Machine learning identifies plasma-derived exosomal miRNAs associated with low muscle mass in Tibetan patients with rheumatoid arthritis

Abstract

Sarcopenia poses a significant health burden in aging populations, particularly among patients with rheumatoid arthritis (RA), but current diagnostic approaches face limitations in early detection. This study aims to identify plasma-derived exosomal miRNA features associated with low muscle mass in Tibetan patients with RA. High-throughput sequencing identified 101 candidate dysregulated miRNAs in Tibetan patients with RA and low muscle mass, including 61 upregulated and 40 downregulated miRNAs, using exploratory thresholds of |log2 fold change| ≥ 0.58 and nominal P < 0.05. Most candidates did not remain significant after Benjamini–Hochberg false discovery rate correction; therefore, the broader candidate set was considered exploratory. Machine learning algorithms consistently identified hsa-miR-1294 and hsa-let-7c-5p as candidate miRNA features. Integrated miRNA–mRNA regulatory network analysis retained 50 candidate target genes for the construction of two direction-specific putative interaction networks involving 43 miRNAs. Functional enrichment analysis identified nominally enriched metabolic pathways, including sphingolipid metabolism and oxidative phosphorylation, while immune microenvironment analysis revealed differences in estimated muscle satellite-cell and mast-cell enrichment scores. Internally evaluated miRNA-based and target gene-based nomogram models both showed preliminary apparent discrimination, with an area under the receiver operating characteristic curve of 0.905 in their respective development datasets. Exploratory computational drug screening identified several candidate compounds for future investigation, including deferasirox, acarbose, taxifolin, pterostilbene, and ercalcitriol. This study prioritized hsa-miR-1294 and hsa-let-7c-5p as candidate exosomal miRNA features associated with low muscle mass in Tibetan patients with RA. These findings provide hypothesis-generating evidence for the further evaluation of minimally invasive miRNA features associated with low muscle mass.

Introduction

Sarcopenia is a progressive and generalized skeletal muscle disorder involving the accelerated loss of muscle mass and function¹. Beyond incurring significant economic costs each year, sarcopenia is linked to heightened risks of falls, fractures, functional impairment, hospitalization, and mortality, thereby profoundly affecting quality of life and independence^2,3. Notably, the prevalence of sarcopenia in the high altitude population over 60 years old was 17.2% in men and 36.0% in women, higher than that reported in low-altitude populations⁴. This elevated prevalence is particularly relevant to Xizang, China, where a substantial proportion of the population is chronically exposed to high-altitude environments. In the context of rapid population ageing, sarcopenia is therefore emerging as an increasingly important public health concern in the region.

Rheumatoid arthritis (RA), an immune-mediated inflammatory disease characterized by chronic and painful joint inflammation, affects approximately 1% of adults worldwide. Sarcopenia is a notable comorbidity of RA, which affects approximately 25% of individuals with RA⁵. Low skeletal muscle mass and density negatively affect the physical performance and level of disability of patients with RA^6,7, with muscle reduction independently predicting poorer quality of life in low-BMI RA patients⁸. These factors collectively identify patients with RA as a critical cohort for sarcopenia research.

Early detection of sarcopenia is crucial for effective management but hampered by non-specific clinical manifestations. Diagnosis relies on assessing three core indicators: muscle mass, muscle strength and physical performance. According to the 2019 Asian Working Group for Sarcopenia (AWGS), both muscle strength and physical performance decline are consequences of reduced muscle mass and have adverse effects on prognosis⁹. Low muscle mass is therefore an important component of sarcopenia assessment, but it alone does not establish a clinical diagnosis of sarcopenia. In the present study, participants were classified according to CT-derived skeletal muscle index; accordingly, the primary phenotype investigated was low muscle mass rather than clinically diagnosed sarcopenia.

Exact quantification of skeletal muscle mass remains challenging. Established methods such as computed tomography (CT), magnetic resonance imaging (MRI), and dual-energy X-ray absorptiometry (DXA)¹⁰ are all costly and largely unavailable in primary care settings, especially in remote and resource-constrained regions like Xizang. Furthermore, these methods assess structural changes occurring relatively late in disease progression, potentially missing opportunities for early intervention¹¹. Therefore, there is a critical need to identify accessible, non-invasive biomarkers for the early sarcopenia detection.

Blood-based biomarkers offer advantages including ease of collection, cost-effectiveness, and potential for serial monitoring^12,13, but conventional protein biomarkers lack adequate sensitivity and specificity for routine clinical application^12,13. MicroRNAs (miRNAs) have emerged as promising biomarker candidates. These small, non-coding RNA molecules (18–25 nucleotides) regulate gene expression post-transcriptionally and play crucial roles in muscle development, differentiation, and maintenance^14,15. Specific miRNAs, termed myomiRs (including miR-1, miR-133a, and miR-29b), are highly enriched in skeletal muscle and regulate fundamental aspects of muscle biology¹⁶. Circulating miRNAs demonstrate remarkable stability in biological fluids due to their association with protein complexes and encapsulation within extracellular vesicles, particularly exosomes¹⁷.

Exosomes (30–150 nm diameter) serve as important mediators of intercellular communication by transferring bioactive molecules¹⁸. Exosomal miRNAs offer advantages over total circulating miRNAs, including enhanced stability, reduced interference from cellular debris, and potential for tissue-specific origin identification¹⁹. Recent evidence suggests muscle-derived exosomes play important roles in muscle homeostasis and pathological responses²⁰, making circulating exosomal miRNAs potential minimally invasive biomarkers reflecting early muscle changes.

Therefore, this study employed a comprehensive approach integrating molecular profiling, bioinformatics analysis, and machine learning methodologies to characterize exosomal miRNA expression profiles in Tibetan rheumatoid arthritis patients with low muscle mass. High-throughput small RNA sequencing was conducted, miRNA data were integrated with transcriptomic datasets to construct regulatory networks and elucidate functional pathways. Multiple machine learning algorithms were employed to prioritize candidate miRNA features associated with low muscle mass, and nomogram-based internally evaluated classification models for identifying low muscle mass were developed for internal evaluation of low muscle mass identification. Through this multifaceted investigation, we aimed to provide novel insights into molecular features associated with low muscle mass in patients with RA while providing a basis for future evaluation of minimally invasive biomarkers and classification approaches for low muscle mass.

Methods

Study design and participants

This case-control study compared plasma exosomal miRNA profiles between Tibetan patients with rheumatoid arthritis (RA) with low muscle mass and those with normal muscle mass. Participants were recruited from Hospital of Chengdu Office of People’s Government of Xizang Autonomous Region between August 2023 and March 2024. Subjects were classified into a low muscle mass group and a normal muscle mass group based on skeletal muscle mass index (SMI) measured by CT according to an SMI < 50 cm2/m2 for men and < 39 cm2/m2 for women²¹. Although handgrip strength and gait speed were collected when available, these measures were not available for all participants and were therefore not used as formal grouping criteria. Accordingly, the primary phenotype investigated in this study was low muscle mass rather than clinically diagnosed sarcopenia.

