Circulating microRNA panels for multi-cancer detection and gastric cancer screening: leveraging a network biology approach

Abstract

Background

Screening tests, particularly liquid biopsy with circulating miRNAs, hold significant potential for non-invasive cancer detection before symptoms manifest.

Methods

This study aimed to identify biomarkers with high sensitivity and specificity for multiple and specific cancer screening. 972 Serum miRNA profiles were compared across thirteen cancer types and healthy individuals using weighted miRNA co-expression network analysis. To prioritize miRNAs, module membership measure and miRNA trait significance were employed. Subsequently, for specific cancer screening, gastric cancer was focused on, using a similar strategy and a further step of preservation analysis. Machine learning techniques were then applied to evaluate two distinct miRNA panels: one for multi-cancer screening and another for gastric cancer classification.

Results

The first panel (hsa-miR-8073, hsa-miR-614, hsa-miR-548ah-5p, hsa-miR-1258) achieved 96.1% accuracy, 96% specificity, and 98.6% sensitivity in multi-cancer screening. The second panel (hsa-miR-1228-5p, hsa-miR-1343-3p, hsa-miR-6765-5p, hsa-miR-6787-5p) showed promise in detecting gastric cancer with 87% accuracy, 90% specificity, and 89% sensitivity.

Conclusions

Both panels exhibit potential for patient classification in diagnostic and prognostic applications, highlighting the significance of liquid biopsy in advancing cancer screening methodologies.

Introduction

Regular cancer screenings can lead to timely diagnosis, enabling effective intervention and increasing the likelihood of successful treatment. Currently approved cancer detection tests, such as endoscopy, colonoscopy, mammography, and cervical cytology, are often aggressive and costly [1]. On the other hand, while clinical signs of cancer progression are often elusive in the early years, underlying molecular changes in the body are still occurring [2, 3]. Hence, exploring non-invasive molecular biomarkers within a broader cancer diagnostic window emerges as a promising strategy. These biomarker panels can be beneficial in evaluating high-risk individuals and serve as the basis for establishing follow-up strategies. It is crucial that they demonstrate high specificity and accuracy to minimize unnecessary follow-up procedures and reduce anxiety.

Multi-cancer screening is challenging due to the diverse nature of cancers in terms of genetic and molecular characteristics. Liquid biopsy is one of the most promising options in noninvasive cancer detection. The term "liquid biopsy" did not appear in PubMed until 2011 [4]. Since then, liquid biopsy, involving the extraction of non-solid biological material from sources like blood, saliva, urine, cerebrospinal fluid, breast milk, and semen, has consistently grown as a research field [5, 6]. These samples are obtained from the body through non-invasive or minimally invasive methods.

Blood is simply reachable and contains a diverse set of analytes, which can be used for medical purposes such as early diagnosis, treatment monitoring, tumor stage determination, disease recurrence risk assessment, and prognosis [7]. Several studies have dedicated efforts to classify cancers using a combination of blood analytes such as proteins, cell-free DNAs, noncoding RNAs, circulating tumor cells (CTCs), and exosome molecular content [8, 9]. For instance, the CancerSEEK project aimed to diagnose and locate operable cancers (ovarian, liver, stomach, pancreas, intestine, lung, and breast) using a multi-element blood test. They classified different cancer samples with an accuracy of between 69 and 98% and were able to determine healthy samples with a high specificity of 99% [1]. In a recent investigation, a platform(I-Biomarker) has been developed for multi-cancer diagnosis and treatment response monitoring. This platform incorporates bioinformatic pipelines for various omics, including RNA-seq, Small RNA-seq, Variants (SNP and Indels), Copy Number Variation, Single Cell RNA-seq, CHIP-Seq, Methylation, and miRNA. They employed Explainable AI to elucidate predictions based on corresponding biomarkers [10].

Transitioning to a focus on microRNAs (miRNAs) these small noncoding RNAs play crucial posttranslational regulatory roles [11]. Regarding their sequence complementarity miRNA can target the 3′-untranslated region (3′-UTR) of messenger RNA and suppress protein translation [12]. Their expression profiles were indicated to be able to classify human cancers [13]. MiRNAs act as oncogenes or tumor suppressors during tumor development and progression. Experimental evidence demonstrates that correcting specific miRNA alterations using miRNA mimics or antagomirs can normalize the gene regulatory network and signaling pathways and reverse the phenotype in cancerous cells [14]. Despite the absence of standardized protocols for the current clinical use of miRNAs, they constitute a reliable tool for future use [15].

