We retrieved serum miRNA expression data and corresponding clinical information from the NCBI Gene Expression Omnibus (GEO) database, applying the following inclusion criteria: 1) Samples from cancer groups were collected from treatment-naive patients with clear diagnoses of malignant tumors; 2) Samples from control groups were obtained from healthy volunteers or individuals with benign diseases; 3) No duplicate samples were included. In total, 20,702 samples from ten individual GEO datasets (GSE106817, GSE112264, GSE113486, GSE113740, GSE122497, GSE137140, GSE139031, GSE85589, GSE164174, GSE124158), were enrolled in this study (Supplementary Table S1). Those datasets included 12,934 cancer-free and 7,768 cancer samples across 12 cancer types, including bladder urothelial carcinoma (BLCA), breast invasive carcinoma (BRCA), colorectal cancer (CRC), esophageal carcinoma (ESCA), glioma (GBM), gastric cancer (GC), hepatocellular carcinoma (HCC), lung cancer (LC), pancreatic adenocarcinoma (PAAD), prostate adenocarcinoma (PRAD), sarcoma (SARC), and ovarian cancer (OV).

The Affymetrix datasets along with corresponding annotation files were downloaded individually. When the probes were converted to gene symbols, if multiple probes corresponded to the same gene symbol and their mean value was taken as the final value. All miRNA expression were integrated once every dataset was processed, normalized, and log2 transformed. To detect and remove the batch effect caused by non-biotechnological bias, we conducted batch correction utilizing the ComBat function from the R package sva (v.3.38.0) (Leek et al., 2012). In total, an expression matrix consisting of 20,702 samples and 2,525 miRNAs was utilized for further analysis.

In this study, a total of 33 m⁷G target miRNAs were extracted from previously published studies (Liu et al., 2019b; Pandolfini et al., 2019) (Supplementary Table S2). To investigate the biological processes in which these miRNAs are involved, we initially predicted their potential targets using the miRWalk3.0 database (http://mirwalk.umm.uni-heidelberg.de/search_genes), with binding probability >95% and binding site position 3′UTR (Sticht et al., 2018). Only experimentally validated miRNA-target interactions from miRTarBase (https://mirtarbase.cuhk.edu.cn/), were considered for miRNA targets.

Differentially expressed m⁷G-harboring miRNAs between the cancer-free and cancer groups were identified using the R package limma (v3.10.3). The false discovery rate (FDR) was controlled using the Benjamini and Hochberg method. miRNAs with FDR <0.05 and |log2FC| >1, were considered significantly differentially expressed. Gene ontology (GO) enrichment and KEGG pathway enrichment analyses were performed to the biological functions and pathways associated with these genes using R package ClusterProfile (v3.6.0) (Yu et al., 2012).

To build a diagnostic signature for cancer detection, all enrolled samples (N = 20,702) were randomly equally divided into training and validation sets using the createDataPartition function of the R package caret (v6.0.88). Each group consisted of 10,351 samples, with 6,467 cancer-free controls and 3,884 cancers in each set. Details on those samples were presented in the Supplementary Table S3. Based on 22 differentially expressed m⁷G-harboring miRNAs (Supplementary Table S4), the least absolute shrinkage and selection operator (LASSO) analyses (100-fold cross-validation) was performed to determine the candidate m⁷G target miRNAs by using R package glmnet (v4.0.2). A diagnostic signature for cancer detection was established using 12 m⁷G target miRNAs (Supplementary Table S4). The model calculation formula is: m 7 G − miRNA score = ∑ i = 0 n C o e f i * x i where Coef_i is the LASSO coefficient of each miRNA, and x _i is the expression level of each miRNA.

All statistical analyses and picture drawing were implemented in R statistical software (v4.0.5). The diagnostic performance of the m⁷G-miRNAs signature was evaluated by receiver operating characteristic (ROC) curve analysis, including the area under the curve (AUC), sensitivity, specificity and accuracy. The relationship of m⁷G-miRNAs and each miRNA was determined using Spearman correlation analysis. Statistical significance between groups was calculated by a Wilcoxon test or Kruskal‒Wallis test at a confidence interval (CI) of 95%, with p < 0.05 considered statistically significant.

A schematic diagram of the study design was presented in Figure 1A. In total, an integrative dataset of 20,702 serum samples and 33 m⁷G target miRNAs was utilized to build a m⁷G target microRNAs diagnostic signature (m⁷G-miRDS) for cancer detection. In the training and validation cohort, the average age of participants in the both was 62.5 years, and the proportion of female was 30.7% and 31.3%, respectively.

GO and KEGG enrichment analysis were performed to explore the biological processes regulated by m⁷G target miRNAs. The results underscored the predominant enrichment of these genes in pathways related to cancer and RNA metabolism, including TGF-β receptor signaling pathway, RNA metabolism process and RNA localization (Supplementary Figures S1D–SE; Supplementary Table S5).

Utilizing expression profiles of 33 pan-cancer datasets from The Cancer Genome Atlas (TCGA) project, we observed aberrant expression of METTL1 and WDR4, known as m⁷G methyltransferase, across multiple cancer types. Specially, the two genes were downregulated in colon adenocarcinoma compared to controls, while upregulated in diffuse large B-cell lymphoma and thymoma. In addition, they demonstrated significant independent prognostic value in cancers such as BLCA, ESCA and OV, as illustrated in Supplementary Figures S1A-C. These findings implied that m⁷G dysregulation was closely related to tumor occurrence and progression.