Sample collection and exosome isolation

Peripheral blood samples (10 mL) were collected from participants after overnight fasting using EDTA-containing tubes. Blood samples were centrifuged at 1,200 × g for 10 min at 4℃ to separate plasma, which was then centrifuged again at 1,800 × g for 10 min to remove cellular debris. The resulting plasma was aliquoted and stored at −80℃ until exosome isolation. Exosomes were isolated from 500 µL plasma samples using an exosome extraction kit (Yeasen Biotechnology, Shanghai, China) according to the manufacturer’s protocol. Briefly, 400 µL of pre-cooled Solution P1 was added to remove contaminating proteins, followed by centrifugation at 12,000 rpm for 20 min at 4℃. After adding 120 µL of Solution P2 for precipitation and centrifugation at 12,000 rpm for 15 min at 4℃, the pellet was resuspended in 200 µL of 1×PBS. Further purification was performed using Exosome Purification Columns by centrifugation at 5,600 rpm for 20 min at 4℃. Purified exosomes were aliquoted and stored at −80℃ for downstream analysis.

Exosome characterization

Isolated exosomes were characterized using transmission electron microscopy (TEM), nanoparticle tracking analysis (NTA), and Western blot analysis according to International Society for Extracellular Vesicles guidelines. For TEM analysis, exosome samples were negatively stained with 2% phosphotungstic acid and examined using a JEM-1400FLASH transmission electron microscope (JEOL, Tokyo, Japan). Size distribution and concentration were quantified using NanoSight NS300 (Malvern Panalytical) with triplicate measurements. Protein characterization was performed by Western blot analysis to confirm exosome identity and purity. Equal amounts of protein (20 µg) were separated by 12% SDS-PAGE using Easy PAGE gel preparation kit (Seveninnovation Biotechnology, Beijing, China) and transferred to PVDF membranes (Bio-Rad, Hercules, CA, USA). After blocking with 5% non-fat milk for 2 h at room temperature, membranes were incubated overnight at 4℃ with primary antibodies against exosome-specific markers including CD63 (Wanleibio, 1:500), HSP70 (Bioss, 1:500), and TSG101 (Wanleibio, 1:500). Calnexin (Wanleibio, 1:500), an endoplasmic reticulum marker, was used as a negative control to confirm the absence of cellular contaminants. HRP-conjugated goat anti-rabbit IgG (ZSGB-BIO, 1:5000) was applied for 1 h at room temperature. Protein bands were visualized using enhanced chemiluminescence detection (Servicebio, Wuhan, China) and imaged using a Bio-Rad XRS + gel imaging system.

RNA extraction and small RNA sequencing

Total RNA was extracted from exosomes using the miRNeasy Serum/Plasma Kit (Qiagen) following the manufacturer’s protocol. RNA concentration was measured using Qubit RNA HS Assay Kit (Thermo Fisher Scientific), and RNA integrity was assessed using Agilent 2100 Bioanalyzer with the Small RNA Kit. Small RNA libraries were constructed using the NEBNext Multiplex Small RNA Library Prep Set for Illumina (New England Biolabs). Briefly, 3’ and 5’ adapters were ligated to small RNAs, followed by reverse transcription and PCR amplification. Library quality was assessed using Agilent 2100 Bioanalyzer, and quantification was performed using qPCR. Libraries were pooled and sequenced on Illumina NovaSeq 6000 platform with single-end 50 bp reads, targeting 10–20 million reads per sample.

miRNA sequencing data processing and analysis

Raw sequencing reads were subjected to quality control using FastQC v0.11.9. Adapter sequences and low-quality reads (Phred score < 20) were removed using Cutadapt v3.4. Reads shorter than 18 nucleotides or longer than 30 nucleotides were filtered out. Clean reads were aligned to the human reference genome (GRCh38) using bowtie2 v2.4.4 with default parameters. miRNA annotation was performed using miRBase v22.1. Read counts for each miRNA were obtained using featureCounts from the Subread package v2.0.1. Differential expression analysis was conducted using the limma package (v3.46.0) in R v4.1.0. miRNAs with mean counts < 10 were excluded. P values were adjusted using the Benjamini–Hochberg procedure. An adjusted P value < 0.05 was used to define FDR-significant differential expression. For hypothesis-generating downstream analyses, miRNAs with |log2 fold change| ≥ 0.58 and nominal P < 0.05 were retained as exploratory candidate dysregulated miRNAs.

Public dataset acquisition and processing

Sarcopenia-related bulk RNA sequencing datasets were systematically searched and downloaded from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/). Datasets were included if they contained at least 10 samples per group (sarcopenia vs. control). The dataset GSE111016 was selected, which comprises transcriptomic data generated on the GPL16791 platform (Illumina HiSeq 2500), including 20 sarcopenia muscle tissue samples and 20 normal control samples. Raw count matrices were normalized and subjected to differential expression analysis using the limma package (v3.46.0) in R v4.1.0. For hypothesis-generating integration, genes with |log2 fold change| ≥ 0.58 and nominal P < 0.05 were retained as exploratory candidate differentially expressed genes. These public muscle transcriptomic data were used for hypothesis-generating integration with the plasma exosomal miRNA results.

Integrative miRNA-mRNA analysis

miRNA target prediction was performed using MiRanda and RNAhybrid algorithms to identify potential mRNA targets of the candidate dysregulated miRNAs. MiRanda utilizes local alignment algorithms for genome-wide miRNA target prediction, emphasizing evolutionary conservation of miRNA-target binding sites and employing RNAFold to calculate thermodynamic stability. RNAhybrid predicts energetically favorable miRNA-target duplexes by identifying optimal target sitesaccording to binding energy while prohibiting intramolecular base pairing within the individual sequences . Only targets predicted by both algorithms were retained for downstream analysis to ensure prediction reliability. The predicted targets were subsequently integrated with candidate dysregulated genes identified in an independent skeletal muscle transcriptomic dataset. Because the circulating exosomal miRNA and skeletal muscle mRNA data were obtained from different participants and tissue sources, this cross-dataset integration was considered hypothesis-generating and did not provide evidence of direct miRNA–mRNA regulation or sample-level correlation. Predicted miRNA targets were then intersected with oppositely expressed genes to identify putative inverse-expression miRNA–mRNA pairs: upregulated miRNAs were matched with downregulated mRNAs, whereas downregulated miRNAs were matched with upregulated mRNAs. Two direction-specific regulatory networks were constructed to preserve the direction of the predicted relationships—an upregulated-miRNA/downregulated-mRNA network and a downregulated-miRNA/upregulated-mRNA network—and were visualized using Cytoscape version 3.8.2.

