Establishing a miRNA panel for hepatocellular carcinoma screening through a multicenter study

Summary

Most patients with hepatocellular carcinoma (HCC) are often diagnosed at advanced stages, limiting the effectiveness of curative treatments. In a multicenter study involving 522 patients with HCC and liver cirrhosis, predominantly with hepatitis B virus infection, we evaluated plasma microRNAs (miRNAs) as potential liquid biopsy biomarkers for early HCC screening. Significant upregulation of 18 miRNAs in HCC patients was confirmed across three stages. We conducted a meta-analysis integrating our results with existing literature based on pooled effect size, successfully confirming the upregulation of 4 miRNAs in our findings. Employing an explainable machine learning approach, we established a 5-miRNA panel (miR-361-5p+ miR-130a-3p+ miR-27a-3p+ miR-30d-5p+ miR-193a-5p) with alpha-fetoprotein (AFP) for HCC screening. This combined panel demonstrated superior diagnostic performance compared to AFP alone (AUC: 0.924 vs. 0.794; p < 0.001) in distinguishing HCC patients in the testing and validation set, highlighting its potential as a promising minimally invasive screening method for HCC.

Graphical abstract

Highlights

• •The 5-miRNA and AFP20 panel is established as a liquid biopsy biomarker for HCC screening
• •This combined panel demonstrates superior diagnostic performance compared to AFP20 alone
• •It enhances early HCC surveillance accuracy in LC patients at high risk
• •This panel effectively identifies HCC in both HBV-positive and HBV-negative populations

Molecular biology; Cancer

Introduction

Hepatocellular carcinoma (HCC) is the most prevalent form of liver cancer, accounting for over 90% of cases.¹ Up to 80%–90% of newly diagnosed HCC patients have preexisting liver cirrhosis (LC).² Among the etiologies of HCC, chronic hepatitis B virus infection (HBV) is the predominant risk factor, particularly in Asia, where it accounts for more than 50% of total cases.³ Despite being the 3rd leading cause of cancer-related deaths globally, the 5-year survival rate for early-stage HCC can exceed 70% with curative treatments such as liver ablation, resection, and transplantation.^1,4 Poor prognosis of HCC is often due to the lack of reliable and effective screening tools for early diagnosis. Approximately 50% patients are diagnosed at advanced stages, where curative therapies are no longer suitable.¹ Unlike other solid tumors, HCC diagnosis primarily relies on non-invasive imaging techniques rather than tissue biopsies due to the risk of bleeding and tumor seeding. The American Association of the Study of Liver Diseases (AASLD) recommends biannual surveillance for at-risk populations, particularly adult patients with LC, through liver ultrasound and serum alpha-fetoprotein (AFP) assessments, which are the most widely used serological biomarker for HCC globally.⁵ Abnormal results indicated by liver nodule ≥10 mm or elevated AFP levels >20 ng/mL (AFP20) prompt further evaluation through multiphase computed tomography (CT) or magnetic resonance imaging (MRI) to confirm the diagnosis of HCC.

Liquid biopsy has gained significant attention as a minimally invasive tool for early cancer screening and longitudinal disease monitoring. MicroRNAs (miRNAs), a class of small non-coding RNAs with around 22 nucleotides in length, regulate gene expression post-transcriptionally by binding to the 3′ untranslated region of target messenger RNAs (mRNA), leading to mRNA degradation or translational repression. The remarkable stability of miRNAs in blood makes them a promising candidate for liquid biopsy applications over other biomarkers.⁶ Numerous studies have sought to identify circulating miRNA targets for HCC diagnosis, but their results have often been inconsistent. For example, circulating miR-122-5p exhibits inconsistent dysregulation patterns in HCC, with some studies reporting upregulation, while others have identified downregulation or no significant change.^7,8,9,10,11 In this large-scale multicenter study, we confirmed multiple miRNAs identified in previous studies and strengthened their significance through a comprehensive meta-analysis. Additionally, we revealed novel miRNA candidates as potential diagnostic biomarkers for HCC. From these identified targets, we constructed a 5-miRNA panel alongside AFP20 for HCC surveillance in the LC population, applicable to both HBV-positive and HBV-negative patients. This panel also demonstrated robust performance in detecting early-stage HCC and AFP-negative HCC. Our constructed miRNA panel could improve the current surveillance method for HCC as a minimally invasive first-line screening, ultimately enhancing patient prognosis by providing them with a reliable early diagnostic method.

