three subtypes in the model. (F) The expression of MRGs in the model in the three subtypes is represented by a heatmap. * < 0.05, ** < 0.01, *** < 0.001, ns is not significant.

Differences in biological function between subtypes

We investigated the differences in the biological functions of the three groups of metabolomic genotypes and performed an enrichment analysis of GO and KEGG. Cluster1 and Cluster2 significantly differed in regulation of response to stimulus and cytokine‐cytokine receptor interaction (Figure 6A); Cluster2 and Cluster3 remarkably differed in immune system process (Figure 6B); Cluster3 and Cluster1 appreciably differed in enriching associated with NF‐kappa B signaling pathway (Figure 6C). Altogether, this study highlighted the functional differences between metabolism molecular subtypes and uncovered a molecular network associated with these functions.

FIGURE 6.

Biological functions of different metabolomic genotypes. (A) GO classification map, GO bubble map, and KEGG bubble map of cluster 1 and cluster 2. (B) GO classification map, GO bubble map, and KEGG bubble map of cluster 2 and cluster 3. (C) GO classification map, GO bubble map, and KEGG bubble map of cluster 1 and cluster 3.

3.1.3. Correlation between metabolism and immune infiltration

MRGs and immune correlation

After obtaining a clinically considerable metabolomics diagnostic model of DUs, our next step was to examine the relationship between MRGs and the immune microenvironment (Figure 7A). It was found that there was a relation between MRGs and factors and immune infiltration, of which LYN and Type 1 helper cell, LYN with Antimicrobials were remarkably positively correlated; CFB and CD56dim natural killer cell, RHOH with TGF‐β Family Members were significantly inversely correlated (Figures 7B, 8B).

FIGURE 7.

Immune cell infiltration obtained with ssGSEA. (A) Comparison of the differences in immune cell infiltration levels in DU and normal samples. (B) Correlation between immune cell infiltration and MRGs in the model, and the differences in the distribution of immune cells and MRGs in DU and normal samples were extracted. * < 0.05, ** < 0.01, *** < 0.001, ns was not significant.

FIGURE 8.

Immune response pathways and MRGs. (A) Comparison of the differences in immune response pathways between DU and normal samples. (B) Correlation of MRGs in immune response pathways and models, and differences in the distribution of immune cells and MRGs in DU and normal samples were extracted. * < 0.05, ** < 0.01, *** < 0.001, ns is not significant.

Metabolic molecular subtypes and immune infiltrate correlation

In addition, comparing the differences in the degree of immune cell infiltration among the three subtypes, 22 immune cells were found to be significantly different (Wilcox test, P < 0.05). Comparing the three subtypes, cluster 3 had a higher degree of immune infiltration (Figure 9A). In addition, the results showed appreciable differences in 12 immune‐related pathways, and cluster 3 had a higher degree of immune infiltration. Results suggested that all three subtypes were related to immune infiltration, and cluster 3 had a higher degree of immune microenvironment infiltration (Figure 9B).

FIGURE 9.

Comparison of the immune microenvironment of different metabolomics genotypes. (A) Differences in immune cell infiltration among the three subtypes. (B) Differences in immune‐related pathways among the three subtypes. * < 0.05, ** < 0.01, *** < 0.001, **** < 0.0001, ns not significant.

3.1.4. QRT‐PCR validated MRGs expression with DU patient skin tissues and STZ‐induced DU mice

To demonstrate the results of the MRGs‐based learning machine model, we further validated the expression of eight target mRNAs using qRT‐PCR analysis based on skin tissues of DU patients (Figure 10) and mice with DU (Figure S1), the results revealed that the expression of all hub genes was significantly upregulated, including protein metabolism regulator GALNT6, ²⁶ , ²⁷ nucleotide metabolism regulator XDH, ²⁸ and a tyrosine‐protein kinase LYN. ²⁹ Meanwhile, several kernel MRGs were proven to be related to inflammation, including GLDC ³⁰ and MMP12. ³¹ Notably, a close relationship had been found between MRGs and immune cells and immune response. CFB has been reported to be involved in the regulation of the immune response in autoimmune diseases, ³² followed by KLK6, which promoted the progression of inflammation in a variety of inflammatory and autoimmune diseases. ³³ In addition, RHOH played a role in T cell development. ³⁴ Together, these outcomes validated the correlation of metabolically relevant biomarkers learned by the random forest model in DUs with the diagnostic and potential therapeutic value.

