A blood-based noninvasive miRNA signature for predicting survival outcomes in patients with intrahepatic cholangiocarcinoma

Abstract

Background

The prognosis in patients with intrahepatic cholangiocarcinoma (ICC) is generally poor. To improve treatment selection, we sought to identify microRNA (miRNA) signature associated with survival outcomes in ICC.

Methods

We first analysed the miRNA expression profiles of primary ICC from two public datasets to identify a miRNA panel to detect patients for short-term survival. We then analysed 309 specimens, including 241 FFPE samples from two clinical cohorts (training: n = 177; validation: n = 64) and matched plasma samples (n = 68), and developed a risk-stratification model incorporating the panel and CA 19-9 levels to predict survival outcomes in ICC.

Results

We identified a 7-miRNA panel that robustly classified patients with poor outcomes in the discovery cohorts (AUC = 0.80 and 0.88, respectively). We subsequently trained this miRNA panel in a clinical cohort (AUC = 0.83) and evaluated its performance in an independent validation cohort (AUC = 0.82) and plasma samples from the additional validation cohort (AUC = 0.78). Patients in both clinical cohorts who were classified as high-risk had significantly worse prognosis (p < 0.01). The risk-stratification model demonstrated superior performance compared to models (AUC = 0.85).

Conclusions

We established a novel miRNA signature that could robustly predict survival outcomes in resected tissues and liquid biopsies to improve the clinical management of patients with ICC.

Introduction

Liver cancer is the sixth most commonly diagnosed cancer and the fourth leading cause of cancer-related deaths globally, with approximately 841,000 new cases and 782,000 deaths annually [1]. Primary liver cancers include hepatocellular carcinoma (75–85% of cases) and intrahepatic cholangiocarcinoma (ICC; 10–15% of cases), as well as other rare types [1–3]. According to current National Comprehensive Cancer Network (NCCN) guidelines, surgical resection with histologically negative margin remains the only potentially curative treatment for patients with ICC [4]. Unfortunately, 5‐year overall survival (OS) rates following curative resection in ICC patients are only 20–30% [5–7], primarily due to the high incidence of tumour recurrence, with rates ranging from 50 to 70% [8–11]. While cancer recurrence occurs frequently, the perioperative treatment options remain arbitrary. For better treatment outcomes, perioperative chemotherapy may benefit these patients by improving tumour resectability and survival outcomes, as has been observed in other malignancies, including pancreatic cancer and rectal cancer; nevertheless, there are no established perioperative treatment options in the NCCN guidelines. Furthermore, due to the rarity of this malignancy, there are no well-defined randomised or prospective studies comparing survival outcomes; highlighting that the prediction of survival outcomes is an important clinical challenge and it may have a significant potential impact for more improved treatment selection of patients with curative resection.

Developing accurate models for the preoperative risk stratification of patients with surgically resectable ICC has important clinical significance, because the resection of ICC often necessitates a major hepatectomy and is associated with a higher risk of positive surgical margins and major post-operative complications than other indications (e.g., liver metastasis) [12, 13]. Staging systems and prognostic models have been used in several malignancies to facilitate treatment decisions and provide guidance based on anticipated long-term outcomes [14, 15]. However, there are few established models to predict survival outcomes for patients with resectable ICC, and even fewer have been used clinically. Furthermore, most models to date have used variables that can only be assessed post-operatively and thus cannot be used to help inform treatment selection [16–20]. Therefore, there is a critical need to develop clinically feasible, predictive models that incorporate preoperative variables to identify patients who are most likely to benefit from surgical resection. Accumulating evidence indicates that the expression patterns of microRNAs (miRNAs) can be used to determine the physiological and pathological status of cancer patients. In addition to demonstrating that the differential expression of specific miRNAs contributes directly to cancer pathogenesis, several studies have emphasised the potential of miRNAs as circulating biomarkers [21–23]. Indeed, our group has also previously reported multiple blood-based miRNA signatures that enabled robust and specific detection of colorectal cancer recurrence and metastasis [24–26].