Functional annotation and pathway analysis

Functional enrichment analysis of the candidate target genes was performed using the clusterProfiler (v4.0.0) package in R. Gene Ontology (GO) enrichment analysis was conducted for the three main categories: Biological Process (BP), Cellular Component (CC), and Molecular Function (MF), to comprehensively investigate the functional characteristics of the target genes. Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis²² was also carried out to identify relevant signaling and metabolic pathways. Given the exploratory nature of the analysis, GO terms and KEGG pathways with nominal P < 0.05 were considered nominally enriched. In addition, Gene Set Enrichment Analysis (GSEA) was implemented in R using clusterProfiler with KEGG pathway gene sets, and normalized enrichment scores (NES) and nominal p-values were calculated. Pathways with nominal P < 0.05 and |NES| > 1 were defined as exploratory enrichment signals.

Machine learning-based feature selection and biomarker identification

Three machine-learning algorithms were applied to explore candidate miRNA feature rankings associated with low muscle mass using R version 4.1.0. Random Forest feature selection was conducted using the randomForest package (v4.6–14) with 1,000 decision trees and default parameters. Variable importance scores were calculated using the mean decrease in Gini impurity criterion to rank miRNAs from the constructed miRNA-mRNA regulatory networks. Extreme Gradient Boosting (XGBoost) was implemented using the xgboost package (v1.4.1.1) with parameters: maximum tree depth = 6, learning rate = 0.3, subsample ratio = 0.8, and 100 boosting rounds. Feature importance was assessed using the gain metric, and variable relative importance scores were calculated and normalized. Support Vector Machine analysis was performed using the e1071 package (v1.7–8) with RBF kernel. Recursive Feature Elimination was implemented using the caret package (v6.0–88) to evaluate feature subsets ranging from 1 to 40 miRNAs, with model performance assessed using 10-fold cross-validation to determine optimal feature numbers. Feature rankings from all three methods were compared using Venn diagram analysis to identify overlapping candidate miRNAs. miRNA-target gene regulatory pairs were constructed by cross-referencing the most important miRNAs with candidate dysregulated genes from transcriptomic datasets. It should be noted that, given the limited sample size (n = 20), Random Forest and XGBoost feature importance rankings were derived from the complete dataset rather than within a nested cross-validation framework, and the potential for information leakage between feature selection and model evaluation is acknowledged as an inherent limitation of the current study.

Nomogram-based model development and internal evaluation

Two exploratory nomogram models were developed to evaluate the identification using molecular features derived from the bioinformatics and machine-learning analyses. The miRNA-based nomogram was developed and internally evaluated in the sequencing cohort (n = 20), using low muscle mass status as the outcome. The target gene-based nomogram was developed and internally evaluated using GSE111016 (n = 40), using sarcopenia status as the outcome. Therefore, the two models were developed in separate datasets and evaluated different but related phenotypes. Multivariable logistic regression was performed for model construction using the “rms” package in R, and nomograms were generated to visualize individualized risk prediction. Apparent discriminatory performance was internally assessed using receiver operating characteristic (ROC) curves and calculating the area under the curve (AUC) with the “pROC” package. Decision curve analysis, implemented with the “rmda” package, was used to determine the exploratory decision-curve results of each nomogram across a spectrum of threshold probabilities.

Immune microenvironment analysis

Estimated cell-type enrichment scores in muscle tissue samples were calculated using the xCell algorithm, which generates enrichment scores for 64 immune and stromal cell types from bulk RNA-sequencing data. xCell analysis was performed on the normalized gene expression matrix from the GSE111016 dataset using the xCell R package (v1.1.0). Enrichment scores for each cell type were calculated for individual samples. Differences in estimated cell-type enrichment scores between participants with sarcopenia and normal controls were assessed using the Wilcoxon test, with nominal P < 0.05 used as the exploratory threshold. Cell types showing differences in enrichment scores were selected for further analysis. Pearson correlation analysis was performed to examine associations between these enrichment scores and target-gene expression levels.

Drug screening and molecular docking analysis

The Connectivity Map database was used to prioritize computationally predicted candidate therapeutic compounds with expression signatures inversely associated with the sarcopenia-related gene signature. Gene expression profiles of candidate dysregulated target genes were uploaded to the CMap portal (https://clue.io/), and connectivity scores were calculated to identify compounds with opposing expression patterns to the sarcopenia signature. Compounds with negative connectivity scores indicating inverse expression-profile associations were ranked and selected for further investigation. To assess the binding interactions between predicted compounds and target proteins, molecular docking simulations were performed using AutoDock Vina. Three-dimensional protein structures of key target genes were retrieved from the Protein Data Bank (PDB), and small molecule structures of the top-ranked compounds were obtained from PubChem database. Molecular docking was performed to evaluate binding affinities and predict protein-ligand interactions. Binding affinities were expressed as docking scores (kcal/mol), with more negative values indicating stronger predicted binding.

Results

Participant characteristics and sample quality

The baseline demographic, anthropometric, RA-related, medication, muscle-related, and inflammatory characteristics are summarized in Table 1. Twenty Tibetan patients with RA were included, comprising 14 participants with low muscle mass and six with normal muscle mass. No statistically significant between-group differences were detected in demographic characteristics, RA-related features, medication history, inflammatory markers, or metabolic comorbidities (all P > 0.05). As defined by the prespecified classification criterion, mean SMI was lower in the low muscle mass group than in the normal muscle mass group (35.88 ± 5.85 vs. 42.29 ± 2.21 cm²/m²).

Table 1.

Baseline demographic and clinical characteristics of Tibetan patients with rheumatoid arthritis stratified by muscle mass status.

Variable	Overall N = 201	Muscle mass		p-value2
		Normal muscle mass N = 6 (30%)1	Low muscle mass N = 14 (70%)1
Age, years	53.45 (9.91)	48.50 (9.99)	55.57 (9.43)	0.17
Sex				> 0.99
Female	19 (95.00%)	6 (100.00%)	13 (92.86%)
Male	1 (5.00%)	0 (0.00%)	1 (7.14%)
BMI, kg/m²	24.14 (3.56)	25.16 (5.27)	23.70 (2.68)	0.54
Waist circumference, cm	87.50 (11.22)	86.83 (10.74)	87.79 (11.81)	0.86
Calf circumference, cm	33.50 [29.75, 35.50]	33.75 [31.00, 36.00]	33.50 [29.50, 35.00]	0.68
SARC-CalF score	2.50 [0.00, 12.00]	7.00 [1.00, 12.00]	1.00 [0.00, 12.00]	0.61
RA disease duration, years	7.00 [4.00, 15.00]	4.00 [3.00, 9.00]	9.00 [6.00, 22.00]	0.17
RAPID-3 score	9.42 (3.45)	9.43 (3.84)	9.41 (3.42)	0.99
Medication history
Any antirheumatic medication use, n (%)	15 (75.0%)	5 (83.3%)	10 (71.4%)	1.000
csDMARD use, n (%)	14 (70.0%)	5 (83.3%)	9 (64.3%)	0.613
Biological DMARD use, n (%)	3 (15.0%)	1 (16.7%)	2 (14.3%)	1.000
Glucocorticoid exposure, n (%)	10 (50.0%)	2 (33.3%)	8 (57.1%)	0.628
Analgesic use, n (%)	5 (25.0%)	1 (16.7%)	4 (28.6%)	1.000
SMI, cm²/m²	37.80 (5.82)	42.29 (2.21)	35.88 (5.85)	0.002
ESR, mm/h	37.00 [24.00, 69.00]	29.50 [23.00, 57.00]	50.00 [25.00, 69.00]	0.41
TNF, pg/mL	16.50 [10.20, 27.55]	15.10 [8.92, 25.10]	19.55 [10.40, 29.60]	0.55
CRP, mg/L	14.38 [3.92, 59.84]	9.32 [5.83, 36.47]	29.11 [6.89, 59.84]	0.33
Diabetes mellitus				0.49
Insulin resistance	1 (5.00%)	1 (16.67%)	0 (0.00%)
No	15 (75.00%)	4 (66.67%)	11 (78.57%)
Yes	4 (20.00%)	1 (16.67%)	3 (21.43%)