FIGURE 10.

qRT‐PCR validated MRGs expression in DU patient skin tissues. The mRNA level of the candidate metabolomics hub genes in DU patient skin tissues compared with the normal skin samples (n = 5). The data are represented as an average ± SD. Four lesions were included in the analysis in each group. * < 0.05, ** < 0.01, *** < 0.001, ns not significant, compared with the control group.

4. DISCUSSION

In our study, we sought to explore the role of metabolic genes in DUs and verify the validity of disease diagnostic models based on MRGs learned from random forests. Moreover, the biological function characteristics of DU subtypes based on MRGs were comprehensively explored through molecular classification. Further, the relationship between immune cells and the screened MRGs was confirmed. Thus, we found that MRGs had potential diagnostic biomarkers value in DUs, and finally verified our view through clinical verifications and in vivo experiments.

Large data repositories can be utilized by machine learning to identify novel risk predictors and how they interact more complexly and define multiple regression models as performance evaluation benchmarks with AUC = 0.7143. ³⁵ Random forest algorithm is one of them whose AUC = 0.787 outperforms other methods in AUC and has a considerable improvement over benchmarks, ³⁵ the primary predictive model in our research to reveal the extent to which metabolic regulatory genes contribute to DUs. The integrated decision tree‐based machine learning algorithm with strong anti‐noise and better stability can be used for feature selection based on features with high prediction accuracy. ³⁶ Our experimental outcomes validated that random forest model results matched the algorithm, which could further reduce the dimension of the variables to construct robust genetic signatures consisting more efficiently of eight MRGs.

The chronic inflammatory process is not conducive to the wound healing of DUs, while metabolism can participate in this link and affect wound healing. ³⁷ MRGs could mediate inflammation development, including CFB, MMP12, and XDH. It has been reported that CFB is a crucial factor involved in the alternative complement pathway, exacerbating the inflammatory response to idiopathic membranous nephropathy concomitant with IgA nephropathy ³¹ ; MMP12, an elastase produced mainly by macrophages, is an important mediator of acute and chronic lung injury and is directly involved in the development of inflammatory reactions (Mohan et al., 2020). ³⁸ The suggestion that CFB and MMP12 may aggravate the DU process is consistent with our findings. Consistently, XDH is a critical enzyme in the purine catabolism pathway, and increased mRNA expression of XDH in DU wounds is associated with excessive ROS production. ³⁹

Similarly, proliferation is also essential to wound healing. Delayed re‐epithelialization and impaired angiogenesis are crucial factors contributing to the complexity of healing DUs. ⁴⁰ Our results revealed that MRGs, including GLDC and KLK6 may affect DUs by mediating angiogenesis and re‐epithelization. GLDC is a pivotal enzyme in the glycine lysis system, whose expression is downregulated in gastric cancer cell lines and tissues, and the proliferation and migration of gastric cancer cells can be inhibited by methylation of GLDC (Min et al., 2016); ⁴¹ KLK6 is a secreted serine protease that downregulates mRNA expression in gliomas and can inhibit tumor growth (Lou et al., 2014). ⁴² The suggestion that GLDC and KLK6 may inhibit wound healing of DUs is consistent with our findings.

In addition, the unique immunological profile of each subtype also validated the feasibility of our approach to classifying the immunophenotypes of various MRGs, suggesting DUs can be subtyped at the molecular or immune level, not just at the phenotypic level, which may help to understand the basic processes of immunoregulation. Henceforward, diagnosing and treating DUs can contemplate metabolic molecular typing, to formulate a more refined and individualized strategy.

We performed a correlation analysis between MRGs and immune infiltration, based on which we found the overall function of MRGs in regulating the immune microenvironment of DUs. For instance, LYN, a tyrosine‐protein kinase, as molecular switches regulating immunoreceptors that direct homeostasis or inflammation ²⁸ ; Type 1 helper T cells (Th1) activate macrophage immune and phagocyte‐dependent protective responses involved in the development of inflammation in organ‐specific autoimmune diseases (Romagnani, 1999). ⁴³ A considerable positive correlation between LYN and Th1 was revealed in the immune microenvironment of MRGs. We hypothesized that LYN may exacerbate the inflammatory response in DUs, inhibiting wound healing.