In the present study, we performed a systematic, genome-wide analysis to comprehensively discover biomarkers and identify a miRNA expression signature to predict survival outcomes following hepatectomy in patients with ICC. Following the identification of a miRNA signature from expression profiling datasets, we rigorously validated and evaluated its performance from resected tissue specimens in large, independent clinical cohorts. Finally, we evaluated the feasibility of translating our biomarkers into a noninvasive, blood-based assay by analysing blood specimens from patients with ICC. Our final risk-stratification model robustly identifies patients with high-risk ICC and can be used to improve overall patient management.

Methods

miRNA biomarker discovery

To comprehensively discover and establish a panel of miRNA biomarkers of poor survival outcomes following hepatectomy in patients with ICC, we analysed miRNA expression profiles from two public datasets: The Cancer Genome Atlas (TCGA) and GSE53870. In total, miRNA expression profiling data were obtained from 53 patients, including 15 patients from the TCGA cohort (5 with short-term survival [cancer-related death within 1 year of hepatectomy] and 10 with long-term survival [surviving more than 2 years after hepatectomy]) and 38 patients from the GSE53870 dataset (19 with short-term survival and 19 with long-term survival), as illustrated in Fig. 1a.

Fig. 1. Genome-wide discovery and validation of a novel 7-miRNA panel to detect patients with ICC at high risk for short-term survival.

a Study overview. b, c ROC curves for the performance of the 7-miRNA panel in distinguishing patients at high risk for short-term survival in the (b) GSE53870 (with short-term survival [cancer-related death within 1 year of hepatectomy] = 19, long-term survival [surviving more than 2 years after hepatectomy] = 19, AUC = 0.80) and c TCGA cohorts (with short-term survival = 5, long-term survival = 10, AUC = 0.88). d, e Comparisons of OS in high- versus low-risk groups, determined based on the 7-miRNA panel, in the d GSE53870 and e TCGA cohorts (P < 0.01). ROC curves and AUC values were used to evaluate the performance of the panel in detecting patients with short-term versus long-term survival. OS times were calculated from the date of surgery to the date of death from any cause, or last follow-up date. We estimated OS per group using the Kaplan–Meier method.

Patient cohorts

To train and validate our miRNA biomarker panel, we analysed 309 clinical specimens, including 241 formalin-fixed paraffin-embedded (FFPE) and 68 plasma samples from three independent ICC patient cohorts. The training cohort (n = 177) included patients treated at Tokushima University and Kumamoto University, and the validation cohort (n = 64) included patients treated at Kyushu University. We also used matched preoperative plasma samples in a subset of patients (n = 48) from the training cohort as an additional validation cohort. Only regions confirmed by imaging were diagnosed as sites of tumour recurrence. The study was conducted in accordance with the Declaration of Helsinki. A written informed consent was obtained from all patients, and the study was approved by the institutional review boards of all participating institutions.

Statistical analysis

The clinicopathologic characteristics of the training and validation patient cohorts are shown in Table 1. The detail was described in Supplementary method.

Table 1.

Clinicopathological characteristics of clinical cohorts.