¹Mean (SD); n (%); Median [Q1, Q3]

²P-values were calculated using Welch’s two-sample t-test for normally distributed continuous variables, Wilcoxon rank-sum test for non-normally distributed continuous variables, and Fisher’s exact test for categorical variables. For categorical variables with sparse cells, Fisher’s exact test with Monte Carlo simulation was used based on 2,000 replicates.

Exosome characterization and RNA quality assessment

Isolated exosomes displayed characteristic morphological features when observed under transmission electron microscopy, presenting as spherical to cup-shaped vesicles with well-defined membrane structures (Fig. 1A). Nanoparticle tracking analysis revealed consistent size distributions between groups, with mean particle sizes showing no statistically significant difference between the low muscle mass group and the normal muscle mass group (Fig. 1B). The particle size distributions exhibited characteristic exosomal profiles with peak concentrations occurring around the expected range for exosomes, confirming successful isolation of extracellular vesicles. Western blot analysis confirmed the molecular identity and purity of isolated exosomes through detection of established exosomal protein markers (Fig. 1C). The analysis revealed robust expression of CD63, HSP70, and TSG101. Importantly, Calnexin, an endoplasmic reticulum-resident protein, showed minimal detection, confirming the absence of cellular contaminants and validating exosome purity.

Fig. 1.

Characterization of plasma-derived exosomes from individuals with low muscle mass and normal muscle mass. (A) Transmission electron microscopy images of isolated exosomes from representative samples (N1, N2: normal muscle mass group; L1, L2: low muscle mass group). Scale bar = 100 nm. (B) Nanoparticle tracking analysis showing size distribution profiles of exosomes from the same representative samples. (C) Western blot analysis of exosome-associated protein markers from the same representative samples. Protein lysates were probed with antibodies against Calnexin, CD63, HSP70, and TSG101. Molecular weights are indicated on the right.

Small RNA sequencing data overview

Small RNA sequencing generated high-quality datasets from all 20 plasma exosome samples. After quality filtering and adapter trimming, 171.1 million clean reads were retained for downstream analysis, with high sequencing quality metrics. A total of 1,097 miRNAs with sufficient expression (mean counts ≥ 10 across all samples) were retained for downstream differential expression analysis (see Supplementary Materials for full detection statistics).

Identification of candidate dysregulated miRNAs

To explore global expression patterns and sample heterogeneity, PCA was first performed on the miRNA expression profiles. The PCA plot showed partial visual separation between the low muscle mass group and the normal muscle mass group, with the first two principal components accounting for 30.3% and 10.8% of the total variance, respectively (Fig. 2A). This visual pattern was considered descriptive and was not interpreted as evidence of statistically significant group separation.

Fig. 2.

Exploratory differential expression analysis of exosomal miRNAs in individuals with low muscle mass and normal muscle mass. (A) Principal component analysis (PCA) plot of exosomal miRNA expression profiles. (B) Volcano plot displaying exploratory candidate miRNAs between the low muscle mass and normal muscle mass groups based on |log2 fold change| ≥ 0.58 and nominal P < 0.05. (C) Heatmap visualization of the exploratory candidate miRNAs across all samples.

After Benjamini–Hochberg correction, only hsa-miR-3942-5p remained statistically significant (FDR-adjusted P = 0.0491). Using the exploratory thresholds of |log2 fold change| ≥ 0.58 and nominal P < 0.05, differential expression analysis retained 101 miRNAs as exploratory candidates, including 61 upregulated and 40 downregulated miRNAs in individuals with low muscle mass (Fig. 2B, C). Among these exploratory candidates, the largest positive fold changes were observed for hsa-miR-122-3p (log2 FC = 5.12, nominal p = 0.01), hsa-miR-802 (log2 FC = 4.75, nominal p = 0.03), and hsa-miR-216a-5p (log2 FC = 4.19, nominal p = 6.4 × 10 − 3). The largest negative fold changes were observed for hsa-miR-519a-5p (log2 FC = −3.31, nominal p = 2.9 × 10 − 3), hsa-miR-515-5p (log2 FC = −3.27, nominal p = 1.9 × 10 − 3), and hsa-miR-6780a-5p (log2 FC = −2.85, nominal p = 4.5 × 10 − 4).

Public dataset analysis and identification of exploratory candidate genes

PCA of the GSE111016 dataset demonstrated an apparent separation between sarcopenia patients and normal controls (Fig. 3A); this pattern was considered descriptive. After Benjamini–Hochberg correction, only NOX1 remained statistically significant (FDR-adjusted P = 0.0152). Using limma for differential expression analysis (Fig. 3B and C), a total of 156 exploratory candidate DEGs were identified from the 40 muscle tissue samples (20 sarcopenia vs. 20 normal controls). These included 73 upregulated genes and 83 downregulated genes in participants with sarcopenia. These genes were identified using the exploratory thresholds of |log2 fold change| ≥ 0.58 and nominal P < 0.05 and were used only as exploratory candidates in subsequent integrative analyses.

Fig. 3.

Exploratory differential gene expression analysis of sarcopenia muscle tissue transcriptome. (A) Principal component analysis of the GSE111016 dataset showing sample distribution. Participants with sarcopenia are represented by triangles and normal controls by circles. (B) Volcano plot showing exploratory candidate DEGs based on |log2 fold change| ≥ 0.58 and nominal P < 0.05. (C) Heatmap showing expression patterns of the exploratory candidate genes across muscle tissue samples.

miRNA-mRNA regulatory network construction

Target prediction analysis using miRanda and RNAhybrid algorithms identified potential mRNA targets for the candidate dysregulated miRNAs. For upregulated miRNAs, miRanda predicted 11,566 targets while RNAhybrid predicted 10,730 targets, with 6,181 overlapping predictions. When these overlapping predictions were intersected with the downregulated DEGs from GSE111016, 25 candidate downregulated target genes were retained (Fig. 4A). Similarly, for downregulated miRNAs, miRanda predicted 12,108 targets and RNAhybrid predicted 11,172 targets, with 6,936 overlapping predictions. Intersection of these predictions with the upregulated DEGs yielded 25 candidate upregulated target genes (Fig. 4B). The retained candidate target genes were subsequently linked to their predicted miRNAs to construct two direction-specific putative miRNA–mRNA interaction networks. The upregulated miRNA regulatory network (Fig. 4C) consisted of 51 nodes (26 miRNAs and 25 target genes) and 63 edges (regulatory interactions), with hub miRNAs including hsa-miR-6764-5p, hsa-miR-6842-3p, and hsa-miR-3199 showing the highest connectivity degrees. The downregulated miRNA regulatory network (Fig. 4D) comprised 42 nodes (17 miRNAs and 25 target genes) and 71 edges, with hsa-miR-3150a-3p, hsa-miR-6780a-5p, and hsa-miR-378 g serving as major hub nodes with extensive target connections.