Results

Baseline characteristics of the study participants

A total of 522 non-hemolyzed plasma samples were collected from the Chongqing (CQ) and Beijing (BJ) cohorts. The clinical and demographic data of the study participants were summarized in Table 1. 354 matched patients, adjusted for confounding factors including age, sex, and HBV status, were included in the validation stage 1. Sample matching was not employed in the validation stage 2 to more accurately represent the actual clinical setting, resulting in a higher proportion of males, older individuals and patients with HBV.

Table 1.

Clinical characteristics of the study participants

	Matched samples in validation stage 1				Validation stage 2
	CQ cohort (n = 203)		BJ cohort (n = 151)		Additional validation set (n = 150)
	HCC (n = 94)	LC (n = 109)	HCC (n = 83)	LC (n = 68)	HCC (n = 29)	LC (n = 121)
Sex
Male	71 (76%)	80 (73%)	58 (70%)	54 (79%)	26 (90%)	43 (36%)
Female	23 (24%)	29 (27%)	25 (30%)	14 (21%)	3 (10%)	78 (65%)
Age	55.1 (13.3)	55.0 (11.7)	57.0 (10.8)	54.8 (10.0)	63.4 (13.9)	56.2 (11.5)
HBV
Yes	71 (76%)	85 (78%)	63 (76%)	50 (74%)	24 (83%)	29 (24%)
No	23 (24%)	24 (22%)	20 (24.1%)	18 (26%)	2 (7%)	81 (67%)
NA	0 (0%)	0 (0%)	0 (0%)	0 (0%)	3 (10%)	11 (9%)
AFP
≤20 ng/mL	29 (31%)	97 (89%)	43 (52%)	66 (97%)	10 (35%)	107 (88%)
>20 ng/mL	65 (69%)	12 (11%)	40 (48%)	2 (3%)	15 (52%)	3 (3%)
NA	0 (0%)	0 (0%)	0 (0%)	0 (0%)	4 (13%)	11 (9%)
BCLC
0 + A	28 (30%)	NA	29 (35%)	NA	6 (21%)	NA
B + C + D	65 (70%)	NA	54 (65%)	NA	21 (73%)	NA
NA	3 (2%)	NA	0 (0%)	NA	2 (7%)	NA
Tumor size
≤3 cm	34 (36%)	NA	47 (57%)	NA	9 (31%)	NA
>3 cm	57 (61%)	NA	29 (35%)	NA	17 (59%)	NA
NA	3 (3%)	NA	7 (8.4%)	NA	3 (10%)	NA
Macrovascular invasion
Yes	32 (34%)	NA	30 (36%)	NA	10 (35%)	NA
No	62 (66%)	NA	53 (64%)	NA	18 (62%)	NA
NA	1 (1%)	NA	0 (0%)	NA	1 (3%)	NA
Metastasis
Yes	14 (15%)	NA	8 (9%)	NA	2 (7%)	NA
No	97 (84%)	NA	74 (90%)	NA	26 (90%)	NA
NA	1 (1%)	NA	1 (1%)	NA	1 (3%)	NA

Data are n (%) or mean (SD). HCC, hepatocellular carcinoma; LC, liver cirrhosis; HBV, chronic hepatitis B virus; AFP, alpha-fetoprotein; BCLC, Barcelona Clinic Liver Cancer; SD, standard deviation; NA, not available.

Identification of 18 upregulated miRNAs in HCC patients

The study design was presented in Figure 1A. Among the 2549 miRNAs interrogated in the discovery stage, 188 miRNAs exhibited differential expression between HCC and LC patients (P_adj < 0.05) (Figure 1B). Comparison of our findings with existing miRNA profiling dataset GSE106817¹² resulted in the selection of 16 upregulated miRNAs (Table S1) along with the addition of 2 previously reported miRNAs. During the validation stage 1, we quantified these miRNAs using RT-qPCR in 354 matched samples. The selection of internal control is crucial for robust relative quantification. miR-16-5p^6,13,14,15 consistently exhibited high expression levels across all samples and demonstrated a strong positive correlation with total quantified miRNA expressions (Pearson’s correlation r = 0.82, p < 0.001). Additionally, no significant difference was observed between two sample groups. Thus, miR-16-5p was determined to be a more suitable normalization control compared to the spike-in target. After normalization with miR-16-5p, the upregulation of 18 miRNAs was confirmed in the validation stage 1, including all 16 miRNAs identified in earlier microarray analysis (Figures 1B and S1). All miRNAs were detected in over 92% of samples and showed no association with HBV status, except for miR-122-5p (Table S2). Among all miRNAs, miR-361-5p, miR-130a-3p, and miR-24-3p emerged as the top 3 significant targets (all with P_adj < 10⁻¹⁵) (Table S3). Unsupervised hierarchical clustering demonstrated that the samples were grouped based on miRNA expression profiles, rather than by AFP status or sample cohorts (Figure 1C). This highlighted the potential of these 18 miRNAs targets as biomarkers for distinguishing HCC from LC patients based on their expression profiles.

