Analysis of exosomal mRNA profile in cancer tissue from patients with colorectal cancer

Abstract

Colorectal cancer (CRC) is one of the most common malignant tumors. Exosomes are key modulators of intercellular communication and important participants in CRC tumorigenesis. However, the influence and underlying mechanism of cancer-secreted exosomes in CRC are not fully understood. In the present study, exosome-enriched extracellular vesicles (EVs) were isolated from paired CRC and adjacent normal control tissue (NC) from patients with CRC using ultracentrifugation, size exclusion chromatography and ultrafiltration. Bowtie software was used to identify mRNA expression patterns. Further analyses included functional enrichment analysis using Gene Ontology and Kyoto Encyclopedia of Genes and Genomes for differentially expressed genes (DEGs) and Gene Set Enrichment Analysis for functional enrichment of the entire transcriptome. Gene Set Variation Analysis was used to identify differentially enriched functions and pathways for each sample. Weighted gene correlation network analysis was used to select candidate key genes. The diagnostic performance of the candidate key genes was validated using the models of multiple logistic regression and a support vector machine with the external CRC exosomal RNA database exoRBase. A total of 814 DEGs were identified in the exosome-enriched EVs from patients with CRC compared with those from NC group. The enriched functions were potentially associated with key steps and pathways involved in CRC tumorigenesis, including cell cycle, DNA replication, RNA activity, protein metabolism and the PI3K/Akt', Wnt, MAPK and KRAS pathways, as well as the regulation of mitochondria, the pentose phosphate pathway and intestinal immunity. A total 14 candidate key genes, namely, p53 effector related to PMP-22, MYC, dyskerin pseudouridine synthase 1), VDAC1 (voltage dependent anion channel 1), adenosylhomocysteinase), TMEM123 (transmembrane protein 123), RPL36A (ribosomal protein L36a), CCT5 (chaperonin containing TCP1 subunit 5), CHCHD2 (coiled-coil-helix-coiled-coil-helix domain containing 2), RPS21 (ribosomal protein S21), KRT18 (keratin 18), TGFBI (transforming growth factor beta induced), ETS2 (ETS proto-oncogene 2) and AXIN2 (axin 2), were identified. The present exploratory study revealed pronounced transcriptomic changes in exosome-enriched EV derived from CRC compared with those derived from paracancerous tissue, and may provide a novel platform for understanding tumor microenvironment interactions.

Introduction

Colorectal cancer (CRC) is the third most common diagnosed cancer and the most common cancer of the gastrointestinal tract. It is also the second leading causes of cancer-related deaths worldwide, with ~1.93 million new cases and 0.90 million deaths estimated in 2022, accounting for 9.6 and 9.3% of all cancer cases and deaths, respectively (1). In China from 2000 to 2019, the incidence of CRC has been increasing approximately doubled among male patients, rising at an average of 2.8% per year, and the annual increase among women was smaller at 1.4% (2,3). CRC is heterogeneous, and its carcinogenesis may be associated with genetic and epigenomic alterations and environmental and lifestyle factors (4). Early diagnosis of CRC is key for the survival of patients with CRC. The overall 5-year survival rate of patients with CRC is 65%; however, this rate increases to 94% if diagnoses are made in early (stage I and stage II) and localized stages (5). By contrast, the 5-year survival rate for patients with CRC with distant metastases is only 12-13% (6,7). Detection methods for the early diagnosis of CRC are needed to improve therapeutic outcomes. Although colonoscopy is an important and widely used for the detection of CRC, its invasion nature limits its use in certain patients and its accuracy depends on the skill level and experience of the endoscopist; False-negative colonoscopy findings have been reported in certain patients (8). More accurate and less invasive screening methods are therefore needed for CRC detection. Novel biomarkers are considered promising diagnostic tools for the detection of cancer (9).

Extracellular vesicles (EVs) are key mediators of intercellular communication. EVs are cell-derived, membrane-bound vesicles that transport bioactive molecules between cells. Based on their origin, size, release pathways, content and function, EVs are generally classified into three primary subtypes: Microvesicles, apoptotic bodies and exosomes (10). Studies have shown that exosomes are involved in the pathogenesis of CRC, including the proliferation, progression and chemoresistance of CRC cells (11,12). Exosomes are small lipid bilayer membrane vesicles, ranging in size from 30 to 150 nm and are secreted by almost all cell types (10). They are present in internal cell spaces and various body fluids, including plasma, urine, cerebral spinal fluid, breast milk, tears, saliva and gastric acid (13). Exosomes are key components of cell signaling and intercellular communication within the tumor microenvironment (TME) (14). They contain bioactive molecules, such as DNA, RNA, protein and lipids, but do not contain cell organelles (15). Previous studies have focused mainly on exosomes derived from body fluid samples or cell lines (16,17). Interest in tissue-derived exosomes has increased, as they can be isolated directly from tumor tissue sources and may better reflect the bioactivity of tumor cells within the TME (18,19).

The present study followed the terminology recommended by the International Society for Extracellular Vesicles, which states that the subcellular origin of exosomes should be demonstrated (20). Therefore, the EVs isolated in this study were described as ‘exosome-enriched EVs’. The term ‘exosome’ was used only for general description, as it remains widely used in even when the subcellular origin has not been confirmed (21,22). In this exploratory study, the mRNA profiles of exosome-enriched EV were analyzed. The candidate key genes were validated by the models of multiple logistic regression (LR) and a support vector machine (SVM) with the external CRC exosomal RNA database exoRBase (exorbase.org/). These findings may provide a novel platform for understanding TME interactions.