Characteristics		Training (n = 177), n (%)	Validation (n = 64), n (%)	P*	Additional validation (n = 48), n (%)
Age (years)	Median (range)	71 (37–96)	67 (39–87)	0.11	70 (37–84)
Gender	Male	107 (60.5)	43 (67.2)		33 (68.8)
	Female	70 (39.5)	21 (32.8)	0.34	15 (31.2)
Survival outcome	Cancer-related death	88 (49.7)	29 (45.3)		27 (56.3)
	Survival	89 (50.3)	35 (54.7)	0.55	21 (43.6)
Hepatitis virus infection	Positive	44 (24.9)	13 (20.3)		13 (27.1)
	Negative	133 (75.1)	51 (79.7)	0.46	35 (72.9)
Tumour size (mm)	≥35	90 (50.9)	39 (60.9)		30 (62.5)
	<35	87 (49.1)	25 (39.1)	0.18	18 (37.5)
CEA (ng/ml)	≥ 5.0	58 (32.8)	29 (45.3)		18 (37.5)
	<5.0	119 (67.2)	35 (54.7)	0.08	30 (62.5)
CA19-9 (U/ml)	≥37	87 (49.1)	32 (50.0)		32 (66.7)
	<37	90 (50.9)	32 (50.0)	0.91	16 (33.3)
Microvascular invasion	Positive	90 (50.9)	29 (45.3)		19 (39.6)
	Negative	87 (49.1)	35 (54.7)	0.42	29 (60.4)
Macro type	MF	120 (67.8)	51 (79.7)		22 (45.8)
	PI	57 (32.2)	13 (14.0)	0.07	26 (54.2)
Adjuvant chemotherapy	Yes	51 (28.8)	26 (40.6)		12 (25.0)
	No	123 (69.5)	38 (59.4)		36 (75.0)
	Missing	3 (1.7)	0 (0)	0.10	0 (0)
Recurrence status	Recurrence	109 (61.6)	37 (57.8)		37 (77.1)
	Non-recurrence	68 (38.4)	27 (42.2)	0.60	11 (22.9)
Differentiation	Well	62 (35.0)	32 (50.0)		14 (29.2)
	Moderately	79 (44.6)	28 (43.8)		19 (39.6)
	Poor	24 (13.6)	4 (6.2)		7 (14.6)
	Others	12 (6.8)	0 (0)	0.16	8 (16.6)

CA19-9 Carbohydrate antigen 19-9, CEA Carcinoembryonic antigen, MF mass forming type, PI periductal-infiltrative type.

*P-values show the difference for training vs. validation cohort using the Chi-Square test or Mann–Whitney U test.

Results

Genome-wide miRNA expression profiling identified a novel 7-miRNA panel for the identification of high-risk patients with ICC

We performed an unbiased, comprehensive genome-wide analysis of two independent miRNA expression profiling datasets (TCGA and GSE53870) to identify a miRNA signature for the detection of survival outcomes in patients with ICC (Fig. 1a). We first compared the miRNA expression profiles of patients with short-term survival (cancer-related death within 1 year of hepatectomy) and patients with long-term survival (alive more than 2 years after hepatectomy) in the GSE53870 cohort. This analysis identified a panel of differentially expressed miRNA biomarker candidates (P < 0.05, |log₂FC| > 1). Subsequent validation of these candidates in the TCGA dataset led us to identify a panel of 7 differentially expressed miRNAs, including miR-10b-3p, miR-26b-3p, miR-27a-3p, miR-106b-3p, miR-219a-3p, miR-338-5p, and miR-421. A Cox’s proportional hazard regression model incorporating these 7 miRNAs to identify patients at high risk (for short-term survival) resulted in an AUC of 0.80 in the GSE53870 cohort (95% confidence interval [CI] = 0.64–0.91; Fig. 1b, Table. S1). Furthermore, the diagnostic performance of the 7-miRNA signature was validated in the TCGA dataset (AUC = 0.88; 95% CI = 0.61–0.99; Fig. 1c). To evaluate the prognostic potential of our miRNA biomarkers, we performed survival analysis to compare the OS of high- versus low-risk patients. High-risk ICC patients in both cohorts demonstrated significantly worse OS (P < 0.01; Fig. 1d, e), highlighting the clinical significance of our biomarker panel.

Successful validation of the miRNA panel in a clinical training and validation cohort for detecting patients with ICC with poor survival outcomes after hepatectomy

Next, we aimed to validate our 7-miRNA biomarker panel using clinical cohorts. To this end, we first confirmed that the patients within our training and validation cohorts possessed similar clinicopathologic characteristics. The training cohort comprised 177 patients, including 88 patients with cancer-related death (49.7%); the median age of patients in this cohort was 71 years. The validation cohort comprised 64 patients, including 29 patients with cancer-related death (45.3%); the median age of patients in this cohort was 67 years. We did not observe any statistically significant differences in the survival outcomes, or any other clinicopathological characteristics, of the two cohorts, limiting the risk of inadvertent bias (Table 1).