Figure 1.

Upregulation of 18 miRNAs in HCC patients

(A) Study design of the current study. CQ, Chongqing; BJ, Beijing.

(B) Volcano plot of differentially expressed miRNAs between HCC patients and LC patients in the discovery stage. The 16 validated miRNA targets were annotated.

(C) Unsupervised hierarchical clustering of 18 miRNA targets across 354 matched samples in validation stage 1.

Validating the miRNA targets through meta-analysis

Literature search in PubMed and Web of Science identified a total of 221 publications relevant to circulating miRNAs for HCC diagnosis (Figure S2). After full text assessment, 41 publications met the inclusion criteria. A total of 93 differentially expressed miRNAs have been reported, most of which were identified in single studies only. Among them, 11 miRNAs with available expression data from 3 or more studies were eligible for meta-analysis (Table S4). Among our 18 miRNA targets, 5 of them have been reported as upregulated in HCC.

miR-122-5p and miR-21-5p were the most frequently reported miRNAs among all identified miRNAs, establishing them as the most reliable candidates for meta-analysis. However, significant heterogeneity was identified for both miRNAs, resulting in inconclusive results in initial meta-analyses that included all studies (Figure 2). No publication bias was identified for either miRNA. The Baujat plots indicated that the main source of heterogeneity was attributed to studies using less common normalization control, particularly the study by Xu et al.¹⁶ for miR-122-5p and Guo et al.¹⁷ for miR-21-5p (Figure S3). After categorizing the studies into different subgroups based on the normalization controls, the heterogeneity within each subgroup was markedly reduced for both miRNAs (Figure 2). This refined subgroup analysis yielded consistent upregulation for miR-122-5p (effect size = 0.66, 95% CI: 0.53 to 0.78; p = 1.17 × 10⁻²⁴) and miR-21-5p (effect size = 0.41, 95% CI: 0.04 to 0.78; p = 0.028) across studies using miR-16-5p or spike-in as common controls. By extending this methodology to additional miRNAs, significant upregulation of miR-192-3p and miR-29a-3p was also identified across studies using common controls. Alternatively, 1 upregulated miRNA (miR-193a-5p) and 2 downregulated miRNAs (miR-26a-3p and miR-223-3p) were identified across studies using mixed controls. Overall, the meta-analysis confirmed the upregulation of 4 miRNAs identified in our earlier findings, specifically miR-122-5p, miR-21-5p, miR-29a-3p, and miR-193a-5p, reinforcing the validity of our platform for identifying dysregulated miRNAs in HCC.

Figure 2.

Forest plot of study-wide significant miRNAs in HCC

CI, confidence interval; SMD, standardized mean difference. Upregulation of 4 miRNAs in our findings was validated through a meta-analysis.

Establishment and validation of a combined 5-miRNA and AFP20 panel for HCC diagnosis

Univariate logistic regression was initially performed to determine the diagnostic performance of individual miRNAs in distinguishing HCC from LC. Among the 18 upregulated miRNAs, miR-361-5p demonstrated the highest area under the curve (AUC) of 0.801, while AFP20 demonstrated an AUC of 0.767 (Table S3). The diagnostic performance of several top-ranked miRNAs was comparable to that of AFP20, while most miRNAs generally showed lower diagnostic performance. None of the 18 miRNA targets showed a correlation with AFP levels (Pearson correlation, all with p > 0.05). This lack of correlation suggested that the upregulation of these miRNAs and AFP occurred through independent mechanisms, highlighting the potential use of circulating miRNAs in parallel with AFP for the diagnosis of HCC.

Next, an explainable machine learning strategy was implemented to integrate multiple miRNAs for the diagnosis of HCC. Using 75% of the 354 matched samples as the training set, logistic least absolute shrinkage and selection operator (LASSO) regression analysis suggested an optimal panel consisting of 5 miRNAs (miR-361-5p+ miR-130a-3p+ miR-27a-3p+ miR-30d-5p+ miR-193a-5p) with AFP20. The diagnostic performance of this panel yielded an AUC of 0.873 (95% confidence interval [CI]: 0.831–0.915) in the training set (Figure S4A). The combined panel maintained robust performance in the remaining 25% of matched samples as the testing set, yielding an AUC of 0.879 (95% CI: 0.809–0.950) (Figure 3B). Further evaluation in an additional validation set resulted in an AUC of 0.957 (95% CI: 0.924–0.990) (Figure 3C). In comparison, the diagnostic performance of AFP20 alone was lower than that of the combined panel across all datasets (AUC = 0.747 [95% CI: 0.698–0.795] in the training set; 0.790 [95% CI: 0.708–0.87] in the testing set; 0.786 [95% CI: 0.687–0.886] in the validation set; Figures 3A–3C).