There are different analysis methods for the classification of patient groups. The preference for network biology over previous methods, such as fold change analysis, lies in its ability to capture and analyze complex interactions within biological systems. While fold change analysis focuses on individual gene expression changes, network biology can bridge individual genes to a system-level view, which considers the relationships and dependencies among genes, miRNAs, proteins, and other molecular entities, which are then expected to be enriched in individual pathways, and processes.

Correlation networks are increasingly used in bioinformatics applications to facilitate network-based methods of identifying candidate biomarkers or therapeutic targets [16]. In gene co-expression networks, each gene corresponds to a node, and an edge connects two genes if their expression values are highly correlated. The Weighted Gene Co-Expression Network (WGCNA) assigns a weight to each link to represent the extent of gene co-expression. WGCNA has already been applied to several cancers, and promising results have been obtained [17]. After detecting clusters (modules) of highly correlated genes, this package can summarize such clusters using the module eigengene or an intramodular hub gene, correlate modules to one another or external sample traits (eigengene network methodology), and calculate module membership measures [16]. In network biology, researchers often explore the preservation of modules across multiple networks and assess how gene pathways are affected under specific conditions. Non-preserved modules can also provide valuable information, reflecting species-specific or disease-specific modules [18]. In the second part of the study, a preservation analysis is adopted to detect specific modules in gastric cancer.

Numerous previous studies have explored the expression profiles of miRNAs in various cancers, aiming to identify associated patterns, as summarized in Supplementary Table 1, Additional File 1. However, it remains unclear whether some specific miRNAs can accurately divide the statistical population into two categories, healthy people and individuals at high risk of cancer in general. Identifying common alterations in miRNAs across multiple cancers could provide valuable insights into the shared regulatory mechanisms of these small RNA molecules.

Considering the important role of multi-cancer screening, this study, conducts an inclusive examination of serum miRNA profiles across thirteen different cancer types, to identify common biomarkers for multi-cancer screening. Employing weighted miRNA co-expression network analysis, the study successfully identifies modules and hub miRNAs that demonstrated shared characteristics across the diverse spectrum of cancers. Subsequently, we assess the capability of this strategy in detecting specific cancers, such as gastric cancer, from other cancers through a preservation analysis. The machine learning outcomes revealed two distinct miRNA panels: one showcasing robust performance in multi-cancer screening (96.1% accuracy, 96% specificity, and 98.6% sensitivity), and the other displaying promise in gastric cancer detection (87% accuracy, 90% specificity, and 89% sensitivity). With potential applications in diagnostic and prognostic contexts, these findings underscore the role of liquid biopsy in advancing cancer screening methodologies with a focus on commonalities among diverse cancer types.

Materials and methods

This study unfolds in two distinct parts. In the initial phase, researchers dedicated their efforts to accurately classify samples into two groups: non-cancer and those suffering from one of thirteen different cancers (regardless of the cancer type). Transitioning into the second phase, the study introduces an approach to pinpointing more precise cancer markers. In this phase, the focus is on identifying miRNAs specifically indicative of gastric cancer as a particular disease. Although gastric cancer ranks seventh among cancers in terms of incidence, it is the fourth leading cause of cancer-related deaths in the world [19]. This might be due to the absence of symptoms in the early stages of the disease, which makes gastric cancer detection difficult. Physicians often do not diagnose the disease until it reaches advanced stages, and at that time, treatment is much more difficult and less responsive. The analysis flow is illustrated in Fig. 1.

Fig. 1

Study analysis flow describing steps in multi-cancer screening and gastric cancer classification

Datasets for training and evaluation

This study independently utilized the GSE113486 dataset for training and the GSE112264 dataset for evaluation, both sourced from the Gene Expression Omnibus [20,21,22,23]. These datasets comprise serum samples from 13 distinct cancer types.

Before conducting any additional analysis, it is imperative to normalize the data. Over the last decade, Limma has emerged as a widely used tool for microarray data analysis. This tool offers an integrated solution for the comprehensive analysis of data derived from gene expression experiments [24]. Using this package, we downloaded and normalized the data.

Multi-cancer screening

In the initial dataset, we categorized all samples into two study groups to build the network based on it. Group 1 comprised 100 non-cancer control samples, while Group 2 consisted of 872 samples representing thirteen cancer types. We constructed the network and prioritized miRNAs based on their significance across various cancers. Subsequently, we rigorously assessed the miRNA candidate list through diverse machine-learning models. Further elaboration on these steps will be provided in subsequent sections.