Materials and methods

Patients and tissue samples

From June 2023 to January 2024, the paired CRC and adjacent normal tissue (NC; ≥5 cm away from the tumor margin) were collected from nine patients (age range was 46-80 years old, with 6 males and 3 females) with CRC who received surgical treatment at the Beijing Rehabilitation Hospital of Capital Medical University (Beijing, China). The inclusion criteria were as follows: Aged >18 years, histopathological diagnosis of CRC (adenocarcinoma), no past medical history of other cancer, no implemented chemotherapy, no immunosuppressive treatment, no preoperative radiotherapy and no other associated somatic disease requiring chronic therapy before surgery. All participants in this study provided written informed consent. The present study was approved by the ethics committee of Beijing Rehabilitation Hospital of Capital Medical University (approval no. 2021bkky-149).

Isolation of exosome-enriched EVs from tissue

Following surgical resection, the tissue samples were immediately stored at -80˚C until use for exosome-enriched EV analysis. Exosome-enriched EVs were isolated from tissue as described by Vella et al (23), with slight modification. The dissociation mixture was prepared using a Miltenyi Human Tumor Dissociation kit (Miltenyi Biotec, cat. no. 130-095-929). Enzymes R, H and A were resuspended according to the manufacturer's instructions. The dissociation mixture containing 2.2 ml RPMI-1640 culture medium (Corning, Inc.), 50.0 µl enzyme, R, 100.0 µl enzyme H and 12.5 µl enzyme A was prepared immediately before use.

A small piece of tissue (~200 mg) was weighed, sectioned (300 µm thickness) on dry ice and incubated in the dissociation mixture for 10-15 min at 37˚C. The dissociated tissue was gently filtered twice through a 70 µm filter to remove residual tissue. The suspension was centrifuged at 300 x g for 10 min at 4˚C, and the supernatant was transferred to a fresh tube and centrifuged at 2,000 x g for 10 min at 4˚C. The cell-free supernatant was centrifuged at 10,000 x g for 20 min at 4˚C, then filtered through a 0.22 µm filter to remove cell debris.

The suspension was then processed by ultracentrifugation (UC) at 150,000 x g for 2 h at 4˚C. Size exclusion chromatography (SEC) was used for separation. The pellet was resuspended in 1 ml PBS and further purified via Exosupur^® columns (13.14x0.81 cm; cat. no. ES9P14e, Echobiotech). The run duration was 30 min at room temperature. Fractions 1-5 were collected and concentrated to 200 µl with 100 kDa molecular weight cut-off Amicon^® Ultra spin filters (Merck KGaA). No EV-depleted reagents were used during the isolation of exosome-enriched EVs from tissue.

Measurement of the size distribution and particle concentration of exosome-enriched EVs

The size distribution and particle concentration of the exosome-enriched EVs were measured using a nanoflow cytometer (N30E Nanoflow Analyzer, NanoFCM Inc.). The side scatter intensity (SSI) was measured by loading standard polystyrene nanoparticles (250 nm) into the nanoflow cytometer. The isolated exosome-enriched EV sample was diluted with PBS to 1-10 ng/µl and loaded into the nanoflow cytometer to measure the SSI. Concentration was calculated as the ratio of the SSI concentration to the particle concentration of the standard polystyrene nanoparticles.

For size measurement of the EVs, mixed-size standard silica nanoparticles (68,91,113 and 155 nm) were loaded into the nanoflow cytometer to generate a standard curve, followed by loading of the exosome-enriched EV sample. The size distribution was calculated according to the standard curve.

Transmission electron microscopy (TEM)

TEM (H-7650, Hitachi Ltd.) was used to observe exosome-enriched EVs. A total of 10 µl exosome-enriched EV mixture was placed on a copper mesh (BZ110223a, Zhongjingkeyi Technology Inc.) and incubated at room temperature for 1 min. Following washing with sterile distilled water, the exosome-enriched EVs were stained with a uranyl acetate (cat. no. TX89992, Yi Xin Biology) solution for 1 min. The sample was dried for 2 min under incandescent light. The copper mesh was imaged via TEM.

Exosome-enriched EV protein characterization via western blot analysis

Total protein was extracted from the supernatant fraction and tissue samples using RIPA buffer (cat. no. CW2333S, Jiangsu CoWin Biotech Co., Ltd.) containing protease inhibitor P6730 (Beijing Solarbio Science & Technology Co., Ltd.). The supernatant was obtained on ice for 20 min and then centrifuged at 16,100 x g for 10 min at 4˚C. The protein concentration was measured using the BCA method. Based on particle concentration and total protein concentration, the particle-to-protein ratio of exosome-enriched EV sample was calculated as 2.04x10⁸ particles/µg. A total of 30 µg protein from isolated exosome-enriched EV was denatured, resolved on 8-10% SDS-PAGE gels and transferred onto PVDF membranes. The same amount of protein from CRC tissue lysate was loaded as negative control for endoplasmic reticulum marker calnexin for contamination assessment. The PVDF membranes were blocked with 5% non-fat milk for 60 min at room temperature and then incubated with primary antibodies at 4˚C overnight as follows: ALG-2-interacting protein X (Alix; cat. no. sc-53540; Santa Cruz Biotechnology, Inc.), CD9 (cat. no. 60232-1-Ig; ProteinTech Group, Inc.), HSP70 (cat. no. ab181606; all 1:1,000; Abcam) and calnexin (cat. no. 10427-2-AP; 1:500; ProteinTech Group, Inc.). The HRP-conjugated goat anti-rabbit IgG (H+L; cat. no. SA00001-1; 1:5,000; ProteinTech Group, Inc.) was used as the secondary antibody after hybridization for 1 h at room temperature. The protein band was visualized after incubating with ECL reagent (cat. no. PE0010, Beijing Solarbio Science & Technology Co., Ltd.) and the image was captured using Tianneng Chemiluminescence Gel Imaging System (Tanon-5200 Multi, Shanghai Tanon Life Science Co., Ltd.).