Figure 3.

Diagnostic performance and risk score of combined panel across different sample sets

(A) In the training set.

(B) In the testing set.

(C) In the validation set.

(D) Risk score distribution of combined panel across patients with different HCC stages. The combined panel demonstrated superior diagnostic performance compared to AFP20 alone across all sample sets. miRNA panel: miR-361-5p+ miR-130a-3p+ miR-27a-3p+ miR-30d-5p+ miR-193a-5p. ∗p < 0.05; ∗∗ p < 0.01; ∗∗∗ p < 10⁻³; ∗∗∗∗ p < 10⁻⁴.

The diagnostic performance of the panels was further evaluated across multiple HCC subsets, including early-stage HCC, HBV-associated HCC, HBV-negative HCC, and AFP-negative HCC (Figures S4A–S4D). Integration of miRNA panel with AFP20 had superior diagnostic performance than continuous AFP in the training set (Figure S4E) and consistently outperformed AFP20 alone in all subsets. The diagnostic performance of the combined panel and AFP20 for overall HCC and early-stage HCC was summarized in Table 2. Risk score was calculated for each patient to classify them into high-risk and low-risk groups. A progressive increase in risk score was observed with advancing HCC stages (Figure 3D), while no significant difference in risk score was identified among patients stratified by age, sex, and HBV status (Figures S4F–S4H). Using Youden’s index as the classification cutoff (0.5572) for the combined panel, both the combined panel and AFP20 achieved specificity exceeding 90% for identifying early-stage HCC. Notably, the combined panel exhibited a marked increase in sensitivity compared to AFP20 (63% vs. 50%). This improvement highlighted the potential of integrating the miRNA panel with AFP20 to enhance HCC screening.

Table 2.

Comparison of the diagnostic performance of combined panel and AFP20 to identify patients with HCC or early-stage HCC in the testing and validation set

	Combined panel	AFP20	p-val
Diagnosing overall HCC (67 HCC vs. 157 LC)
Cutoff	0.557	0.500
AUC (95%CI)	0.924 (0.887–0.960)	0.794 (0.734–0.855)	6.20 × 10−8
Sensitivity	0.776	0.627
Specificity	0.911	0.962
Diagnosing early-stage HCC (24 early-stage HCC vs. 157 LC)
AUC (95%CI)	0.876 (0.803–0.949)	0.731 (0.628–0.834)	5.57 × 10−4
Sensitivity	0.625	0.500
Specificity	0.911	0.962

Early-stage HCC was defined as patients with BCLC stage 0 and A. The comparison was performed in the same set of patients. Samples with missing AFP level were excluded. DeLong’s test was used to compare the AUC of combined panel and AFP20. HCC, hepatocellular carcinoma; LC, liver cirrhosis; HBV, chronic hepatitis B virus; AFP, alpha-fetoprotein; BCLC, Barcelona Clinic Liver Cancer; AUC, area under curve.

Associations of miRNAs with clinical parameters and predictions of target genes

We further examined the association of miRNA expressions with various tumor-associated characteristics in HCC patients. Consistent with the impressive performance of our miRNA panel in identifying early-stage HCC, 17 miRNAs were significantly upregulated in early-stage HCC patients (Figure 4A), underscoring their potential as diagnostic biomarkers for early HCC detection. Additionally, 11 miRNAs showed increased expression in patients with larger tumor size (Figure 4B), while 5 miRNAs were elevated in patients with tumor invasion (Figure 4C). Among the 5 miRNAs included in the miRNA panel, the expression of 3 miRNAs (miR-130a-3p, miR-361-5p, and miR-27a-3p) was increased in correlation with advancing HCC stage and tumor size. We utilized 3 bioinformatics platforms, TargetScan 8.0,¹⁸ miRDB,¹⁹ and miRWalk,²⁰ to predict the target genes of the 5 miRNAs and subsequently constructed an interaction network (Figure 4D). KEGG enrichment analysis of the overlapping target genes from 3 platforms revealed significant enrichment in the PI3K-Akt, MAPK, and HCC signaling pathways (Figure 4E).

Figure 4.