Co‑expression network analysis

Using the WGCNA package, an unsigned weighted miRNA co-expression network was constructed by comparing the miRNA expression profiles of the two mentioned groups. Given that miRNAs are regulatory factors, and any change in their expression may play a critical role, an unsigned network was chosen to capture both positive and negative correlations.

In the network, the nodes represent miRNA expression profiles, and the edges denote pairwise correlations between them. The adjacency matrix, aij, is a symmetric n × n matrix with continuous values between 0 and 1. Specifically, aij = ∣cor(xi,xj)∣ represents the adjacency in the unsigned network, where Pearson's correlation coefficient was used to construct the correlation matrix. Pearson correlation is appropriate for analyzing the relationship between two quantitative variables, assuming they may exhibit a positive, negative, or no linear relationship.

Next, a weighted network adjacency was defined through soft thresholding, which involves raising the absolute value of the correlations to a power greater than one. This approach emphasizes stronger correlations while downplaying weaker ones [25].

Co-expressed miRNAs were then clustered into modules, and the eigengene dendrogram of each module was generated. Following this, the correlation between the module eigengenes and cancer presence was examined using the Pearson correlation.

The filtering of miRNAs was based on two criteria: |mTS| (miRNA Trait Significance) > 0.5 and |mMM| (miRNA Module Membership) > 0.71 within their respective module. This threshold, labeled "Alpha" was applied to the network to yield a list of potential miRNA candidates.

Evaluation by machine learning

Before evaluating the chosen miRNAs through machine learning, balancing sample numbers and adjusting distribution within each group is crucial. This can be achieved by random sampling from each cancer type. Group 1 consisted of 100 non-cancer control samples, and Group 2 labeled as Cancerous, included eight randomly selected samples from each cancer type, totaling 104 cancerous samples.

We assessed the performance of eight algorithms on the candidate miRNA list to identify the most suitable algorithm for our specific purpose and dataset. The algorithms are as follows:

• CART: Classification and Regression Trees

• GLMNET: Regularized Linear Models

• KNN: K Nearest Neighbor

• LDA: Linear Discriminant Analysis

• LR: Linear Regression

• MLP: Multi-Layer Perceptron

• RF: Random Forest

• SVM: Support Vector Machine

To this end, the Classification and Regression Training (Caret) package in the R programming language was used [26, 27]. The evaluation parameters were ROC, sensitivity, and specificity.

Then, by narrowing the threshold alpha, the number of miRNAs was decreased to 25, 17, and 4, and each list was evaluated using the chosen algorithm.

Gastric cancer screening

To pinpoint specific miRNAs associated with a particular cancer type, such as gastric cancer, the second phase of the study focused on constructing a Weighted MiRNA Co-Expression Network (WMiCN) tailored to this specific cancer. The revised sample selection included Group 1 (non-cancer control with 100 samples) and Group 3 (gastric cancer with 40 samples).

Co-expression-network analysis

The construction of the unsigned WMiCN (Weighted miRNA Co-expression Network) was based on the expression profiles of samples from Group 1 and Group 3. As in the initial phase, co-expressed miRNAs were clustered into modules, and the correlation between these modules and the trait of interest—gastric cancer incidence—was explored.

Preservation analysis

In many applications, researchers seek to determine whether network module properties change across different conditions. In our case, we aim to investigate which modules in the gastric cancer network are specifically related to gastric cancer and not preserved in the network representing other cancer types. To identify these gastric cancer-specific modules, we conducted a preservation analysis by comparing circulating miRNA expression profiles in gastric cancer with those in other cancer types.

Various preservation statistics are available to measure this preservation among networks, and since each one measures different aspects, their results may not always align. Therefore, composite preservation metrics offer a more comprehensive and efficient evaluation of module preservation across networks.

The overall significance of the observed preservation statistics can be assessed using Zsummary, which is calculated as the mean of Z.density and Z.connectivity. According to Langfelder et al. [28], the following thresholds apply:

• Zsummary > 10: Strong evidence of module preservation

• 2 < Zsummary < 10: Weak to moderate evidence of preservation

• Zsummary < 2: No evidence of module preservation

It is important to note that Zsummary depends on module size, meaning that the preservation of larger modules tends to be statistically more significant. However, to treat all modules equally, regardless of size, we focused on the medianRank composite statistic. Analogous to Zsummary, medianRank is defined as the mean of medianRank.density and medianRank.connectivity. Unlike Zsummary, medianRank is not affected by module size, making it useful for comparing relative preservation among modules. Modules with lower medianRank values demonstrate stronger evidence of preservation compared to those with higher values.