To confirm the diagnostic robustness of our panel, we systematically interrogated its ability to predict short-term survival after hepatectomy in ICC patients in the training cohort. Using Cox’s proportional hazard regression analysis, we developed the miRNA panel with the following weights assigned to each biomarker: Risk score = (0.08018*miR-10b-3p) + (−0.1054 *miR-26b-3p) + (0.3543*miR-27a-3p) + (−0.3636*miR-106b-3p) + (−0.2799*miR-219-3p) + (−03098*miR-338-5p) + (−0.4045*miR-421). This model robustly identified high-risk patients with ICC in our training cohort (AUC = 0.83, 95% CI = 0.74–0.90, Fig. 2a). Encouraged by these results, we evaluated the robustness and accuracy of our statistical model by applying it to an independent validation cohort. We observed that the diagnostic performance of our model in the validation cohort was comparable to that observed in the training cohort (AUC = 0.82, 95% CI = 0.64–0.94; Fig. 2b).

Fig. 2. Training and validation of the panel for identifying survival outcomes in patients with ICC.

a, b ROC curves for the performance of the miRNA panel in tissue specimens from patients in the a training (n = 177, AUC = 0.83) and b validation cohorts (n = 64, AUC = 0.82). c–f Comparisons of c, e OS and d, f RFS in high- versus low-risk groups, determined based on the panel, in the c, d training and e, f validation cohorts. ROC curves and AUC values were used to evaluate the performance of the panel in detecting patients with short-term versus long-term survival. OS and RFS times were calculated from the date of surgery to the date of death from any cause or recurrence, or last follow-up date. We estimated OS and RFS per group using the Kaplan–Meier method.

The miRNA panel exhibited robust prognostic potential for predicting survival outcomes in patients with ICC

To evaluate the prognostic potential of our miRNA panel, we performed survival analysis to compare OS and RFS between patients categorised as high-risk versus low-risk. The median follow-up times were 26.3 months (95% CI = 23.2–44.9) in the training cohort and 26.1 months (95% CI = 24.7–54.4) in the validation cohort. Importantly, we observed that high-risk patients had a significantly worse prognosis than low-risk patients in the training cohort (OS: P < 0.01; RFS: P < 0.01; Fig. 2c, d) and the validation cohort (OS: P = 0.02; RFS: P = 0.04; Fig. 2e, f). Furthermore, we conducted a multivariate Cox’s proportional hazard regression analysis to establish a risk-stratification model incorporating our miRNA panel with other clinicopathological factors, and patients categorised as high-risk by the miRNA panel had a significantly worse OS than low-risk patients in both independent cohorts (training cohort: hazard ratio [HR] = 3.50, 95% CI = 2.19–5.59, P < 0.01; validation cohort: HR = 2.23, 95% CI = 1.02–5.02, P = 0.04; Table 2]. These results indicate that, in addition to its predictive ability to detect short-term survival in patients with ICC, our miRNA panel has significant prognostic potential as well.

Table 2.

Univariate and multivariate Cox’s proportional hazard regression analysis to detect short-term (high risk) versus long-term (low risk) survival.