Fig. 4.

Construction of miRNA-mRNA regulatory networks. (A) Venn diagram showing the overlap among targets of upregulated miRNAs predicted by miRanda and RNAhybrid and candidate downregulated genes, identifying 25 candidate downregulated target genes. (B) Venn diagram showing the overlap among targets of downregulated miRNAs predicted by miRanda and RNAhybrid and candidate upregulated genes, identifying 25 candidate upregulated target genes. (C) Putative interaction network comprising upregulated miRNAs and predicted downregulated target genes. miRNAs are represented by red circles and target genes by green circles. (D) Putative interaction network comprising downregulated miRNAs and candidate upregulated target genes. Edges indicate computationally predicted miRNA–mRNA associations. Although each three-way Venn intersection contained 25 candidate target genes, the corresponding networks represented opposite expression-direction combinations and were therefore analysed as distinct, direction-specific networks.

Functional enrichment and pathway analysis

To elucidate the potential biological functions of the miRNA-mRNA regulatory networks, comprehensive functional enrichment analysis was performed on the putative target genes retained in the two networks. Gene Ontology enrichment analysis revealed distinct functional patterns between upregulated and downregulated miRNA targets across three ontological categories (Fig. 5A and D). For upregulated miRNA targets, biological process analysis showed nominal enrichment in cellular differentiation and development processes, particularly “keratinocyte differentiation”, “epidermal cell differentiation”, “skin development”, and “epidermis development” etc. Cellular component analysis demonstrated enrichment in membrane-related structures, including “intrinsic component of membrane”, “anchored component of membrane”, and “endoplasmic reticulum membrane” etc. Molecular function analysis highlighted enzymatic activities, with “oxidoreductase activity” and “vitamin D 24-hydroxylase activity” showing the strongest enrichment, along with “dATP phosphatase activity”, “dCTP phosphatase activity” etc. Conversely, downregulated miRNA targets demonstrated different patterns of nominal enrichment. Biological process analysis revealed nominal enrichment in cardiovascular function-related processes, including “regulation of heart contraction”, “heart contraction”, “blood circulation”, “circulatory system process” etc. Cellular component analysis showed enrichment in specialized membrane structures such as “CBM complex”, “voltage-gated sodium channel complex”, “photoreceptor disc membrane”, and “immunological synapse” etc. Molecular function analysis highlighted “molecular function regulator”, “receptor ligand activity”, “receptor regulator activity”, and various transporter activities etc.

Fig. 5.

Functional enrichment and pathway analysis of miRNA target genes. (A) Gene Ontology (GO) enrichment analysis results for targets of upregulated miRNAs, categorized by biological processes (BP), cellular components (CC), and molecular functions (MF). (B) KEGG pathway enrichment analysis for upregulated miRNA targets. (C) Pie chart visualization of KEGG pathway categories for upregulated miRNA targets. (D) GO enrichment analysis results for targets of downregulated miRNAs. (E) KEGG pathway enrichment analysis for downregulated miRNA targets. (F) Pie chart of KEGG pathway categories for downregulated miRNA targets. (G, H) Gene Set Enrichment Analysis (GSEA) plots showing enrichment patterns for various biological pathways. KEGG pathway information was obtained from the Kyoto Encyclopedia of Genes and Genomes (KEGG) database²² and used with permission from Kanehisa Laboratories.

KEGG pathway enrichment analysis identified several pathways potentially associated with sarcopenia-related molecular alterations (Fig. 5B, C and E, and 5F). Upregulated miRNA targets were enriched in metabolic pathways including “Sphingolipid metabolism”, “Sphingolipid signaling pathway”, “Metabolic pathways”, “Retinol metabolism”, and “Various types of N-glycan biosynthesis”, suggesting possible alterations in lipid and general metabolic processes. Downregulated miRNA targets showed enrichment in specialized pathways such as “2-Oxocarboxylic acid metabolism”, “Arginine biosynthesis”, “ABC transporters”, “Neuroactive ligand-receptor interaction”, and “Phototransduction”, suggesting potential alterations in metabolic and signaling processes.

GSEA revealed coordinated transcriptomic differences associated with sarcopenia status (Fig. 5G and H). Upregulated gene sets showed nominal enrichment in multiple metabolic stress pathways, including arachidonic acid metabolism, drug metabolism via cytochrome P450, sphingolipid metabolism, linoleic acid metabolism, retinol metabolism etc. (NES > 1.0, nominal p < 0.05). In contrast, downregulated gene sets showed negative enrichment of fundamental cellular processes, including “Oxidative phosphorylation”, “Ribosome”, “Cardiac muscle contraction”, “Aminoacyl-tRNA biosynthesis”, “Citrate cycle (TCA cycle)” etc. (NES < −1.0, nominal p < 0.05). Collectively, these exploratory pathway analyses suggested pathway-level differences involving lipid metabolism, energy-production-related processes, protein synthesis, and muscle contraction. These enrichment findings do not establish corresponding functional metabolic alterations.

Machine learning-based feature selection and biomarker identification

Three machine learning algorithms were employed to prioritize candidate miRNA features associated with low muscle mass rather than to establish validated biomarkers for identifying low muscle mass. Random Forest analysis demonstrated that hsa-miR-1294 exhibited the highest variable importance score, followed by multiple miRNAs with varying degrees of importance (Fig. 6A). XGBoost feature selection revealed hsa-let-7c-5p as the most important feature with the highest variable relative importance (approximately 1.2), followed by hsa-miR-1294 and hsa-miR-1306-5p (Fig. 6B). SVM analysis with recursive feature elimination showed that the highest apparent cross-validation accuracy was achieved with approximately 38 features, reaching a 10-fold cross-validation accuracy of 0.942 (Fig. 6C and D). The SVM feature selection process demonstrated that model performance plateaued after including the top 30–35 most informative miRNAs. Integrated analysis of the three machine learning algorithms was performed to identify overlapping candidate miRNA features (Fig. 6E). Their putative target relationships were subsequently examined to provide potential biological context, including hsa-miR-1294 predicted to target TRARG1, and hsa-let-7c-5p predicted to target SLC38A11, RNF207, FGF11, and ABCG4 (Fig. 6F).