We used these two statistics to identify non-preserved modules. The hub miRNAs of the non-preserved modules were then selected for further scrutiny based on the criteria |mTS|> 0.8 and |mMM|> 0.71 within their respective modules. This threshold, referred to as "Beta," was applied to the network, resulting in a refined list of 51 candidate miRNAs.

Evaluation by machine learning

To address the need for balanced sample sizes in classification problems, two groups were carefully formed to ensure an optimal distribution: Group 3, comprising 40 samples representing gastric cancer, and Group 4, consisting of 48 samples randomly selected, with four samples chosen from each of the remaining twelve cancer types.

Similar to the preceding phase, eight distinct machine learning algorithms were scrutinized, and the most effective one was chosen. Employing the selected algorithm, we assessed the suitability of the candidate list for classifying both Group 3 and Group 4. By refining the Beta threshold, we subsequently reduced the number of miRNAs first to 17 and then to 4, evaluating their ensuing performance.

In the next step, we validated both multiple and particular cancer screening models using the independent test dataset. Performance reports of both models determine the panel's ability in cancer classification.

Enrichment analysis

In the next step, functional enrichment and pathway analyses were performed for the candidate miRNA list in both study phases. To investigate the miRNAs’ function and disease associations, TAM 2.0 can be used, which performs the enrichment analysis of these miRNAs in the curated miRNA sets using the statistical overrepresentation analysis [29]. The other tools used in this study were miRPathDB, miRNET, Mienturnet, and Enrichr [30,31,32,33]. Finally, we reviewed the previously published studies on the two final miRNA panels.

Results

The samples’ general characteristics and distribution in both train and test datasets are presented in Tables 1 and 2, respectively. Within the first dataset, the original study focused on bladder cancer, which involved random division of data into two equal groups for training and testing (n = 486 per group). They ultimately proposed a panel of seven miRNAs (miR-6087, miR-6724-5p, miR-3960, miR-1343-5p, miR-1185–1-3p, miR-6831-5p, miR-4695-5p) for classifying bladder cancer samples against normal and other cancer samples, achieving an AUC of 0.97, sensitivity of 95%, and specificity of 87%. Descriptive variables were compared using the χ2 test, while continuous variables were compared using the independent t-test. Also, Fisher's linear discriminant analysis was used to analyze the expression of multiple miRNAs. The returned values obtained from the discriminant functions were used to define the index. A score of ‘zero’ or above represented bladder cancer, and scores below zero represented other cancers or clinical conditions [20].

Table 1 The general characteristics of train and test datasets

Table 2 The distribution of the disease in the samples across the independent training and test datasets

Using the second dataset Urabe et al. investigated the potential of serum miRNAs as a precise diagnostic tool for patients suspected of having prostate cancer. Candidate miRNAs for prostate cancer detection were identified in the discovery set (n = 123). Using combinations of these candidate miRNAs, a diagnostic model was constructed in the training set (n = 484), and its performance was assessed in the validation set (n = 484). The resulting robust diagnostic model featured two miRNAs, namely miR-17-3p and miR-1185–2-3p. It demonstrated high diagnostic performance with a sensitivity of 90% and a specificity of 90% in the validation set, irrespective of the Gleason score and the clinical tumor-node-metastasis stage [20, 21].

In our study, we utilized both datasets differently, with the first used for training and the second for testing. Unlike the original studies, which focused specifically on bladder and prostate cancers, our first phase investigates 13 different cancer types, and the second phase focuses on gastric cancer. While we proposed two distinct panels of four miRNAs for MCD(Multi-Cancer Detection) and GCD(Gastric Cancer Detection), the original studies developed panels for classifying bladder and prostate cancers. Methodologically, we employed co-expression network analysis, whereas the original studies applied statistical techniques such as Fisher's linear discriminant analysis and χ2 tests to analyze miRNA expression and classify samples.

Multi-cancer screening results

In the first step, patient clustering was performed. To achieve a co-expression network with a balance between scale independence and mean connectivity, both metrics were calculated across power values ranging from 1 to 20 using the soft-connectivity algorithm. The optimal power value was selected when the scale independence reached 0.9. Finally, the co-expression matrix was computed using this determined power value.

In the first phase of our study, a power of 13, satisfied the criteria for topological scale-freeness, as shown in Supplementary Figure S1A, Additional File 2. Based on this power the adjacency matrix was calculated and the weighted miRNA correlation network was constructed. Finally, a cluster dendrogram of the modules was generated and seven modules were identified with the labels red, brown, turquoise, yellow, blue, green, and gray, containing 99, 213, 258, 206, 242, 145, and 1402 miRNAs, respectively, as shown in Supplementary Figure S1B, Additional File 2.