RNA isolation and quality analysis

Total RNA was extracted from tissue exosome-enriched EVs using Exosome RNA Purification kit (cat. no. 5202050, Simgen Biotech) according to the manufacturer's protocol. Nucleotide length was paired-end 150 bp. NovaSeq 6000 SP Reagent kit (150 cycles; cat. no. 20028312; Illumina Inc.) was used for the sequencing. Final library loading concentration was 10-15 nmol/l. Raw bcl files were converted to FASTQ format using bcl2fastq software (v2.20, Illumina Inc., USA), TopHat2/Bowtie2 (https://tophat.com/ and https://bowtie-bio.sourceforge.net/bowtie2/index.shtml) for genome alignment, StringTie (https://ccb.jhu.edu/software/stringtie/index.shtml) for gene expression quantification (FPKM) and R (v4.1.3, https://www.r-project.org/), prcomp function for PCA. The RNA concentration and purity were evaluated using the Qubit microRNA Assay kit, a Qubit fluorometer (Thermo Fisher Scientific, Inc.; cat. no. Q32851) and an RNA Nano 6000 Assay kit with an Agilent Bioanalyzer 2100 System (Agilent Technologies, Inc.) (cat. no. 5067-1511).

Library preparation and sequencing

For the mRNA libraries, a total of 4.5 ng RNA/sample was used as input material for the RNA sample preparations. The sequencing libraries were generated using a SMARTer stranded total RNA-Seq kit V2 (Takara Bio, Inc.) (cat. no. 634413) according to the manufacturer's instructions. Finally, library quality was assessed via the Agilent Bioanalyzer 2100 and quantitative PCR using SYBR Green I, KAPA Library Quantification kit (cat. no. 07960140001, Roche Diagnostics). Primer sequences were as follows: (5'-3'): Forward primer (P5): AATGATACGGCGACCACCGA Reverse primer (P7): CAAGCAGAAGACGGCATACG. Thermocycling conditions: Initial denaturation at 95˚C for 3 min; followed by 40 cycles of 95˚C for 10 sec, 60˚C for 30 sec; quantification method: Library molar concentration was determined using the standard curve method generated from serially diluted reference DNA. Clustering of the index-coded samples was performed on an acBot Cluster Generation System using TruSeq PE Cluster kitv3-cBot-HS (Illumina, Inc.; cat. no. PE-401-3001) according to the manufacturer's instructions. Following cluster generation, the library preparations were sequenced on an Illumina NovaSeq 6000 platform, and paired-end reads were generated by EchoBiotech Co. Ltd.

Data preprocessing and quantification of mRNAs

Raw data reads in fastq format were processed through in-house Perl scripts. After removing reads containing adapters, reads containing poly-N sequences and low-quality reads from the raw data, the clean reads were obtained. The percentage of bases with Phred quality score ≥20, (Q20), ≥30 (Q30), GC content and sequence duplication level of the clean data were calculated. All downstream analyses were based on high-quality, clean data. Paired-end clean reads were aligned to the reference genome GRCh38 (ncbi.nlm.nih.gov/projects/genome/assembly/grc/human/) using TopHat2/Bowtie2 (tophat.com/; bowtie-bio.sourceforge.net/bowtie2/index.shtml). Mapped reads were used for the quantification of gene expression and differential expression analysis.

StringTie (v2.2.1; ccb.jhu.edu/software/stringtie/index.shtml) was used to estimate gene expression levels as fragments/kilobase of exon/million mapped fragments (FPKM) for each sample. The FPKM value was calculated based on the length of the fragments and the number of reads mapped to each fragment. FPKM values were used for expression quantification and visualization. Principal component analysis (PCA) of RNA sequencing data was performed using the ‘prcomp’ function in R (R v4.1.3; r-project.org).

Pathway enrichment analysis

Gene Ontology (GO; geneontology.org/) (24) enrichment analysis of the target genes of the differentially expressed mRNAs was performed using the topGO R package (v2.52.0; bioconductor.org/packages/topGO/) (25). Kyoto Encyclopedia of Genes and Genomes (KEGG) (26) integrates genomic, chemical, and pathway information to facilitate the functional interpretation of large-scale molecular data, particularly through enrichment analysis of biological pathways; (genome.jp/kegg/). clusterProfiler R package (v4.6.1; bioconductor.org/packages//release/bioc/html/clusterProfiler.html was used to test the statistical enrichment of differentially expressed genes (DEGs) in KEGG pathways.

Gene expression profiles and phenotype groupings in the CRC and NC groups were analyzed using Gene Set Enrichment Analysis (GSEA). GSEA was performed using the msigdbr' R package (version 7.5.1; gsea-msigdb.org/gsea/msigdb/) and Molecular Signatures Database (MSigDB, v7.5.1; gsea-msigdb.org/gsea/msigdb/) (27), available via the GSEA website (software.broadinstitute.org/gsea/index.jsp). MSigDB 7.5.1) was used to identify potential enrichment items. P<0.001 and false-discovery rate (FDR) <0.05 were considered statistically significant. The Reactome database (reactome.org/) was used to analyze the gene set enrichment.