In the training set, 22 of 33 m⁷G target miRNAs were significantly differentially expressed when comparing the cancer group with the cancer-free controls (FDR<0.05 and log₂|fold change| > 1). Among these, 21miRNAs exhibited significant upregulation in cancers, and only miR-125a-3p was significantly downregulated (Figure 1B and Supplementary Table S4). Then, we identified a total of 12 candidate m⁷G target miRNAs using the LASSO (Supplementary Figures S1F,G). Supplementary Table S4 listed 12 m⁷G target miRNAs and their corresponding coefficients. Based on the expression profiles of the 12 candidate miRNAs, unsupervised hierarchical clustering demonstrated a distinct separation between cancers and cancer-free controls (Figure 1C). This observation was further supported by principal component analysis (Figure 1D). These results underscored significant differences in expression patterns between two groups, providing valuable insights for the development of a diagnostic signature for cancer detection.

The diagnostic performance of each candidate miRNA was assessed individually for cancer detection. Each miRNA exhibited a significant diagnostic capability in distinguishing cancers from cancer-free controls, with AUC values ranging from 0.742 for miR-193a-5p to 0.955 for miR-663a (Figure 1E). The similar finding was also observed in the validation cohort (Figure 1F). Those results suggested that identified candidate m⁷G target miRNAs hold promise as biomarkers for cancer detection.

Utilizing the 12 identified m⁷G target miRNAs, we built a diagnostic signature for cancer detection, named m⁷G-miRDS. The m⁷G-miRNAs score in cancers was significantly higher than that in cancer-free controls (Figure 2A). The distribution of m⁷G-miRNA scores in each cancer type showed that BRCA patients had the lowest median m⁷G-miRNA scores, while BLCA patients had the highest median m⁷G-miRNA scores (Figure 2B). The m⁷G-miRDS model, a 12-m⁷G-miRNAs set, showed a higher diagnostic power than each candidate miRNA alone in distinguishing cancer samples from cancer-free controls in the training cohort, with an AUC and 95% confidence interval (CI) of 0.974 (0.971–0.977) (Figure 2C). Similar to the training cohort, the m⁷G-miRDS model also showed a high diagnostic performance in the validation cohort, with an AUC of 0.972% and 95% CI, 0.969 to 0.975 (Figure 2D). Additionally, the m⁷G-miRDS demonstrated a satisfactory diagnostic value in the entire set (Figure 2E).

By Spearman correlation analysis, we observed a remarkable positive correlation between m⁷G-miRNAs and 12 miRNAs, except hsa-miR-125a-3p (Figure 2F and Supplementary Table S6). Previously published studies reported the important value of hsa-miR-93 and hsa-miR-122 in the diagnosis and prognosis of various cancer types (Luna et al., 2017; Gao et al., 2019; Faramin Lashkarian et al., 2023). In decision curve analyses, m⁷G-miRNAs demonstrated an absolute superiority net benefit within a wide range of decision-making threshold probabilities compared to hsa-miR-93 and hsa-miR-122 neither in training set not validation set (Figures 2G,H).

Considering the effect of patient clinicopathological characteristics on diagnostic efficacy, such as gender, age and pathological stage, we tested the diagnostic performance of m⁷G-miRDS stratified by patient gender. No significant difference in m⁷G-miRNA scores was observed between female and male patients (p = 0.45, Figure 3A). Similar to gender, we also did not observe a significant correlation between m⁷G-miRNA scores and patient age (R = 0.07, Figure 3B). According to the subgroups classified by tumor stage, there was no significant difference in the m⁷G-miRNA scores between different pathological stages, except for stage I and stage III (Figure 3C). In conclusion, m⁷G-miRDS was not interfered by the gender and age of patients, underscoring its robustness as a biomarker for distinguishing cancers from controls.

To evaluate the discriminative ability of m⁷G-miRNAs across various cancer types, we mixed each cancer type separately with cancer-free controls. As expected, m⁷G-miRDS showed superior discriminatory capability, with the highest AUC and 95% CI for bladder cancer in the training and validation cohort (0.982, 0.979–0.986 and 0.981, 0.978–0.985, respectively), and the lowest AUC and 95% CI for breast cancer in both sets (0.954, 0.946–0.963 and 0.955, 0.947–0.963, respectively) (Figure 3D and Supplementary Table S7). Despite a slight diminish in diagnostic ability of m⁷G-miRDS for individual cancer type, it maintained very high sensitivity. The results suggested that m⁷G-miRDS could identify over 92% of patients with a given cancer type, resulting in a lower incidence of missed diagnosis.

Among ten datasets, the dataset GSE164174 included samples derived from early gastric cancer, and the training and validation cohort consisted of 698 cancer-free controls and 711 samples from patients with early gastric cancer, with 705 stage I and 5 stage II samples. In Figure 3E, the m⁷G-miRDS demonstrated promising performance in diagnosing early gastric cancer, with an AUC of 0.973 (95% CI, 0.970–0.975), a specificity of 0.910, a sensitivity of 0.984 and an accuracy of 0.947.

In addition, the dataset GSE113740 included 46 chronic hepatitis and 93 liver cirrhosis serum samples, thus we also investigated the capability of m⁷G-miRDS in distinguishing between cancer and non-tumor diseases. The model exhibited a remarkable accuracy in discriminating HCC individuals from patients with chronic hepatitis or liver cirrhosis (Figures 3F–H), with an AUC of 0.846 (95% CI: 0.805–0.888), surpassing the performance of traditional biomarkers, such as alpha-fetoprotein (AUC, 0.65; specificity, 51.4%; sensitivity, 73.3%) (Yamamoto et al., 2020). The above results indicated that chronic disorders might not significantly affect the diagnostic performance of m⁷G-miRDS.