By examining the correlation between the modules and cancer in general, the highest correlation coefficients belonged to the yellow (0.62), turquoise (0.59), green (0.54), and brown (0.52) modules (Fig. 2A). To confirm the validity of this assessment, we also looked into each module and examined the correlation between the miRNA module membership factor and its significance in cancer. In this evaluation, the highest correlation rates were observed for the yellow (0.77), turquoise (0.7), green (0.7), and blue (0.63) modules (Fig. 2B).

Fig. 2

Correlation analysis of modules and cancer. (A) Correlation analysis between each module eigen miRNA and cancer as a trait. (B) Correlation analysis between miRNA module membership and miRNA significance in cancer

Following the application of the alpha threshold, 70 miRNAs were chosen. Supplementary Table 2, Additional File 3, provides the list of these miRNAs, along with their distribution across different modules. Figure 3 illustrates the comparative results of the performance of the selected machine learning algorithms. This evaluation was based on specificity, sensitivity, and the area under the ROC curve. The random forest algorithm was selected according to its appropriate performance regarding the three evaluation criteria for this dataset.

Fig. 3

Comparison of performance between different Machine Learning Algorithms using multi-cancer screening miRNA panel with a confidence level of 0.95

The multi-cancer screening model performance results using the candidate miRNA lists are presented in Table 3. Classifying cancer samples from non-cancer control samples, the best result was achieved by 4-miRNA panel containing "hsa-miR-614", "hsa-miR-1258", "hsa-miR-548ah-5p", and "hsa-miR-8073" with 96.1, 96, and 98.6 specificity, sensitivity, and ROC on original data by fivefold cross-validation and 94.9, 94.8, 97.5 for accuracy, specificity, and sensitivity on the independent dataset.

Table 3 The performance results of the multi-cancer screening model on candidate miRNA lists

Considering that the previous evaluation was performed collectively and given that the accuracy might be different for each cancer, further evaluation was performed for different cancers separately. Figure 4 illustrates the model performance results of each cancer type.

Fig. 4

Comparison of model performance for each cancer type separately using 70-miRNA and 4-miRNA models

Gastric cancer screening results

In the process of classifying gastric cancer, prior to creating the network, we initially detected an outlier among the non-cancer samples, specifically with the ID gsm3107261. Consequently, this outlier was excluded from the analysis. Subsequently, the network was established with a power setting of 9 to achieve scale-freeness, as determined by mean connectivity and scale independence metrics (Supplementary Figure S2A, Additional File 4). A dendrogram representing the miRNAs was generated, leading to the identification and labeling of seven distinct modules as green, yellow, brown, blue, turquoise, and gray, each comprising 45, 96, 202, 603, 747, and 872 miRNAs, respectively (Supplementary Figure S2B, Additional File 4).

In the examination of module correlations with gastric cancer, four modules, namely blue, turquoise, green, and yellow, exhibited the strongest correlations (Fig. 5). It's worth noting that, following the guidelines outlined in the WGCNA manuals, the gray module was disregarded in this analysis.

Fig. 5

Detecting most correlated miRNA modules to Gastric cancer. (A) Correlation between module eigen miRNAs and Gastric cancer incidence. (B) Correlation between miRNAs module membership and their significance in Gastric cancer

In the preservation analysis, the blue module holds the 5th position and displays the lowest preservation level among other cancer types, regardless of the gray module, as demonstrated in Table 4. Accordingly, it stands out as the most distinctive module specific to gastric cancer. The comparison between the median rank and Zsummary criteria is represented in Fig. 6.

Table 4 Preservation analysis results, the blue module shows the lowest level of preservation among other cancers

Fig. 6

Preservation analysis indicates which modules are more specific in gastric cancer. (A) Comparison based on Median Rank. (B) Comparison based on preservation Zsummary

Following applying the Beta threshold, a total of 51 miRNAs were chosen for inclusion in the candidate list. Out of these, 50 miRNAs were sourced from the blue module, recognized as the most distinct module in gastric cancer, while one miRNA originated from the green module (Supplementary Table 3, Additional File 5).

In the machine learning assessment, eight different algorithms were compared regarding their classification performance with the candidate miRNAs as presented in Fig. 7. The GLMNET algorithm showed the best performance overall. Table 5 shows the results for the candidate list consisting of 51, 17, and 4 miRNAs, respectively.