Gene set variation analysis (GSVA) of exosomal DEGs. GSVA is a non-parametric and unsupervised method to evaluate the enrichment of transcriptome gene sets. By scoring the gene sets, GSVA transforms gene-level alterations into pathway-level changes and cell types to judge the biological function of the sample. GSVA scores were calculated between the CRC and NC groups using the R package limma (v3.50.3; bioconductor.org/packages//release/bioc/html/limma.html) to identify differentially enriched functions and pathways. GSVA was used to calculate a signaling pathway variation score for each sample using the GSVA R package (v1.42.0) (https://bioconductor.org/packages//release/bioc/html/GSVA.html) (28), with P<0.05 considered to indicate a statistically significant difference.

Weighted gene correlation network analysis (WGCNA)

Exosome mRNA co-expression networks were constructed using the WGCNA package in R (V1.72.1) (29). After the expression data were filtered with transcript per million >1 and coefficient of variation >0.5, the remaining mRNAs were used for WGCNA. A soft threshold was chosen to build a network. The adjacency matrix was set to a continuous value from 0-1, thus, the network conformed to a power-law distribution and more closely resembled an actual biological network. The block modules were used to build a scale-free network, and module partition analysis was to identify co-expression modules, which were used to group genes with similar expression patterns. The dynamic tree cutting algorithm was used to determine the module by dividing the cluster tree into branches and assigning colors. The module eigengenes (ME) was used to summarize the gene expression associated with all modules.

To identify clinically meaningful modules, the correlation between modules and clinical trait was analyzed. For analysis, two methods were used to identify key modules in the network. In the first method, the Pearson correlation coefficient between the ME of each module and clinical feature was calculated, thereby identifying modules that were significantly correlated with the features (P<0.05). In the second method, the Pearson correlation coefficient [gene significance (GS)] between the expression of each gene and each clinical trait was calculated; the average absolute value of GS for all genes within the module was calculated. The larger the absolute value, the stronger the correlation between the module and clinical trait.

Screening and validation of key genes

Candidate key genes were selected from the known mRNAs according to the following criteria: An absolute value of log₂ FC value >1, an average FPKM value >10, an area under the curve (AUC) value >0.8 and expression levels >0 in 13 samples. The direction of regulation was consistent with The Cancer Genome Atlas (TCGA) RNA expression analysis data for rectum adenocarcinoma (READ) colon adenocarcinoma and the genes belonged to the dark olive green and cyan WGCNA modules. Candidate key genes were analyzed using the Wilcoxon test with data from TCGA-READ (cancer.gov/tcga) (30). The overlapping genes were subsequently subjected to protein-protein interaction (PPI) analysis STRING (v12.0) (https://string-db.org/).

Following univariate logistic regression analysis, candidate key genes with statistical significance (P≤0.05) were selected. Their diagnostic performance was evaluated using an independent CRC cohort from the exoRBase database (Version 2.0; exorbase.org:8080/exoRBaseV2/download/toIndex (31), which contains exosomal RNA profiles from patients with CRC and healthy controls.

The exoRBase dataset was randomly divided into a training set (70%, CRC, n=25, healthy controls, n=83, total, n=108) and an independent testing set (30%, CRC, n=11, healthy controls, n=36, total, n=47). Candidate molecular panels comprising 3-14 genes were systematically evaluated using the training set alone. Feature selection was embedded within a nested three-fold cross-validation workflow: Feature screening, panel performance ranking and hyperparameter tuning were independently conducted within each fold of the training cohort. The independent testing set was excluded from all feature selection and model optimization procedures and reserved for final external validation to completely avoid data leakage.

For each candidate panel, both SVM and LR models were independently constructed. LR models were developed using the lrm function from the R rms package (v6.8.0) (hbiostat.org/r/rms/), with maximum likelihood estimation adopted for model fitting to generate predicted probabilities for diagnostic evaluation. SVM models were implemented using LIBSVM (v3.23; csie.ntu.edu.tw/~cjlin/libsvm/) (32) using a C-support vector classification (C-SVC) classifier, employing a radial basis function (RBF) kernel. In this formulation, C is the penalty parameter that balances margin maximization against training error minimization, and γ is the RBF kernel coefficient that controls the kernel's width. Both hyperparameters were optimized within a nested cross-validation framework. Platt sigmoid scaling was applied to obtain probabilistic outputs from the SVM models to align with the LR probability prediction format.

For each candidate panel, the average performance across the three cross-validation folds was calculated, including the AUC, precision, recall, F1 score, which is a standard statistical metric calculated as the harmonic mean of precision and recall, and Matthews correlation coefficient. Candidate panels were ranked primarily according to the mean AUC, with the remaining metrics serving as complementary criteria to select the optimal molecular signature for each algorithm. The 95% confidence intervals (CIs) of the AUC values for both models were computed using the DeLong method (33).

The final optimized SVM and LR models were validated using the independent testing set. Receiver operating characteristic (ROC) curves and confusion matrices were generated to assess overall predicting performance. Bootstrap-corrected calibration curves of the independent testing set (based on 100 bootstrap replicates) were plotted to evaluate the calibration degree of both models. For the LR model, a diagnostic nomogram was constructed and Shapley Additive Explanations (SHAP) analysis was performed to quantify the contribution weight and importance ranking of each molecule in the final prediction model.