Factors	Univariate analysis			Multivariate analysis
	HR	95% CI	p	HR	95% CI	p
Training cohort (n = 177)
Age
(≥70 vs. <70)	1.18	0.77–1.80	0.44
Gender
(Male vs. Female)	1.28	0.83–1.97	0.26
Hepatitis virus infection
(Positive vs. Negative)	0.57	0.34–0.96	0.03	0.66	0.39–1.12	0.12
Tumour size
(≥35 mm vs. 35 mm)	2.05	1.34–3.15	<0.01	1.57	1.00–2.45	0.05
CEA
(≥5.0 ng/mL vs. <5.0 ng/mL)	1.50	0.96–2.35	0.07
CA19-9
(≥37 U/mL) vs. <37 U/mL)	2.53	1.63–3.91	<0.01	2.19	1.40–3.43	<0.01
Microvascular invasion
(Positive vs. Negative)	2.47	1.60–3.82	<0.01	2.19	1.40–3.44	<0.01
Adjuvant chemotherapy
(Yes vs. No)	1.51	0.96–2.37	0.07
miRNA panel
(High risk vs. Low risk)	3.48	2.18–5.54	<0.01	3.50	2.19–5.59	<0.01
Validation cohort (n = 64)
Age
(≥70 vs. <70)	2.08	0.98–4.40	0.06
Gender
(Male vs. Female)	0.87	0.40–1.88	0.72
Hepatitis virus infection
(Positive vs. Negative)	0.51	0.18–1.48	0.21
Tumour size
(≥35 mm vs. 35 mm)	1.12	0.52–2.42	0.77
CEA
(≥5.0 ng/mL vs. <5.0 ng/mL)	2.30	1.08–4.88	0.03	2.04	0.95–4.41	0.07
CA19-9
(≥37 U/mL) vs. <37 U/mL)	1.74	0.84–3.64	0.14
Microvascular invasion
(Positive vs. Negative)	1.37	0.65–2.87	0.41
Adjuvant chemotherapy
(Positive vs. Negative)	0.61	0.28–1.34	0.22
miRNA panel
(High risk vs. Low risk)	2.53	1.15–5.59	0.02	2.23	1.02–5.02	0.04

Bold values indicate statistical significance p < 0.05.

CA19-9 Carbohydrate antigen 19-9, CEA Carcinoembryonic antigen, CI confidence interval, HR hazard ratio.

Translation of the miRNA panel into a noninvasive assay for preoperative risk-assessment in patients with ICC

An ideal clinical application of our miRNA biomarker signature would be in a noninvasive, liquid biopsy diagnostic platform. Such an assay would obviate the need to analyse tissue specimens, which are not generally available from most patients in the preoperative setting. Therefore, in this study, we focused on translating our tissue-based biomarkers into a blood-based assay. Using matched preoperative plasma samples from a subset of 48 patients with ICC from the training cohorts, we confirmed the robustness and accuracy of our model, applying the same statistical coefficients used in our tissue-based assessments to calculate the 7 miRNA-based risk score in the additional validation cohort. The diagnostic accuracy was consistent with our findings using tissue specimens (AUC = 0.78, 95% CI = 0.59–0.91, Fig. 3a, b), and ICC patients in the high-risk group had a significantly worse prognosis than patients in the low-risk group (OS: P < 0.01; RFS: P < 0.01; Fig. 3c, d). In addition, high-risk patients categorised using a multivariate Cox’s proportional hazard regression model exhibited significantly worse OS vs. low-risk patients (HR = 4.92, 95% CI = 1.64–14.78, P < 0.01; Table. S2). Finally, we used matched pre- and post-operative plasma samples (n = 20) to evaluate the utility of the candidate miRNAs as blood-based biomarkers. Consistent with our findings that the 7-miRNA panel in ICC tissues and preoperative plasma samples was associated with poor prognosis, we found that risk scores based on the miRNA panel were significantly higher for the post-operative samples of patients with poor outcomes than in the samples of patients with prolonged survival (p < 0.05; Fig. 3e). These findings highlight the clinical significance of our miRNA panel in identifying survival outcomes in patients with ICC.

Fig. 3. Validation of the panel in clinical plasma samples and the establishment of a final risk-stratification model for patients with ICC.

a ROC curve for the performance of the miRNA panel in matched plasma samples from a subset of 48 patients in the additional validation cohort. b Violin plots of risk scores, based on the panel analysis of plasma samples, in patients with short- versus long-term survival; *P < 0.05. c, d Comparisons of c OS and d RFS in high- versus low-risk groups, determined based on the panel analysis of plasma samples. e Comparison of the risk scores in matched pre- and post-operative plasma samples from a subset of 20 patients in the clinical training cohort. *P < 0.05. f ROC curve for the performance of the combination model, incorporating the 7-miRNA panel and CA19-9 levels, compared to each factor alone.