Association of miRNAs with clinical parameters

(A–C) Association with (A) different HCC stages, (B) tumor size, and (C) tumor invasion. 3 miRNAs (miR-130a-3p, miR-361-5p, and miR-27a-3p) in the panel demonstrated further upregulation in patients with late-stage HCC and larger tumor size. ns, not significant; ∗P_adj < 0.05; ∗∗ P_adj < 0.01; ∗∗∗ P_adj < 10⁻³; ∗∗∗∗ P_adj < 10⁻⁴.

(D) Interaction network of the 5 miRNAs in the panel and their target genes. Only genes targeted by 3 or more miRNAs are shown.

(E) KEGG enrichment analysis for target genes of the 5-miRNA panel.

Discussion

The current study achieved significant advancements in HCC detection through liquid biopsy by establishing a combined panel of miRNAs with AFP20, enabling early identification of HCC patients within the at-risk population. This approach provides a distinct advantage over current population screening method that relies solely on AFP20 as a blood biomarker, which has limited sensitivity in detecting early-stage HCC and often leads to delayed diagnosis and adverse clinical outcomes. We identified 18 upregulated miRNAs in HCC, with 4 miRNAs further validated through a meta-analysis. From these targets, we developed a combined panel of 5-miRNA with AFP20, which could effectively distinguish HCC patients from LC patients, including the detection of early-stage HCC. This 5-miRNA panel can also function independently for HCC screening in patients without available AFP levels. The current study represented a large-scale, high-throughput, and multicenter analysis of circulating miRNAs for HCC diagnosis. Our findings highlighted the potential of circulating miRNA as minimally invasive diagnostic biomarkers for the early detection of cancer. It is not our intention to replace existing imaging-based diagnostic methods for HCC, but to improve HCC surveillance by integrating circulating miRNAs with AFP as a first-line screening method.

As a liver-specific miRNA,²¹ dysregulation of circulating miR-122-5p in HCC has been widely reported with contradictory findings,^7,8,9,10,11 primarily due to the use of different normalization control as indicated by our meta-analysis. The dysregulation of circulating miR-122-5p is not exclusive to HCC; it has also been observed in other liver chronic diseases, such as HBV, chronic hepatitis C infection, nonalcoholic fatty liver disease, and drug-induced liver injury.^22,23,24,25 In contrast, downregulation of liver miR-122-5p is consistently observed in HCC.²¹ Our findings revealed an intriguing inverse relationship between circulating and liver miR-122-5p levels. The upregulation of miR-122 showed minimal or no significant correlation with the liver enzymes aspartate aminotransferase (AST) and alanine aminotransferase (ALT) (r = 0.10, p = 0.03 for AST; r = 0.03, p > 0.05 for ALT), suggesting that its release into the circulation is unlikely to be driven solely by hepatocyte cell death. A similar inverse relationship has been noted in non-alcoholic steatohepatitis (NASH), where Pirola et al. also suggested that lower expression of liver miR-122-5p may result from its release into the circulation rather than a decrease in overall expression.²⁶ Although the upregulation of key miRNAs in our panel, miR-361-5p, miR-130a-3p, and miR-27a-3p, have rarely been reported in HCC, their elevated expressions were consistent with the expression data from Yokoi et al. (Table S1).^12,27 Notably, obesity-induced upregulation of serum miR-27a-3p in pregnant mice has been shown to intergenerationally increase hepatic miR-27a-3p levels and HCC susceptibility in their offspring,²⁸ highlighting the potential role of circulating miR-27a-3p in promoting HCC.

The limited overlap of miRNA targets from previous studies may arise from variations in study design, including differences in biofluid selection, at-risk populations and normalization controls. Currently, there is no established internal normalization control for the relative quantification of circulating miRNAs in HCC. We selected miR-16-5p as the internal control due to its relatively stable expression across our samples. This was further validated by a meta-analysis, which showed consistent results across studies employing similar normalization controls. However, some studies reported downregulation of miR-16-5p in HCC patients,^9,29 suggesting that the selection of a consensus normalization control needs further exploration. We initiated our study with a high-throughput screening and confirmed the results in multiple samples sets using RT-qPCR, where the results of RT-qPCR could be easily transformed into a clinical in vitro diagnostic test.

This study systematically identifies 18 upregulated miRNAs in HCC patients and strengthens these findings through a comprehensive meta-analysis. The key success of this study is the establishment of a 5-miRNA panel with AFP to distinguish HCC patients from the at-risk population. The panel demonstrates robust diagnostic performance across multiple HCC subsets, including early-stage HCC, HBV-associated HCC, and AFP-negative HCC, highlighting its promise as a reliable tool for population-based screening. Our minimally invasive combined panel can enhance the early detection of HCC, thereby providing the patients with increased treatment options and improving their clinical outcomes.