Fig. 7

Performances comparison between the machine learning algorithms using the gastric cancer screening panel with a confidence level of 0.95

Table 5 The performance results of the gastric cancer classification model using different candidate miRNA lists

The panel of 4 miRNAs can discriminate between a set of 12 cancers along with normal samples and gastric cancer patients using the GLMNET algorithm with the specificity, sensitivity, and area under ROC of 87, 90, and 89, respectively. The accuracy, specificity, and sensitivity in the independent dataset were 79, 78, and 86, respectively. The model performance results appraisal using the independent test data for each cancer type are presented in Fig. 8.

Fig. 8

Comparison of the trained model's performance in distinguishing gastric cancer from other cancer types and non-cancer controls using the 51-miRNA model and the 4-miRNA model

In the final step, functional enrichment was performed. After examining the enrichment results, sufficient evidence was provided for the effect of selected miRNAs in cancer in both phases of the study.

The miRNA-mRNA interaction network analysis with MiRNet online software revealed that the miRNAs and their target genes were primarily involved in different cancers, including lung, breast, ovarian, and gastrointestinal cancers, especially colorectal and gastric cancers, with the number of microRNAs decreased by four, and these cancers remained prominent on the list. In the second phase of the study, the assessment of microRNAs showed that this specific list was more associated with small intestine, colon, pancreas, and gastric cancers. Detailed results are available in Additional File 6.

The miRPathDB analysis showed that the miRNAs and their target genes were mainly implicated in cancer-related pathways, including chronic myeloid leukemia and prostate cancer, and significant cancer signaling pathways such as p53, TGF-beta, and Ras signaling in both study phases.

Discussion

To summarize, our study aimed to develop a Multi-cancer Detection (MCD) method capable of not only distinguishing between patients with specific cancer types and normal individuals but also effectively differentiating gastric cancer from 12 other cancer types. This approach expands beyond single cancer assays, offering a comprehensive multi-cancer detection assay. We compared our methodology and results with three other recent studies. The first study focuses on multi-cancer early detection (MCED), while the two subsequent studies introduce panels specifically for gastric cancer detection.

• Study 1: The study by Zhang et al. focused on identifying miRNAs for discriminating lung cancer from normal samples and then utilized the same list as an MCED (multi-cancer early detection) panel to distinguish each of the other cancer types from normal samples [34]. While our study employed network biology and WGCNA, they utilized differential gene expression analysis. Despite methodological differences, our model's performance was comparable to theirs for both multi-cancer and gastric-cancer detection. In our study, we achieved a sensitivity of 98% and a specificity of 98% with the MCD panel for gastric cancer, which made it possible to differentiate between patients with gastric cancer and normal samples. However, for other cancer types, the sensitivity and specificity varied from cancer type to cancer type, as shown in Fig. 9.

Fig. 9

Performance comparison between our study and three other studies. (MCDP stands for Multi-cancer Detection Panel, and GCP is the abbreviation for Gastric Cancer Detection Panel.)

• Study 2: Abe et al. aimed to identify serum miRNAs distinguishing early gastric cancer (EGC) samples from non-cancer controls [35]. While our study employed network biology, WGCNA, and the Random Forest algorithm, they utilized Fisher’s linear discriminant analysis with a greedy algorithm to establish a diagnostic model. Additionally, they employed DeLong’s test for selecting the best discrimination model. Despite these methodological differences, our model's performance was comparable to theirs for gastric cancer detection.

As their panel is specifically designed to discriminate between gastric cancer samples and normal samples, without consideration for other cancer types, we compared the performance of our MCD panel, designed for general cancer screening, with theirs. Our findings suggest comparable performances between the two panels in discriminating between gastric cancer and normal samples.

Our study achieved a sensitivity of 98% and specificity of 98% using MCD panel for gastric cancer when considering only normal samples, as shown in the following table, In their study's discovery set, they achieved a sensitivity of 98% and a specificity of 97%.

In the second part of our study, we focused on classifying gastric cancer in comparison to various cancer types, while they only evaluated their model’s ability to classify gastric cancer in comparison to Esophageal squamous cell carcinoma(ESCC) and Colorectal cancer (CRC). However, they only compared 10 patients with Stage III locally advanced gastric cancer with 50 ESCC and 50 CRC patients, resulting in imbalanced data. They report The AUCs of the EGC index in Stage III gastric cancer, ESCC, and CRC were 1.00, 0.64, and 0.44. While our second proposed panel demonstrates a sensitivity of 86% and specificity of 100% for discriminating between gastric cancer and esophageal cancer, and a specificity of 74% for colorectal cancer as shown in Fig. 8, it is also noteworthy for its ability to differentiate gastric cancer from pancreatic cancer patients, achieving a specificity of 96%. Our model not only performs better in discriminating between gastric cancer and esophageal and colorectal cancer but also demonstrates proficiency in discriminating gastric cancer from 10 other types of cancer.