A combination signature incorporating miRNA biomarkers and CA19-9 levels demonstrated significantly higher accuracy than the miRNA signature alone in patients with ICC

Because CA19-9 is a widely established and important biomarker in ICC, we next examined whether combining levels of this glycoprotein with our 7-miRNA panel would improve the accuracy of our model for detecting survival in patients with ICC. The AUC of the combined risk model was 0.78, 0.81, 0.79 and 0.82 in age, gender, tumour size, and CEA, respectively (Fig. S1). Interestingly, the risk-stratification model demonstrated superior predictive accuracy compared to the miRNA panel alone (AUC = 0.85, 95% CI = 0.67–0.96, Fig. 3f). Specifically, we observed that its sensitivity, specificity, positive predictive value, and negative predictive values were 91.7%, 81.3%, 78.6%, and 92.9%, respectively (Table 3), highlighting the superiority of the combination signature for identifying poor outcomes in patients with ICC. These findings further highlight the potential of our liquid biopsy-based risk-stratification model for clinical use in patients with high-risk ICC.

Table 3.

Performance of the 7-miRNA model in estimating the risk of survival in patients from the clinical cohorts.

Variable	Value (95% CI)
	Training cohort	Validation cohort	Additional validation cohort	Risk-stratification model
Sensitivity, %	77.4 (58.9–90.4)	100.0 (63.1–100.0)	83.3 (51.6–97.9)	91.7 (61.5–99.8)
Specificity, %	81.0 (68.6–90.1)	72.7 (49.8–89.3)	75.0 (47.6–92.7)	81.3 (54.4–96.0)
AUC, %	83.4 (74.0–90.4)	82.4 (64.2–93.8)	78.1 (58.5–91.4)	85.4 (67.0–95.8)
PPV, %	68.6 (55.4–79.3)	57.1 (40.3–72.5)	71.4 (50.8–85.8)	78.6 (56.6–91.2)
NPV, %	87.0 (77.6–92.9)	100.0	85.7 (62.1–95.6)	92.9 (66.2–98.9)

AUC area under the curve, CI confidence interval, NPV negative predictive value, PPV positive predictive value.

Discussion

Many prognostic staging systems, including the commonly used AJCC TNM system, have focused primarily on risk factors obtained postoperatively through the pathologic assessment of tumour specimens. Risk stratification and the prediction of prognosis after resection are important for evaluating post-operative treatment options; however, the resection of ICC often necessitates a major hepatectomy [12, 13]. In addition, ICC is aggressive and often associated with a poor prognosis, even among patients undergoing resection with curative intent [27–29]. As such, better preoperative prognostic stratification of patients with ICC can provide useful information to help direct clinical decision-making and counsel patients.

The preoperative model would allow for risk stratification at the time of treatment selection, facilitating treatment decisions (e.g. the selection of neoadjuvant or conversion therapy or nonoperative treatment modalities) [30–33]. Several previous reports have called attention to the potential impact of perioperative treatments in patients with ICC [34–41]. NCCN guidelines indicate that neoadjuvant therapy should be considered in patients with gallbladder and pancreatic cancer, and although there is a lack of supporting clinical data, neoadjuvant chemotherapy (NAC) has also been considered as potential treatment option in patients with ICC, according to resectability status. Thereafter, we have established the preoperative predictive models to identify patients who are most likely to benefit from surgical resection in resectable ICC. Although the use of preoperative systemic therapy for patients with resectable ICC is not established, the patients who underwent NAC prior to surgical resection experienced improved OS [42]. In the low-risk patients with postoperative survival outcomes, radical resection should be recommended, while the clinicians can consider the treatment strategies such as radical surgery without lymphadenectomy or chemotherapy for the high-risk patients with the elderly or severe comorbidities. Although more data is needed, preferably through well-designed prospective clinical trials, to define the indications for NAC in resectable ICC patients, our risk model robustly identifies patients with high-risk ICC preoperatively and can be used to improve overall patient management. In addition, preoperative assessments would enable surgeons to provide patients with a more accurate prognostic perspective at the time of the initial consultation.