Fig. 6.

Machine learning-based exploratory feature prioritization. (A) Random Forest feature importance ranking for miRNAs from the constructed regulatory networks. Bar chart shows variable importance scores for individual miRNAs. (B) XGBoost feature importance analysis displaying variable relative importance for miRNA features. (C) Support Vector Machine (SVM) cross-validation accuracy plot across different numbers of features. (D) SVM recursive feature elimination plot showing model performance optimization process. (E) Venn diagram comparing feature rankings from three machine learning algorithms to identify consensus features. (F) Sankey diagram illustrating miRNA-target gene regulatory relationships for top-ranked features.

Nomogram-based exploratory diagnostic model development and internal evaluation

Two nomogram models were constructed as exploratory tools in separate development datasets. The first nomogram incorporated the top-ranked miRNAs, including hsa-miR-1294 and hsa-let-7c-5p, identified through the integrated machine learning approach (Fig. 7A). The second nomogram was constructed using target genes of the key miRNAs, incorporating SLC38A11, ABCG4, FGF11, TRARG1, and RNF207 (Fig. 7D). Internal model evaluation showed preliminary discriminatory performance for both nomogram models. The miRNA-based nomogram achieved an area under the ROC curve (AUC) of 0.905 with a 95% CI of 0.765–1.000 (Fig. 7B), while the target gene-based nomogram showed comparable performance with an AUC of 0.905 with a 95% CI of 0.814–0.996 (Fig. 7E). Decision curve analysis suggested apparent net benefit within the respective development datasets across a wide range of threshold probabilities (Fig. 7C and F). The miRNA-based model showed apparent net benefit within the current cohort, while the target gene model maintained consistent net benefit across threshold probabilities from 0.1 to 0.7. These internally evaluated findings were considered exploratory and require external validation.

Fig. 7.

Development and internal evaluation of exploratory nomogram models. (A) Nomogram model based on selected candidate miRNAs for low muscle mass identification. The nomogram provides a scoring system for individual risk assessment. (B) Receiver operating characteristic (ROC) curve for the miRNA-based nomogram model. (C) Decision curve analysis showing net benefit of the miRNA-based model across different threshold probabilities. (D) Nomogram model incorporating target genes of key miRNAs. (E) ROC curve for the target gene-based nomogram model. (F) Decision curve analysis showing the apparent net benefit of the gene-based model within its development dataset.

Immune microenvironment analysis

xCell analysis of muscle-tissue transcriptomic data revealed differences in estimated cell-type enrichment scores between participants with sarcopenia and controls (Fig. 8A). Muscle satellite-cell enrichment scores were lower in participants with sarcopenia (nominal P < 0.05), whereas mast-cell enrichment scores were higher (nominal P < 0.05). Plasmacytoid dendritic-cell enrichment scores were also lower in participants with sarcopenia (nominal P < 0.05). Correlation analysis revealed associations between candidate target-gene expression and cell-type enrichment scores (Fig. 8B). RNF207 (r = 0.46, P < 0.001), FGF11 (r = 0.38, P < 0.001), and ABCG4 (r = 0.30, P < 0.001) were positively correlated with pDC enrichment scores. Target-gene expression was negatively correlated with mast-cell enrichment scores, with correlation coefficients ranging from r = − 0.23 to − 0.30 (P < 0.001). Muscle satellite-cell enrichment scores were positively correlated with ABCG4 expression (r = 0.21, P < 0.05). These findings indicated exploratory associations between candidate target-gene expression and estimated cell-type enrichment scores but did not establish changes in cell abundance or regulation of immune-cell recruitment and function.

Fig. 8.

Immune microenvironment analysis in sarcopenic muscle tissue. (A) Violin plots showing xCell-derived cell-type enrichment scores in participants with sarcopenia and normal controls. (B) Correlation heatmap between target gene expression levels and cell-type enrichment scores. Color scale represents correlation coefficients, with nominal P-value indicators shown for different correlation strengths.

Exploratory drug screening and molecular docking analysis

Connectivity Map analysis identified computationally predicted compounds potentially associated with the low-muscle-mass-associated gene expression signature. Among the identified compounds, deferasirox showed the most negative connectivity score (−0.5917), followed by acarbose (−0.5613), taxifolin (−0.5582), pterostilbene (−0.5522), and ercalcitriol (−0.5499). These compounds represent diverse therapeutic classes including iron chelators (deferasirox), antidiabetic agents (acarbose), flavonoids (taxifolin and pterostilbene), and vitamin D analogues (ercalcitriol), and should be regarded as exploratory candidates for future validation.

Molecular docking simulations were performed to assess potential binding interactions between the predicted compounds and key target proteins. AutoDock Vina analysis produced predicted docking scores ranging from − 6.0 to −9.5 kcal/mol (Fig. 9A). Deferasirox showed the most favourable predicted docking score with SLC38A11 (−9.495 kcal/mol), while acarbose showed optimal interaction with ABCG4 (−8.754 kcal/mol) and SLC38A11 (−8.633 kcal/mol). Detailed structural analysis revealed predicted binding poses for selected ligand–protein complexes, including TRARG1-deferasirox (Fig. 9B), ABCG4-acarbose (Fig. 9C), FGF11-deferasirox (Fig. 9D), SLC38A11-acarbose (Fig. 9E), and RNF207-acarbose (Fig. 9F). These docking results were used for computational prioritization and did not establish physical binding or therapeutic activity.

Fig. 9.

Molecular docking analysis of computationally prioritized compounds and key target proteins. (A) Heatmap showing predicted docking scores between computationally prioritized compounds (deferasirox, acarbose, taxifolin, pterostilbene, and ercalcitriol) and target proteins (TRARG1, ABCG4, FGF11, SLC38A11, and RNF207). (B) Three-dimensional structure of TRARG1-deferasirox complex showing protein backbone and ligand molecule with key binding site residues labeled. (C) ABCG4-acarbose complex displaying protein structure and acarbose molecule with critical amino acid residues at the binding interface. (D) FGF11-deferasirox complex illustrating protein structure and deferasirox molecule with important contact residues highlighted. (E) SLC38A11-acarbose complex showing protein structure and acarbose molecule with binding site residues indicated. (F) RNF207-acarbose complex presenting protein structure and acarbose molecule with key interacting amino acids labeled.

Discussion

This study presents a comprehensive investigation of exosomal miRNA profiles in RA patients with low muscle mass in Xizang, using high-throughput sequencing, bioinformatics analysis, and machine learning methodologies. We identified altered exosomal miRNA expression patterns, prioritized hsa-miR-1294 and hsa-let-7c-5p as exploratory candidate miRNA features, and explored potential molecular mechanisms associated with low muscle mass. Our findings suggest dysregulation of miRNA expression patterns in plasma-derived exosomes from Tibetan RA patients with low muscle mass, revealing 101 candidate dysregulated miRNAs identified using exploratory thresholds.