Statistical analysis

All data are presented as the mean ± SD. Data analysis was performed using R (version v4.1.3) (34) and differential expression between CRC and NC was analyzed using a paired design in limma package (v3.50.3; bioconductor.org/packages/release/bioc/html/limma.html). Raw read counts were normalized using the Trimmed Mean of M-values method, followed by variance modeling with the voom function in the limma package. A paired design matrix was incorporated into the linear model to account for matched sample pairs. Differentially expressed genes were identified using thresholds of P≤0.05 and |log₂ fold change|≥0.58. The Benjamini-Hochberg procedure was applied to calculate FDR-adjusted P-values for all genes, but it was not used for screening differentially expressed molecules. P<0.05 was considered to indicate a statistically significant difference. A Wilcoxon test was performed for PPI network analysis and to analyzed the expression levels of all genes in patients with CRC were significantly different from those in healthy individuals in TCGA-READ dataset.

Results

Characterization of exosome-enriched EV

Exosome-enriched EV were isolated from tissue of patients with CRC (Table I) using UC, SEC and ultrafiltration. Firstly, Nanoflow cytometry showed that most particles from both the CRC and NC groups were within the expected size range of exosome-enriched EV sizes, with the highest peak observed at 70 nm (Fig. 1A). TEM was used to assess the morphology and size of exosome-enriched EV from both CRC and NC tissue. The exosome-enriched EV showed a typical cup-shaped morphology, with lipid membrane-bound particles ranging from 50-150 nm in size (Fig. 1B). Western blotting showed the expression of commonly used exosomal markers, including CD9, HSP70 and Alix, in the extracted exosome-enriched EV, whereas the expression of the negative exosome marker calnexin was detected only in the supernatant fraction (Figs. 1C and S1A-D). Together, these morphological features, biomarker expression and size distribution indicated the successful extraction of exosome-enriched EVs, consistent with previous reports (10,35).

Table I.

Characteristics of patients with colorectal cancer.

Patient no.	Sex	Age, years	Tumor location	Differentiation grade	Stage
1	Female	46	Rectum	Middle-low	II
2	Male	54	Descending colon	Middle-low	III
3	Male	69	Transverse colon	Middle-low	II
4	Female	61	Ascending colon	High	II
5	Male	71	Ascending colon	Middle-low	III
6	Male	80	Sigmoid colon	Middle-low	III
7	Female	60	Ascending colon	High	III
8	Male	49	Sigmoid colon	Middle-low	III
9	Male	66	Ascending colon	Middle-low	II

Figure 1.

Characteristics and identification of exosome-enriched EV in CRC and NC tissue. (A) Nanoflow cytometry analysis of nanoparticle size distribution of the tissue exosome-enriched EV. (B) Representative exosome-enriched EV TEM. Scale bar, 0.5 µm. (C) Western blotting was used to detect the expression of Alix, CD9, HSP70 and calnexin. CRC tissue lysate was included as controls. EV, extracellular vesicle; CRC, colorectal cancer; NC, normal control; TEM, transmission electron microscopy; Alix, apoptosis-linked gene 2-interacting protein X; HSP70, heat shock protein 70 kDa.

Identification of exosomal mRNA profiles in CRC and NC tissue

Bowtie software was used to identify the mRNA expression patterns in the exosome-enriched EV from the CRC compared with NC tissue. A total of 9,426 mRNAs were detected in the two groups. Using the limma R package, 814 DEGs were identified, including 479 up- and 335 downregulated mRNAs in CRC (Fig. 2A). The heatmap showed mRNA DEGs from individual pairwise comparisons between CRC and NC, all of which were statistically significant (Fig. 2B). PCA of the DEGs was performed to assess differences between the paired CRC and NC samples. A separation tendency indicated differences in mRNA levels between the CRC and NC groups (Fig. 2C). The P-values, corrected FDRs and fold-changes for all genes were summarized in Table SI.

Figure 2.

Exosomal mRNA profile in CRC and NC tissue (n=9/group). (A) Volcano plots of the differentially expressed mRNAs identified using limma P package. (B) Heatmap indicating DEGs from individual pairwise comparisons between CRC and NC (variables important in projection ≥1 and absolute Log2 FC≥1, P<0.05). The scale indicates the z-score. Blue, low expression; red, high expression. (C) Principal component analysis of DEGs from exosome-enriched EV showed a separation tendency between CRC and NC. CRC, colorectal cancer; NC, normal control; DEG, differentially expressed gene; EV, extracellular vesicle; FC, fold change; Dim, dimension.

Functional enrichment of DEGs

GO and KEGG analyses were performed on these DEGs. The most significantly enriched biological processes (BPs) were associated with ‘cell division’, ‘protein stabilization’, ‘rRNA processing’ and ‘translational initiation’. At the cellular component (CC) level, the significant terms included ‘membrane’, ‘nucleolus’, ‘cytosol’, ‘extracellular exosome’, ‘nucleoplasm’ and ‘focal adhesion’. For molecular function (MF), ‘RNA binding’, ‘structural constituent of ribosome’, ‘collagen binding’, ‘protein kinase binding’ and ‘unfolded protein binding’ were among the significantly enriched items (Fig. 3).

Figure 3.

Gene Ontology functional enrichment analyses of differentially expressed genes.

KEGG pathway analysis showed that the most commonly enriched pathways were associated with ‘cell cycle’, ‘focal adhesion’, ‘ribosome’, ‘RNA transport’, ‘transcriptional misregulation in cancer’, ‘oxidative phosphorylation’, ‘PI3K-Akt signaling pathway’, ‘Wnt signaling pathway’, ‘Hippo signaling pathway’, ‘Ras signaling pathway’, ‘MAPK signaling pathway’ and ‘chemokine signaling pathway’. Other enriched pathways included ‘carbon metabolism’, ‘purine metabolism’, ‘biosynthesis of amino acids’ and ‘protein digestion and absorption’ (Fig. 4). These pathways are key for the development and progression of CRC (36,37).