Limitations of the study

As circulating miRNAs can be released from various cell types throughout the body, it is challenging to draw functional conclusions without knowing the origin of the detected circulating miRNAs. Additionally, different etiologies of LC drive the development of HCC through distinct pathological pathways. As the current study excluded patients with other viral infections and recruited participants solely from clinical centers in China, validation in other at-risk populations and ethnic groups is recommended before generalizing the current findings to other populations. Caution must be taken when comparing results across studies with different study designs.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Mengsu Yang (bhmyang@cityu.edu.hk).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •The MicroRNA microarray data have been deposited in GEO database with accession number GSE294055.
• •This paper does not report original code.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact on request.

Acknowledgments

Part of the work described in this paper was conducted by Dr. Zihan Yang, Jockey Club Global STEM postdoctoral fellow supported by The Hong Kong Jockey Club Charities Trust. This work was supported by grants from Hetao Shenzhen-Hong Kong Science and Technology Innovation Cooperation Zone Shenzhen Park project (HZQB-KCZYZ-2021017).

Author contributions

K.K.K.N.: conceptualization, data curation, formal analysis, validation, investigation, methodology, visualization, writing – original draft; Y.L.: resources, writing – review and editing; J.Q.: investigation; Z.Y.: investigation; X.Z.: resources.; H.X.: data curation, clinical data collection, resources; Z.Z.: conceptualization, data curation, clinical data collection, validation, investigation, resources, methodology, supervision, writing – original draft, writing – review and editing; M.Y.: resources, supervision, funding acquisition, writing – review and editing.

Declaration of interests

K.K.K.N. and M.Y. have filed patents/patent applications based on the data generated from this work.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Critical commercial assays
mirVana miRNA Isolation Kit	Thermo Scientific	Cat #AM1560
miRNeasy Serum/Plasma Kit	Qiagen	Cat #217184
miRCURY LNA RT Kit	Qiagen	Cat #339340
miRCURY LNA miRNA Probe PCR Assays	Qiagen	Cat #339350
Deposited data
MicroRNA microarray data	This paper	GEO: GSE294055
Software and algorithms
R software (version 4.4.1)	R Foundation for Statistical Computing	https://www.R-project.org

Experimental model and study participant details

Study cohorts

In this multicenter cohort study, plasma samples were collected from a total of 522 patients (male:female = 344:178; median age: 56 years old, range: 17–86) across three stages: discovery stage, validation stage 1 and validation stage 2. In the discovery stage, microarray analysis was performed on 90 samples (50 patients with HCC and 40 patients with LC) to identify differentially expressed miRNAs between HCC patients and LC patients. In the validation stage 1, significant miRNAs from the discovery stage were selected and validated with reverse transcription quantitative real-time PCR (RT-qPCR) in a total of 354 matched samples, comprising 72 samples from the discovery stage and an additional 282 samples. Blood samples for the discovery stage and validation stage 1 were collected from a Chongqing (CQ) cohort from the Second Affiliated Hospital of Chongqing Medical University (N = 227) and a Beijing (BJ) cohort from Chinese PLA General Hospital (N = 151) from November 2020 to April 2022. An additional 150 samples were recruited from the Second Affiliated Hospital of Chongqing Medical University from September 2022 to December 2022 as validation stage 2. Most patients had a history of chronic hepatitis B virus infection (HBV). HCC patients were diagnosed through computed tomography (CT) or magnetic resonance imaging (MRI), part of which were further confirmed histologically by at least two independent histopathologists according to the American Association of the Study of Liver Diseases (AASLD) guidelines.⁵ Tumor staging was classified according to the Barcelona Clinic Liver Cancer (BCLC) staging system,³⁰ with stage 0 and stage categorized as early-stage HCC and the remaining stages categorized as late-stage HCC. None of the patients with HCC had previously undergone surgical operation, chemotherapy or radiotherapy before blood collection. Patients with LC were diagnosed with ultrasound and included as at-risk controls. Chronic HBV infection was defined as chronic liver disease caused by persistent HBV infection (positive HBsAg >6 months with detectable serum HBV DNA). Clinical characteristics were collected for subsequent analysis. Informed consent was obtained from all patients. The study was approved by the Hospital Ethical Committee of the Second Affiliated Hospital of Chongqing Medical University (2022-80) and adhered to the principles of the Helsinki Declaration.