To further evaluate the diagnostic potential of our circulating miRNA panels in detecting early-stage gastric cancer, we employed the GSE164174 dataset as a third independent validation source. This dataset provides detailed disease stage information, enabling us to examine miRNA expression across different phases of cancer progression. The results demonstrated that the MCD panel achieved a sensitivity of 96.57 and a specificity of 97.18. The Gastric Cancer panel showed a sensitivity of 86.87 and a specificity of 88.33. These findings indicate that both panels maintain robust performance even in the early stages of cancer. Further details are available in Additional File 7.

• Study 3: So et al. developed a serum miRNA panel for identifying patients with all stages of gastric cancer in a high-risk population from normal samples [36]. Their 12-miRNA panel demonstrated a sensitivity of 87.0% at a specificity of 68.4%, with an AUC of 84%, indicating good discriminatory power. They conducted a cross-reactivity test against other common cancers, namely: lung, breast, colorectal, liver, esophageal, prostate, and bladder cancer. They tested for cross-reactivity against these specified cancer types in a total of 90 patients, where 11 of them were mistakenly classified as having a high-risk score for gastric cancer.

As they suggest, their panel can effectively differentiate between gastric cancer samples and normal samples. Therefore, we can compare the performance of both of our 4-miRNA panels with their 12-miRNA panel. Our MCD panel achieves a sensitivity of 98% and a specificity of 98% in the classification of gastric cancer patients versus normal samples. Our gastric cancer detection panel is also promising with a sensitivity of 89% and a specificity of 90%, especially considering that it is a heterogeneous population that includes multiple cancer types and not just gastric cancer and normal samples.

Panels comparison

The panels introduced in the above studies are summarized in Table 6. The only direct commonality is the presence of hsa-miR-1343 in different forms (5p and 3p) between the Study1 MCED panel and Our GC panel. Other miRNAs do not seem to overlap directly between the studies. Each study has identified distinct miRNAs for their respective panels, reflecting possibly different methodologies, target populations, and cancer types being considered. Differences in sample collection, processing techniques, and statistical analysis methods can also contribute to these inconsistencies.

Table 6 Comparison of miRNAs panels: our study versus three other studies

Upon assessing the multi-cancer screening panel, we observed that several studies in the literature have also documented their significance across a range of cancer types, including breast, ovarian, colorectal, and esophageal squamous cell carcinoma (ESCC). A comprehensive review of these studies can be found in Table 7.

Table 7 Evaluation of the multi-cancer screening panel in literature

In parallel with these established findings, our study concurs, as we identified a significant increase in the expression of miR-8073, miR-614, miR-548-3p, and miR-1258 among individuals with cancer, compared with the non-cancer control group. Figure 10 illustrates this consistent overexpression across the spectrum of cancers examined in this study. In our investigation, we demonstrated that the multi-cancer screening model effectively classifies cancer patients when compared to the non-cancer control group.

Fig. 10

Multi-cancer screening panel Expression values in different cancer types and non-cancer control group. The cancers-types legend is sorted ascending by expression median [55]

Furthermore, the model exhibits varying effectiveness in distinguishing between different cancer types and the non-cancer control group. Glioma, lung cancer, bladder cancer, and gastric cancer exhibit the highest classification performance, with correct classification rates of 1, 1, 0.98, and 0.98, respectively, as illustrated in Fig. 4. Conversely, the weakest classification performance is observed for sarcoma, with a correct classification rate of 0.74.

Regarding the proposed panel for gastric cancer, it became apparent that some studies in existing literature have highlighted their relevance in the context of gastric cancer. Table 8 provides an in-depth overview of these studies for reference.

Table 8 Evaluation of the gastric cancer screening panel in literature

In line with the three previous studies reported in Table 8, we also showed that the downregulation of miR-1228 was associated with gastric cancer progression compared to other cancers and non-cancer controls. Since miR-1228 plays an important role in the stages of angiogenesis and EMT and can inhibit these stages, its reduction in the individual can be one of the first signs of threat in tumor formation. The downregulation of miR-1228 is a promising candidate in gastric cancer patients. Nonetheless, by exploring the correlation between miR-1228 and other miRNAs, a more precise understanding of the onset of gastric cancer can be obtained. As a result, we propose a panel of four miRNAs, including miR-1228, to enhance the accuracy of the predictive model.