Figure 4.

Kyoto Encyclopedia of Genes and Genomes functional enrichment analyses of differentially expressed genes.

GSEA was performed on the entire transcriptome to validate the functional enrichment results. The GSEA results were similar to those of the GO and KEGG enrichment analyses of the DEGs. The enriched GO terms included ‘chromosome segregation’ and mitotic nuclear division for BP, ‘chromosomal region’, ‘condensed chromosome’ and ‘protein/DNA complex’ for CC and ‘catalytic activity acting on a nucleic acid’, ‘catalytic activity acting on a tRNA’ and ‘tRNA methyltransferase activity’ for MF (Fig. S2A-C). The enriched KEGG terms included ‘cell cycle’ and ‘DNA replication’. In the Reactome, the enriched genes were associated with ‘DNA replication’ and cell cycle mitotic (Fig. S3A and B).

GSVA of exosomal DEGs

Using hallmark gene sets, GSVA showed that the changes in signaling pathway enrichment in exosome-enriched EV were associated with ‘oxidative phosphorylation’, ‘unfolded protein response’, ‘E2F targets’, ‘G2M checkpoint’, ‘WNT-beta-catenin signaling’, ‘NOTCH signaling’, ‘mTORC1 signaling’, ‘KRAS signaling DN’, ‘MYC targets V2’ and ‘DNA repair’ (Fig. S4A). The enriched GO and KEGG terms included ‘cell cycle’, ‘DNA replication’, ‘TORC1 complex’, ‘tRNA transport’, ‘regulation of mitochondrial translation’, ‘SUMO ligase complex’, ‘pentose phosphate pathway’, ‘glyoxylate and dicarboxylate metabolism’ and ‘intestinal immune network for IgA production’ (Fig. S4B and C).

Identification of candidate key genes associated with CRC progression by WGCNA

Sample dendrograms and trait heatmaps were constructed. The best soft threshold power (β) was selected as 9 to build the scaleless network. A total of 27 modules were identified via average hierarchical clustering and dynamic tree clipping. A better result was obtained using the merged module tree (Fig. 5). The matrix with the module-trait associations and corresponding P-values are shown in a heatmap (Fig. 6). The cyan (r=0.64, P=0.004) and dark olive green module (r=0.65, P=0.003) demonstrated the greatest correlations with BPs, such as ‘cell cycle’, ‘metabolism of RNA’, protein catabolic process’, ‘cellular response to chemical stress’ and ‘miRNA targeted genes in lymphocytes (Fig. S5A and B). In total, 21 candidate key genes were obtained from the intersection with the cyan module or dark olive green module identified by WGCNA.

Figure 5.

Clustering dendrogram of CRC and NC samples (both n=9) with a trait heatmap in weighted gene co-expression network analysis construction and identification of key modules. A total of 27 modules were identified following threshold merging. CRC, colorectal cancer; NC, normal control.

Figure 6.

Heatmap of the correlations between MEs and clinical traits of patients with CRC (n=9) in weighted gene co-expression network analysis construction and key modules. Pearson correlation coefficients were calculated. P-values are shown in bracket. CRC, colorectal cancer; NC, normal control; ME, module eigengene.

The PPI network was plotted according to the analysis of the STRING database. In total, four proteins did not interact with other proteins, whereas MYC was the most highly expressed gene, followed by NME2 (non-metastatic cells 2), VDAC1 (voltage dependent anion channel 1), CCT5 (chaperonin containing TCP1 subunit 5) and DKC1 (dyskerin pseudouridine synthase 1) (Fig. 7).

Figure 7.

Protein-protein interaction network. STRING database analysis revealed interactions among candidate key genes in the dark olive green and cyan weighted gene co-expression network analysis modules.

Validation of the candidate key genes with the external exosomal RNA data of exoRBase

Wilcoxon analysis revealed that the expression levels of all 21 genes in patients with CRC (n=650) were significantly different from those in healthy individuals (n=51) in TCGA-READ dataset (Table II). Following univariate logistic regression analysis, 14 candidate key genes were identified for validation with external exosomal RNA data from exoRBase (n=108). These genes included PERP (p53 effector related to PMP-22 (PERP), MYC, DKC1, VDAC1, AHCY (adenosylhomocysteinase), TMEM123 (transmembrane protein 123), RPL36A (ribosomal protein L36a), CCT5, CHCHD2 (coiled-coil-helix-coiled-coil-helix domain containing 2), RPS21 (ribosomal protein S21), KRT18 (keratin 18), TGFBI (transforming growth factor beta induced), ETS2 (ETS proto-oncogene 2) and AXIN2 (axin 2) (Fig. 8). LR was conducted on the same data from the exoRBase. The confusion matrix analysis showed relatively high efficiency in the test set (n=47). The ROC analysis confirmed the predictive power of the model in both the training (AUC=0. 890, 95% CI: 0.828-0.959) and the test set (AUC=0. 932, 95% CI: 0.818-0.962; Fig. 9A-D). The calibration curves of LR are shown in Fig. 10. A nomogram showed that the expression levels of VDAC1, RPL36A, AXIN2 and AHCY had the greatest difference between CRC and healthy samples (Fig. 11).

Table II.

Relative expression levels of 21 candidate key genes in The Cancer Genome Atlas-rectum adenocarcinoma and colon adenocarcinoma.