Method details

Blood collection and RNA extraction

Peripheral blood samples were collected from all subjects using 4-mL BD Vacutainer EDTA tubes (Becton Dickinson) and stored at 4°C. Plasmas were isolated within 8 h from collection and centrifuged at 4,000 rpm for 15 min at 4°C. Plasma samples were aliquoted in sterilized Eppendorf tubes and stored at −80°C until RNA extraction. The absorbance at 414 nm (A414) was measured for all plasma samples using NanoDrop ND-2000 (Thermo Scientific) to evaluate the degree of hemolysis.³¹ Plasma samples with A414 value greater than 0.2 were removed from downstream analysis. For microarray analysis, miRNA was extracted from 400 μL plasma using mirVana miRNA Isolation Kit (Thermo Scientific) following the manufacturer’s instructions. Total RNA was quantified by NanoDrop ND-2000 (Thermo Scientific) and RNA integrity was assessed by Agilent Bioanalyzer 2100 (Agilent Technologies). For qPCR analysis, miRNA was extracted from 200 μL plasma using miRNeasy Serum/Plasma Kit (Qiagen). The extraction was performed according to the manufacturer instructions with minor modifications: 1) to normalize the extraction variability, a spike in control was added to the sample during the RNA isolation phase; 2) to improve RNA isolation efficiency, bacteriophage MS2 carrier RNA (Thermo Scientific) was added to the sample during the RNA isolation phase (1 μg per mL of QIAzol). Extracted RNAs were stored at −80°C until further use.

Microarray: Discovery stage

The Agilent Human miRNA Microarray Kit, Release 21.0, 8 × 60K (DesignID: 070156) experiment and data analysis were conducted by OE Biotechnology Co., Ltd. (Shanghai, China), according to the Agilent miRNA Microarray System with miRNA Complete Labeling and Hyb Kit protocol (Agilent Technologies). The slides were scanned with the Agilent scanner G2505C (Agilent Technologies). Raw data was extracted using Feature Extraction software (version 10.7.1.1, Agilent Technologies). Only miRNAs with detected signal in at least 50% of any sample group were included in further data analysis. The included data was normalized using the quantile normalization.

RT-qPCR: Validation stage 1 and 2

3 μL of extracted RNA was reverse transcribed using the miRCURY LNA RT Kit (Qiagen) in 10 μL reactions. cDNA was diluted 20× before qPCR. qPCR was performed using miRCURY LNA miRNA Probe PCR Assays (Qiagen). The amplification conditions consisted of a heat activation step at 95°C for 2 min, followed by 40 cycles of 5 s at 95°C and 30 s at 56°C. qPCR reactions were run on the QuantStudio 7 Pro Real-Time PCR System (Applied Biosystems). A cycle threshold (Ct) value of 40 was imputed for the undetected reactions. miR-16-5p was used as an endogenous control for the normalization,^6,13,14,15 and relative quantifications of miRNAs were calculated by 2^−ΔCt, where ΔCt = Ct_target–Ct_control. To ensure consistent quantifications throughout all reactions, three interplate controls were included in each PCR reaction to account for plate-to-plate variation. The results of each reaction were normalized against the interplate controls.

Literature search for meta-analysis

A systematic literature search was conducted in PubMed and Web of Science up to Dec 14, 2023, to identify studies on circulating miRNAs for HCC diagnosis. The search utilized a combination of keywords and medical subject headings including “hepatocellular carcinoma”, “circulating microRNA”, “diagnosis” and “HBV”. Additional manual search was performed using references from retrieved articles and relevant reviews. Only articles published in English and peer-reviewed journals were considered. After removing duplicate records, the abstracts were screened to identify relevant articles. Studies were included based on the following criteria: (1) original studies; (2) clinical studies evaluating circulating miRNAs for the diagnosis of HCC; (3) miRNA profiling studies using serum or plasma as specimens. Following a thorough full-text review, studies were excluded if they did not meet the following criteria: (1) not a case-control study; (2) did not employ RT-qPCR as the quantification method; (3) included patients with viral co-infection or viral infection other than HBV as control. This study was conducted according to the PRISMA statement.

Data extraction

Data extracted from eligible studies included first author, year of publication, region, specimen type, sample size, sample characteristics, study design, normalization control, examined miRNAs, types of dysregulations, effect sizes and p-values. For studies that included comparisons against multiple control groups, effect size data versus the control group with higher risk of HCC was extracted (HBV-LC > LC > HBV). Both significant and non-significant results were included in the meta-analysis. If the effect size data was not directly reported, expression data was obtained from the graphical plot using PlotDigitizer (https://plotdigitizer.com/). The following conservative conversions were applied to p values reported in a predefined significance threshold: p ≥ 0.05 and p ≥ 0.01 were converted to p = 0.5, p < 0.05 to p = 0.025, p < 0.01 to p = 0.005, p < 0.001 to p = 0.0005, and p < 0.0001 or p = 0.0000 to p = 0.00005. All relevant studies were assessed by two independent researchers (K.N., Z.Z.). Any discrepancy was resolved through consultation with the lead researcher (M.Y.).