Multiple studies presented in Table 8 provide substantial evidence regarding the involvement of miR-1343-3p in Gastric cancer. We also showed that the upregulation of miR-1343-3p was associated with gastric cancer compared to the non-cancer control group. Furthermore, this molecule was more expressed in gastric cancer compared to other cancers; hence, it can differentiate gastric cancer from other cancers.

In our investigation, the downregulation of hsa-miR-6765-5p in gastric cancer compared to other cancers and non-cancer controls was significantly associated with gastric cancer controls, and this molecule was down-expressed in gastric cancer compared to other cancers. Limited evidence was discovered concerning the involvement of hsa-miR-6765-5p in gastric cancer, with only one study addressing its association with venous thrombosis. However, additional research investigating the correlation between hsa-miR-6765-5p and various cancers, particularly gastric cancer, is recommended.

The second panel effectively classifies gastric cancer among other cancers and non-cancer controls. The group of four microRNAs could accomplish this classification with high sensitivity and specificity, while individual microRNAs alone did not exhibit this characteristic. As depicted in Fig. 11, miR-1228-5p and miR-6765-5p exhibited comparatively lower expression levels, whereas miR-6787 and miR-1343-3p demonstrated relatively higher expression levels compared to other cancer types and the control group. The proximity in serum miRNA profile patterns between Lung cancer and gastric cancer is also evident in this illustration.

Fig. 11

Gastric cancer screening panel expression level in different cancer types and non-cancer control group. The cancers-types legend is sorted ascending by expression level median [55]

Moreover, the performance of the gastric cancer model varies across different cancer types. The best classification performance is observed for esophageal cancer, non-cancer controls, sarcoma, and pancreatic cancer, with correct classification rates of 1, 1, 1, and 0.98, respectively (Fig. 8). Conversely, the weakest classification performance is observed for lung cancer, with a correct classification rate of 0.18. This discrepancy could be attributed to the close similarity in serum miRNA expression profiles between lung and gastric cancers. We attempted to construct a distinct weighted miRNA co-expression network just between gastric and lung cancer cases. However, the network criteria were not met, and no relevant modules could be detected. The serum miRNA profiles of these two cancers appear to be closely related. The other reason might be the heightened likelihood of metastasis from the lung to the gastric region in cancer patients.

Conclusions

The multi-cancer screening panel comprises four miRNAs with common expression altered in the thirteen studied cancers. The gastric cancer screening panel can differentiate the blood samples of gastric cancer patients from other cancers and non-cancer control groups. In this context, candidate miRNA panels identified through this in silico study hold potential for patient screening and other therapeutic purposes. The approach utilized for gastric cancer discrimination can also be adapted and generalized to other cancers and diseases.

According to the pathogenesis of gastric cancer, imaging methods like endoscopy have achieved high screening efficiency for advanced cases. Over the past decade, this has shifted the focus toward identifying novel biomarkers for early detection. Given the limited clinical data available in the two primary datasets used in this study, we adopted an alternative approach by validating our miRNA panel using a third independent dataset(Additional File 7). Our findings demonstrate that both miRNA panels exhibit robust diagnostic performance, even in the early stages of at least three cancer types: colorectal cancer, esophageal cancer, and gastric cancer. These results highlight the potential of the panels in early cancer detection.

However, a comprehensive assessment of the panels’ performance across all studied cancer types could not be conducted due to limitations in the datasets. This limitation presents an opportunity for future studies to investigate diagnostic performance based on other clinical information. To facilitate this, the publication of standardized datasets that integrate clinical and genomic data across various platforms is essential and could significantly enhance research in the field of liquid biopsy.It's important to note that our research primarily relies on bioinformatics analysis of publicly available datasets, utilizing the standard platform of 3D-gene microarray. While we have successfully developed and evaluated two models — one for multi-cancer detection and another for gastric cancer detection —the absence of experimental studies under wet laboratory conditions limits our ability to gain a deeper understanding of the predictive role of these miRNAs in carcinogenesis. Furthermore, additional validation of these two models using an independent cohort, ideally employing alternative technical platforms such as qRT-PCR, followed by ROC and survival curve analysis is necessary before considering their potential clinical application. Addressing these aspects will be the focus of our future work, as they are beyond the scope of the current study.