The identification of 101 candidate dysregulated miRNAs should be interpreted cautiously, as most candidates did not remain significant after false discovery rate correction. Therefore, these findings represent exploratory molecular signatures rather than definitive low-muscle-mass-associated miRNA profiles. Although the relevance of these specific miRNAs to low muscle mass remains unclear, previous studies have implicated miRNA dysregulation in muscle-wasting conditions. Previous research on cancer cachexia demonstrated alterations in circulating miRNA profiles, with one study reporting the downregulation of miR-122-5p in patients with cachexia and its association with weight loss²³. While our findings show upregulation of miR-122-3p in RA patients with low muscle mass in Xizang, this difference may reflect distinct pathophysiological mechanisms between cancer-associated cachexia and sarcopenia.

Machine learning-based feature selection consistently prioritized hsa-miR-1294 and hsa-let-7c-5p across multiple algorithms, although they were not necessarily the miRNAs with the largest fold changes. This observation suggests that feature prioritization may depend not only on differential-expression magnitude but also on the contribution of each miRNA to classification within the current dataset. Similarly, hub miRNAs identified from the miRNA-mRNA regulatory network, such as hsa-miR-3150a-3p, hsa-miR-6780a-5p, and hsa-miR-378 g, may reflect high network connectivity rather than diagnostic classification ability. Taken together, these findings prioritize hsa-miR-1294 and hsa-let-7c-5p as candidate miRNA features for further evaluation, while the highly dysregulated and hub miRNAs may provide complementary information regarding disease-associated expression changes and predicted network connectivity. Given the limited sample size and the absence of nested feature selection, these findings should be considered exploratory and require validation in independent cohorts.

The integration of plasma exosomal miRNA profiles with a public muscle transcriptomic dataset was intended as a hypothesis-generating strategy. Although the two datasets differ in sample origin, population background, and technical platform, cross-dataset integration provided an indirect approach for prioritizing candidate miRNA–mRNA relationships relevant to skeletal muscle. However, the tissue origin of the plasma-derived exosomes was not determined, and they therefore cannot be assumed to be muscle-derived. Because low muscle mass primarily reflects skeletal muscle alterations, cross-referencing predicted targets of circulating exosomal miRNAs with muscle tissue DEGs may help prioritize miRNA-mRNA pairs with greater relevance to muscle biology. However, this cross-dataset integration cannot replace validation in paired plasma and muscle samples or functional experiments.

Although these specific miRNAs have not been directly reported in sarcopenia literature, the let-7 family has been extensively studied in muscle biology and aging. Wang et al. summarized previous evidence showing elevated expression of several let-7 family members in aged human skeletal muscle and their potential involvement in pathways related to cellular proliferation and differentiation²⁴. Furthermore, research by Maki Itokazu et al. revealed that adipose-derived exosomes deliver let-7d-3p to muscular stem cells, blocking their proliferation by targeting HMGA2, demonstrating how let-7 family members can contribute to sarcopenia development through intercellular communication²⁵. Regarding hsa-miR-1294, a longitudinal case report involving one patient with hypophosphatasia observed increased circulating hsa-miR-1294 during asfotase alfa treatment, accompanied by improvements in functional measures²⁶. Interestingly, our analysis revealed downregulation of hsa-miR-1294 in Tibetan patients with RA and low muscle mass, which may suggest a different regulatory mechanism in age-related muscle loss compared to other musculoskeletal disorders. The same case report also observed increased hsa-let-7i-5p during treatment, providing only indirect contextual evidence for the involvement of let-7 family members in musculoskeletal conditions²⁶.

The development of nomogram-based diagnostic models represents an exploratory step toward the potential clinical translation of our findings. Both the miRNA-based and target gene-based nomograms achieved high apparent discriminatory performance with AUCs of 0.905, within the datasets used for model development. However, this result should be interpreted cautiously. The sample size was small, no independent training and testing cohorts were available, and no external validation dataset was included. Therefore, the observed AUC values may be affected by optimistic bias, overfitting, and potential information leakage. Decision curve analysis suggested potential net benefit across a range of threshold probabilities, but the practical clinical utility of these models remains unconfirmed and requires prospective external validation before clinical implementation.

The construction of putative miRNA-mRNA regulatory networks revealed predicted molecular interactions potentially associated with muscle-related phenotypes. The upregulated miRNA network demonstrated nominal enrichment in cellular differentiation and membrane-related processes, while the downregulated miRNA network showed nominal enrichment in cardiovascular function-related processes. This finding may be relevant to the established association between sarcopenia and cardiovascular dysfunction in elderly populations^27,28. The GSEA results provide exploratory pathway-level findings, including positive enrichment of several metabolic pathways and negative enrichment of energy-production-related pathways. Specifically, our analysis identified nominal enrichment of several pathways potentially relevant to muscle homeostasis. The positive enrichment of sphingolipid metabolism-related pathways was consistent with pathway-level differences involving lipid metabolism, which has been implicated in muscle insulin resistance and inflammatory responses^29,30. Sphingolipids, particularly ceramides, are known to interfere with insulin signaling through inhibition of Akt/PKB phosphorylation, leading to impaired protein synthesis and enhanced protein degradation in skeletal muscle^31,32. The negative enrichment of oxidative phosphorylation pathways observed in our study aligns with the well-documented mitochondrial dysfunction in sarcopenic muscle, affecting the electron transport chain and ATP synthesis, which results in reduced cellular energy availability for muscle contraction and protein synthesis^33,34. Concurrently, the downregulation of ribosomal pathways and aminoacyl-tRNA biosynthesis may indicate alterations in protein-synthesis-related processes but does not directly establish dysregulation of the mTOR pathway³³.

The enrichment of neuroactive ligand-receptor interaction pathways among downregulated targets may implicate pathways related to neuromuscular signaling, which may contribute to the loss of motor units and muscle denervation observed in aging muscle. Recent molecular investigations have provided compelling evidence for the mechanisms underlying neuromuscular junction dysfunction in aging. Balanyà-Segura et al., demonstrated significant molecular adaptations in BDNF/NT-4 neurotrophic and muscarinic pathways at aging neuromuscular synapses, revealing an imbalance in the stoichiometry of BDNF/NT-4, altered TrkB receptor ratios, and reduced M2-subtype muscarinic receptors³⁵. Furthermore, Zhao et al., identified a critical mechanism whereby aging-related increases in interleukin-6 expression inhibit acetylcholine receptor β-subunit (AChR-β) expression through the IL-6/IL-6R-ERK1/2-PGC1α/MEF2C pathway³⁶. This inflammatory-mediated downregulation of acetylcholine receptors provides a molecular explanation for the compromised neuromuscular transmission observed in sarcopenia, as reduced AChR expression directly impairs the muscle’s ability to respond to neural stimulation. These findings align with our pathway analysis showing suppression of neuroactive ligand-receptor interactions and suggest that the inflammatory milieu characteristic of sarcopenia may be consistent with alterations in pathways relevant to neuromuscular signaling^35,36. Additionally, the suppression of cardiac muscle contraction pathways not only reflects the cardiovascular complications associated with sarcopenia but also may reflect overlapping pathway signatures between skeletal- and cardiac-muscle-related processes. Recent comprehensive reviews have highlighted the bidirectional relationship between sarcopenia and cardiovascular diseases, revealing complex pathophysiological mechanisms including chronic low-grade inflammation, hormonal dysregulation, and compromised calcium signaling pathways that affect both cardiac and skeletal muscle function³⁷. In heart failure patients, proposed mechanisms linking heart failure to peripheral muscle wasting include abnormal mitochondrial function, dysregulated proteostatic signaling, and systemic inflammation, all of which contribute to progressive muscle loss and compromised excitation-contraction coupling³⁸. The suppression of cardiac muscle contraction pathways observed in our analysis aligns with clinical evidence demonstrating that sarcopenia and cardiovascular conditions create a vicious cycle.