Quantification and statistical analysis

Statistical analysis

miRNA expressions from RT-qPCR were standardized by Z score normalization. Differential expression analyses were performed using either a two-sided Student’s t test or Wilcoxon rank-sum test. A p-value of <0.05 and fold change >|2| were considered as statistically significant. P-values were adjusted for multiple hypothesis testing using the Benjamini-Hochberg method. Propensity score matching (PSM) was employed to select matched samples in validation stage 1 based on sex, age and HBV status as key confounding factors, resulting in 177 matched pairs of HCC and LC patients. To ensure that the above sample size had sufficient power to identify differential miRNAs expressions, a statistical power analysis was performed using the software G∗Power.³² Considering the average effect sizes of significant miRNAs from validation stage 1, we set the effect size Cohen’s D of miRNAs to be 0.69. Our sample size of 177 pairs exceeded the minimum requirement of 47 paired samples for identifying differential miRNA expressions under a power of 80% and Type I error of 5%.

Effect sizes and corresponding measures of variance (p value, 95% confidence intervals [95% CIs], or standard errors) were extracted from included studies. If means and standard deviations were not provided, they were calculated from the corresponding measure of variance or estimated from the median and interquartile range.³³ Whenever the effect sizes of a miRNA were available from three or more studies, Hedges’ g was calculated as pooled standardized mean differences (SMD) between HCC and LC patients for the meta-analysis. Heterogeneity among the studies was assessed using Cochran’s Q test and Higgin’s inconsistency index (I²). A p value of less than 0.05 in the Cochran’s Q test indicated significant heterogeneity. The DerSimonian-Laird random-effects model was applied for miRNAs with significant heterogeneity³⁴; otherwise, a fixed-effects model was used. Meta-analyses and heterogeneity tests were performed in R with the package “metafor” (version 4.2.0). The SMD and 95% CIs from the meta-analyses were presented with forest plots. Baujat plot and influence analysis were employed to identify studies with extreme effect sizes within the meta-analysis. Subgroup analyses were performed based on differences in study design. Normalization controls frequently used across multiple studies were classified as common controls, while RNU6 was classified as a distinct group and other controls were classified as less common controls. Potential publication bias was examined by visual inspection of the funnel plot and Egger’s test. Unless otherwise specified, a two-sided p value of less than 0.05 was considered statistically significant. Significance in the meta-analysis was adjusted for multiple testing using Bonferroni correction (i.e., α = 0.05/11 = 4.55 × 10⁻³).

Univariate logistic regression was performed to evaluate the performance of individual miRNAs and AFP20. Matched samples in validation stage 1 were randomly split into a training set (75%) and testing set (25%). Following 100 iterations of repeated 5-fold cross-validation, LASSO regression analysis was performed to establish an optimal miRNA panel with AFP20 in the training set using the caret package. The miRNA set with the highest AUC was selected as the final panel for further validation in the testing set and additional validation set. Multicollinearity among variables in the panel was examined by variance inflation factor, with a score of <5 indicating acceptable multicollinearity. Risk scores were calculated for all patients, defined as the sum of each miRNA expression weighted by its corresponding logistic regression coefficient in the model. The risk score cutoff was determined by Youden’s index to differentiate patients into high-risk and low-risk groups. Subgroup analyses were performed across multiple sample subsets to evaluate the robustness and the diagnostic performance of the established miRNA panel. Diagnostic performance, sensitivity and specificity of the models were evaluated using AUC and receiver operating characteristics (ROC) analysis. PSM was performed using the matchit function in the MatchIt package. Statistical comparisons of ROC curves were performed using Delong’s test in the pROC package. Statistical analyses were performed using R, version 4.4.1.

miRNA target genes prediction and network analysis

Target genes of miRNAs were predicted using 3 platforms, TargetScan8.0,¹⁸ miRDB¹⁹ and miRWalk.²⁰ Functional enrichment of mutually predicted target genes was performed using the Kyoto Gene and Genome Encyclopedia (KEGG) database. Predicted targets from 3 or more miRNAs were used to construct the interaction networks using Cytoscape v3.10.2.³⁵

Published: June 23, 2025