Gene	Colorectal cancer (n=650)	Healthy (n=51)
AHCY	130.0	37.3
CHCHD2	307.5	196.0
KRT18	441.3	219.7
PERP	131.5	50.2
TGFBI	144.3	13.7
DKC1	41.8	13.8
POMP	125.3	60.0
MYC	62.5	14.6
HMGA1	248.9	86.0
CCT5	46.8	24.5
TMEM123	92.6	53.7
ETS2	121.2	58.4
TKT	55.1	23.0
DDX21	41.1	17.0
MCM7	66.1	25.3
AXIN2	32.9	5.3
RPS21	951.5	354.4
HMGB1	44.7	27.7
VDAC1	133.3	95.3
RPL36A	1.1	0.3
NME2	3.2	0.2

Candidate key genes were analyzed using the Wilcoxon test with data from TCGA-READ (cancer.gov/tcga). The analysis revealed that the mean expression levels of all 21 genes in patients with CRC were significantly different from those in healthy individuals; all P<0.001.

Figure 8.

Univariate logistic regression analysis of the candidate key genes. A total of 14/21 candidate key genes were considered statistically significant (P<0.05) and selected for validation. OR, odds ratio.

Figure 9.

Multiple logistic regression analysis. (A) Confusion matrix (CRC, n=25; healthy control, n=83, total, n=108) and (B) ROC curve graph of training sets. (C) Confusion matrix (CRC, n=11; healthy control, n=36; total n=47) and (D) ROC curve of test sets. CRC, colorectal cancer; ROC, receiver operating characteristic; AUC, area under the curve.

Figure 10.

Bootstrap-corrected calibration curves of multiple logistic regression. Calibration curve of (A) training (CRC, n=25; healthy control, n=83, total, n=108) and (B) test sets (CRC, n=11; healthy control, n=36; total n=47). CRC, colorectal cancer.

Figure 11.

Nomogram for predicting candidate key genes in tumorigenesis of colorectal cancer. Each gene corresponded to a scaled quantitative risk score derived from regression coefficients of the model. Total score indicated the contribution magnitude of each molecular feature in the prediction model.

To validate the candidate key genes using exoRBase data, a SVM analysis model was established. Following 3-fold cross validation on the training set, a nine-molecule panel was selected, comprising VDAC1, MYC, ETS2, DKC1, AHCY, RPL36A, TMEM123, PERP and KRT18. The confusion matrix analysis showed relatively high efficiency of the test set. ROC analysis showed an AUC of 0.948 (95% CI: 0.883-1.000) in the training set and 0.881 (95% CI: 0.767-0.996) in the test set (Fig. 12A-D). The calibration curves of LR are shown in Fig. 13. SHAP values were calculated to assess the contribution of each gene to the model; VDAC, MYC and ETS2 were the top three contributors (Fig. 14).

Figure 12.

Bootstrap-corrected calibration curves of support vector machine. Calibration curve of (A) training sets (CRC n=25; healthy control, n=83; total n=108) and (B) test sets (CRC, n=11; healthy control, n=36; total n=47). (C) Confusion matrix and (D) ROC curve of test sets. CRC, colorectal cancer; AUC, area under the curve.

Figure 13.

Analysis of support vector machine. (A) Confusion matrix (CRC n=25; healthy control, n=83; total n=108), ROC curve of the training set. (C) Confusion matrix (CRC, n=11; healthy control, n=36; total n=47), ROC curve of the test sets. CRC, colorectal cancer; ROC, receiver operating characteristic curve.

Figure 14.

SHAP values of the candidate key genes. The SHAP values were calculated to indicate the contribution of each gene to the model. VDAC, MYC and ETS2 were the top three genes that contributed the most to the model. SHAP, Shapley additive explanations; VDAC, voltage-dependent anion channel; ETS2, ETS proto-oncogene 2.

Discussion

The present exploratory study analyzed exosomal mRNA. Analysis of tissue-derived exosome-enriched EV enabled comparison of exosome components from tumor and adjacent NC tissue. These exosome-enriched EVs are present in the extracellular interstitium and function as mediators of intercellular signal transduction. They may therefore more accurately reflect the pathophysiological characteristics and behavior of cells compared with exosomes derived from body fluid samples or cell lines (19).

Transcriptomic analysis identified 814 differentially expressed mRNAs between CRC and normal tissue exosome-enriched EV, suggesting that the mRNA cargo of exosome-enriched EVs is markedly altered during CRC progression (38). Exosomes play important roles in the TME by modulating immune responses, supporting metastasis and influencing angiogenesis, EMT, metabolic reprogramming and drug resistance (39,40). Functional enrichment analysis revealed that these differentially expressed mRNAs were primarily associated with ‘cell cycle’, protein synthesis and signal transduction, including the PI3K/AKT/mTOR, Wnt/β-catenin and MAPK signaling pathways. These pathways drive CRC proliferation, survival and metastasis, supporting the biological relevance of the present findings (7,37,40). Similar enrichment of pathways involved in ribosome biogenesis, ‘DNA replication’ and focal adhesion indicates that exosomal mRNAs may reflect the active metabolic and proliferative state of CRC cells (42).

GSEA and GSVA confirmed that transcriptional dysregulation in CRC-derived exosome-enriched EV extended beyond classical oncogenic signaling, with alterations in mitochondrial activity, oxidative phosphorylation and mTORC1-mediated translation control. These findings suggested that the exosomal mRNA cargo influences both metabolic reprogramming and stress response mechanisms within the TME (40). NOTCH and KRAS signaling and MYC are key players in the pathophysiogenesis of CRC (37,43,44). The regulatory and metabolic functions of MTORC1 signaling are associated with cancer cell proliferation (45). Cancer cells require substantial energy for proliferation, and their competition with neighboring tissue for substances may contribute to metabolic abnormality in these tissues (46). In addition, metabolic changes in surrounding cells may be influenced by secretions produced through metabolic activity within the TME (47).