The xCell analysis revealed differences in estimated cell-type enrichment scores between participants with sarcopenia and controls. Muscle satellite-cell enrichment scores were lower in participants with sarcopenia, which may be consistent with altered muscle-regenerative biology. However, xCell does not directly quantify cell abundance or function³⁹. Mast-cell enrichment scores were higher, whereas plasmacytoid dendritic-cell (pDC) enrichment scores were lower in participants with sarcopenia. Although mast cells and pDCs have recognized roles in immune processes^40–42, the biological implications of these enrichment-score differences remain uncertain. Correlation analysis revealed positive associations between candidate target-gene expression and pDC enrichment scores and negative associations with mast-cell enrichment scores. These findings indicated exploratory associations between candidate target-gene expression and estimated cell-type enrichment scores but did not establish changes in cell abundance or regulation of immune-cell recruitment and function.

The Connectivity Map and molecular docking analyses were performed as exploratory computational screens rather than experimental validation of therapeutic efficacy. Deferasirox, an iron chelator, showed the strongest negative connectivity score and was therefore prioritized as an exploratory candidate; however, this result does not establish that iron metabolism contributes to low muscle mass⁴³. Acarbose, an antidiabetic agent, was also computationally prioritized but cannot be considered to have demonstrated therapeutic potential⁴⁴. The identification of flavonoids (taxifolin and pterostilbene) and vitamin D analogues (ercalcitriol) provides additional candidates for future investigation but does not demonstrate therapeutic activity or confirm the involvement of their associated pathway^45,46. Nevertheless, these compounds should be described as computationally predicted candidate compounds, not as validated therapeutic agents. CMap analysis only suggests reversal of disease-associated expression signatures, and molecular docking only estimates potential ligand-protein binding. Neither approach confirms drug efficacy, target engagement, dose-response relationships, pharmacokinetic feasibility, or safety in RA patients with low muscle mass.

Several limitations should be acknowledged when interpreting our findings. The relatively small sample size may limit the generalizability of our results, and validation in larger, more diverse cohorts is essential. The sequencing cohort was single-centre, predominantly female, and imbalanced in group size, which further limits generalizability. The limited sample sizes may have reduced the statistical power to detect additional associations after FDR correction; therefore, findings based on nominal P values should be interpreted as exploratory. In particular, the exosomal miRNA cohort was not divided into independent training and testing sets, and no external validation cohort was included. Although SVM-RFE used 10-fold cross-validation for internal feature-selection assessment, feature selection and model performance evaluation were not fully separated across independent datasets. Therefore, the reported diagnostic performance may be affected by optimistic bias, overfitting, and potential information leakage. The case-control study design precludes assessment of temporal relationships and causal inferences regarding miRNA expression changes and disease progression. In addition, the primary phenotype in the sequencing cohort was low muscle mass rather than clinically confirmed sarcopenia, because complete handgrip strength and gait speed data were not available for all participants. The sequencing cohort and GSE111016 also differed in population background, tissue source, and phenotype definition; therefore, the cross-dataset integration should be considered hypothesis-generating rather than external validation. While our integrated approach combining multiple machine learning algorithms supports the prioritization of candidate miRNA features, experimental validation of the predicted miRNA-mRNA interactions through functional studies would provide additional mechanistic insights. Accordingly, hsa-miR-1294 and hsa-let-7c-5p should be regarded as candidate miRNA features that require further confirmation in larger, multicenter, independent cohorts before clinical translation. Nevertheless, the future identification and validation of minimally invasive blood-based biomarkers for low muscle mass identification could address an important clinical need by facilitating sample collection and serial monitoring. The exploratory nomogram models may provide clinically interpretable frameworks for future evaluation, but their practical implementation will require further external validation and prospective evaluation.

Conclusion

In conclusion, this exploratory study identified plasma-derived exosomal miRNA features associated with low muscle mass in patients with RA. The integration of high-throughput sequencing, bioinformatics analysis, and machine learning approaches prioritized candidate miRNAs and putative regulatory networks for further investigation. The preliminary discriminatory performance of our nomogram-based models, combined with exploratory computational drug-screening results, provides hypothesis-generating evidence for future biomarker and mechanistic studies. However, the diagnostic performance observed in this study should be interpreted as preliminary because independent training and testing datasets were not available. Further external validation in larger cohorts is required before these candidate features or nomogram models can be applied for clinical identification of low muscle mass or sarcopenia-related risk.

Author contributions

All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by Li Tang, Zhigang Lin, Linmeng Zou, Yongmei Hu, Shan Gao and Ling Wei. Wen Pan contributed to the acquisition and curation of part of the study data and participated in the review of the revised manuscript. The first draft of the manuscript was written by Zhigang Lin, and all authors commented on previous versions of the manuscript. All authors, including Wen Pan, read and approved the final manuscript and agreed to be accountable for their respective contributions.

Funding

This research was supported by Science and Technology Projects of Xizang Autonomous Region, China (No. XZ202301ZY0050G and XZ202501YD0013) and Chengdu Medical Research Program, China (No. 2022359).

Data availability

The Gene expression dataset (GSE111016) was downloaded from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/). Processed gene expression profiles of sarcopenia-associated DEGs were submitted to the CMap portal (https://clue.io/) for connectivity analysis. The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.

Code availability

The custom scripts used in this study, together with workflow documentation, software and package versions, key parameter settings, and instructions for reproducing the principal analyses, are publicly available in Zenodo (https://doi.org/10.5281/zenodo.21525716). The same materials are also provided as supplementary attachments. The deposited files contain the custom analysis scripts and documentation for the analytical steps performed using established third-party software. No new algorithm or standalone software package was developed in this study.

Materials availability

All materials are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

The study was conducted according to the guidelines of the Declaration of Helsinki, and approved by the Ethics Committee of Hospital of Chengdu Office of People’s Government of Xizang Autonomous Region (Ethical Approval number: (2023) Scientific Research No. 43). Informed consent was obtained from all subjects involved in the study.

Consent for publication

All participants provided informed consent for publication of study results.