Likewise, RHOH is a member of the Rho GTPase family and regulates thymocyte development and T‐cell receptor signalling ⁴⁴ ; transforming growth factor β (TGF‐β) as an immunosuppressant stimulates tissue differentiation, angiogenesis, and epithelial‐mesenchymal transformation. ⁴⁵ More importantly, a deficiency of TGF‐β would lead to delayed healing of chronic diabetic wounds (Zubair and Ahmad, 2019), ⁴⁶ whose expression was downregulated in STZ‐induced diabetic mouse models, while DU wounds healing could be accelerated by upregulating the expression of TGF‐β (Sun et al., 2020). In contrast, RHOH and TGF‐β Family Members were appreciably inversely correlated in the immune microenvironment of MRGs. Accordingly, we speculated that RHOH may be detrimental to wound healing in DUs. These speculations are consistent with our results.

A comprehensive understanding of the characteristics of metabolomics is pivotal for elucidating the underlying mechanisms of DUs. The current study systematically investigated the interactions between MRGs and immune responses. However, the specific functions of three MRGs‐based DU subtypes and their core targets should be undertaken to dig deeper into. The functions of hub MRGs during chronic wounds are our future direction.

5. CONCLUSIONS

In the present study, we explored the role of MRGs in DUs through random forest models and experimental validation. Eight MRGs obtained in the random forest were verified by clinical verifications and animal experiments, inhibiting these targets could be targeted to promote DUs healing. In addition, it is worthwhile to classify three DU subtypes based on MRGs, which encourages the implementation of precise diagnostic and treatment strategies. Our work furnishes a reference for delving into the function and significance of MRGs in other sicknesses.

AUTHOR CONTRIBUTIONS

Xiao‐Xuan Ma, Ying Zhang, and Jing‐Si Jiang contributed equally. Yi‐Mei Tan and Le Kuai conceived and designed the study. Jian‐Kun Song, Yue Luo, Xiao‐Ya Fei, and Yi Ru performed data curation. Ying Luo, Xin Ma, and Bin Li performed experimental work. Xiao‐Xuan Ma, Ying Zhang, and Jing‐Si Jiang prepared the original draft. Yi‐Mei Tan and Le Kuai reviewed and edited the manuscript. All authors have read and approved the final manuscript.

FUNDING INFORMATION

This study was supported by the NSFC of China (No. 82204954, 82004235, 82174383, 81973860), the National Key Research and Development Program of China (No. 2018YFC1705305), Shanghai Clinical Key Specialty Construction Project (shslczdzk05001), Shanghai Development Office of TCM (ZY(2018‐2020)‐FWTX‐1008, ZY(2021‐2023)‐0302), Shanghai Science and Technology Committee (21Y21920101, 21Y21920102), Youth Talent Promotion Project of China Association of Traditional Chinese Medicine (2021‐2023) Category A (CACM‐2021‐QNRC2‐A10), Health Young Talents of Shanghai Municipal Health Commission (2022YQ026), Xinglin Youth Scholar of Shanghai University of Traditional Chinese Medicine (No. RY411.33.10), “Chen Guang” project supported by Shanghai Municipal Education Commission and Shanghai Education Development Foundation (22CGA50), Shanghai Sailing Program (No. 21YF1448100, 22YF1450000, 22YF1441300, 23YF1439800), Clinical research and cultivation project in hospital and Clinical transformation incubation program in hospital (No. lczh2021‐05).

CONFLICT OF INTEREST STATEMENT

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as potential conflicts of interest.

Supporting information

Data S1. Supporting Information.

Supplementary File 1. Approved by the Ethics Committee of Yueyang Hospital of Integrated Traditional Chinese and Western Medicine Affiliated to Shanghai University of Traditional Chinese Medicine.

Supplementary File 2. The laboratory animal ethics review approval of Yueyang Hospital of Integrated Traditional Chinese and Western Medicine of Shanghai University of Traditional Chinese Medicine Ethics Committee.

Supplementary Table 1. Primers for qRT‐PCR of DU patient skin tissues.

Supplementary Table 2. Primers for qRT‐PCR of STZ‐induced DU mice tissues.

Supplementary Figure 1. qRT‐PCR validated MRGs expression in STZ‐ induced mice. The mRNA level of the candidate metabolomics hub genes in STZ‐induced DU mice compared to the control group (n = 5). The data is represented as an average ± SD. Four lesions were included in the analysis in each group. * < 0.05, ** < 0.01, *** < 0.001. ns, not significant, compared to the control group.

Data S4. Supporting Information.

Ma X‐X, Zhang Y, Jiang J‐S, et al. Metabolism‐related biomarkers, molecular classification, and immune infiltration in diabetic ulcers with validation. Int Wound J. 2023;20(9):3498‐3513. doi: 10.1111/iwj.14223

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available from the corresponding author upon reasonable request.