WGCNA and univariate logistic regression identified 14 candidate key genes associated with pathogenesis of CRC, namely PERP, MYC, DKC1, VDAC1, AHCY, TMEM123, RPL36A, CCT5, CHCHD2, RPS21, KRT18, TGFBI, ETS2 and AXIN2. WGCNA was exploratory because of the small sample size. Using external datasets from exoRBase, these key exosomal mRNAs were validated in predictive modeling via logistic regression analyses and SVM. exoRBase is a repository of human biofluid exosomal RNA data, whereas the data in the present study were tissue-derived. Tissue-derived exosomes differ from biofluid-derived exosomes in both source and composition (19). Therefore, differences between biofluid and tissue-derived exosomes may affect the accuracy of the validation.

The majority of proteins encoded by the candidate key genes are reported to be involved in CRC tumorigenesis (48). Notably, PERP is a plasma membrane protein involved in p53-mediated apoptosis, desmosomal cell-cell adhesion and stratified epithelial integrity (49) and it plays a role in the biological process of CRC (50,51). MYC and AXIN2 are classical regulators of Wnt/β-catenin signaling (52,53), a well-established key pathway for CRC. MYC is involved in the carcinogenesis of numerous types of cancer, including CRC (54). c-MYC, a member of MYC family, is associated with resistance to EGFR inhibitors in CRC (55). DKC1, a key enzyme involved in ribosomal RNA modification, promotes colon cancer progression by regulating ribosomal proteins and the RAS/RAF/MEK/ERK signaling pathway (56). VDAC1 is critical for mitochondrial metabolism and apoptosis regulation, supporting its mechanistic importance in CRC tumorigenesis (57). AHCY regulates intercellular transmethylation and may be a therapeutic target in CRC (58). TMEM123 is a key player in immune surveillance in CRC (59). CCT5 is involved in protein folding, stability and cell structure maintenance. CCT5 serves as a prognostic biomarker for CRC (60,61). KRT18 is an oncogene for CRC (62). TGFBI is an extracellular protein that regulates cell adhesion and serves a significant role in cancer progression (63). ETS2 encodes a transcription factor that regulates genes involved in development and apoptosis and is involved in CRC tumorigenesis (64). RPL36A is a component of the ribosomal protein complex. It is significantly upregulated in CRC tissue and associated with poor prognosis, likely through the suppression of the MYC and MAPK/ERK pathways (65).

To the best of our knowledge, there are no reports showing that RPS21 and CHCHD2 are associated with CRC. RPS21 serves a role in protein synthesis (66). CHCHD2 is a mitochondrial function-associated protein and is mainly involved in physiological processes such as cell migration, antioxidant stress and apoptosis regulation (67). Therefore, the biological relevance of RPS21 or CHCHD2 with tumorigenesis of CRC should be further explored based on the findings of the present study.

Taken together, these findings indicate that exosomal mRNAs in CRC tissue carry distinct molecular signatures reflective of tumor biology, immune modulation and oncogenic signaling. These findings highlight the potential of tissue-derived exosomal transcriptomics as a novel platform for understanding TME interactions. However, the present study has several limitations, including its small sample size and lack of additional RT-qPCR verification and functional validation of key candidate genes. Future studies with larger cohorts and mechanistic investigations are needed to establish the causal roles of these exosomal mRNAs in CRC development and to explore their utility in non-invasive diagnostic settings, such as plasma exosome analysis. Furthermore, biological replicate validation for western blotting was not performed in the present study. This is a technical limitation as it does not account for biological variation across samples, which may limit the generalizability and robustness of the protein expression data. Another technical limitation is that total membrane protein staining was not conducted to verify the sample loading uniformity. Replicate validation for western blotting and total protein staining are required to strengthen data reliability validation and loading consistency.

In conclusion, the present exploratory study provides a detailed profile of tissue exosomal mRNAs in patients with CRC. These findings may expand understanding of CRC exosome biology and provide a foundation for validation in larger experimental studies.

Funding Statement

Funding: The present study was supported by the Scientific Research Fund of Beijing Rehabilitation Hospital, Capital Medical University (grant nos. 2023-10, 2025-21 and 2025-22) and the Scientific Research Fund of Beijing Anorectal Society (grant no. 2020ABCP002).

Availability of data and materials

The data generated in the present study may be found in the Genome Sequence Archive in National Genomics Data Center, China National Center for Bioinformation under accession number HRA016508 or at the following URL: ngdc.cncb.ac.cn/gsa-human.

Authors' contributions

QG, CBK, HYS and XWL designed the study. XWL, MX, FYS, JZ, MQZ, YW, DYD and JFS performed the experiments. QG, CBK, YMK, DYD, YW and XWL analyzed the data. QG, CBK, XWL and HYS wrote the manuscript. QG and YMK revised the manuscript. QG, CBK, YMK and XWL confirm the authenticity of all the raw data. All authors have read and approved the final manuscript.

Ethics approval and consent to participate

All procedures used in this research were approved by the Ethics Committee of Beijing Rehabilitation Hospital, Beijing, China (approval no. 2021bkky-149). Written informed consent was obtained from all participants. All methods were performed in this study were in accordance with the Declaration of Helsinki.

Patient consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.