In the present study, using a systematic and comprehensive biomarker discovery approach, we successfully identified a novel miRNA panel that robustly detects survival outcomes after hepatectomy in patients with ICC. Accumulating reports have shown that some candidate miRNAs could predict the prognosis in ICC [43–45]. However, since these reports lack clinical validation, our present study reports validation of a miRNA signature in large cohorts. More importantly, we used pretreatment blood specimens and resected tissue specimens to confirm our developed miRNA panel for predicting survival outcomes. We found that the majority of the patients in post-surgical blood specimens exhibited lower risk scores based on the miRNA panel compared to that of pre-surgical blood specimens, highlighting its cancer-specificity (Fig. 3E). Moreover, the risk score of patients with short-term survival was higher than that of patients with long-term survival in the post-operative specimens. These findings may be a possibility to strengthen the prediction of survival outcomes after operation. The successful validation of our panel in pre- and post-operative plasma samples, yielding results comparable to those obtained using surgically resected specimens, underscores its potential for clinical translation to improve survival in patients with ICC.

From a functional viewpoint, various miRNAs in our developed biomarker panel have been shown to be involved in cancer pathogenesis. For instance, miR-10b-3p is a direct target of sperm-associated antigen 5, which has been identified as an important proliferation marker and chemotherapy-sensitivity predictor in breast cancer [46]. MiR-27a-3p directly targets Acsl1 and Aldh2, which increased cell proliferation and tumour growth in hepatocellular carcinoma [47]. In ICC, miR-27a offers a novel perspective on the molecular pathogenesis and may serve as a potential target for antimetastatic molecular therapies [48]. Other reports demonstrated that miR‑106b‑3p overexpression promoted migration, invasion and epithelial‑mesenchymal transition by regulating Wnt/β‑catenin signalling pathway of oesophageal squamous cell carcinoma [49]. In addition, miR-106b was reported to reverse chemotherapy resistance and related to poor prognosis in ICC, suggesting its potential role as a novel and powerful strategy for ICC chemotherapy [50].

We would like to acknowledge some potential limitations of the present study. First, our retrospective study design might introduce selection bias. Furthermore, we had somewhat smaller number of plasma samples available; hence, a prospective clinical trial is required to further validate our results and further analysis such as the cost-benefit analysis and nomogram will be needed for the clinical practice. Second, our study used training and validation cohorts of patients from Japan who had similar clinicopathologic characteristics; such characteristics could potentially vary if we were to analyse patient populations from other countries. Therefore, it will be important to validate the selected biomarkers and our risk-stratification model in patient cohorts from other countries to further reinforce the generalisability of our findings. Third, the intra-tumour heterogeneity of ICC tissue specimens complicates the analysis of their molecular profiles. To minimise potential sampling bias, we conducted biomarker discovery and evaluated the performance of our miRNA signature using multiple independent, public datasets, with validation using a liquid biopsy approach. Therefore, we believe our study provides compelling evidence for the clinical significance of our miRNA signature for predicting the prognosis of patients with ICC, potentially constituting a major step towards the robust, widespread analysis of molecular biomarkers for the clinical risk-assessment and management of lethal malignancies.

In conclusion, using genome-wide expression profiling, we identified and developed a novel miRNA signature associated with prognosis in patients with ICC that was successfully validated in resected tissue specimens, as well as pre- and post-operative plasma samples. Pending validation in future prospective studies, our findings highlight the potential clinical impact of this model for conducting more appropriate patient selection and improving individualised treatment strategies for patients with ICC.

Author contributions

Y Wada: Study concept and design; specimen provider; acquisition of clinical data; analysis and interpretation of data and statistical analysis; drafting of the manuscript. M Shimada: Study concept and design; specimen provider; acquisition of clinical data; drafting of the manuscript. Y Morine: Study concept and design; specimen provider; acquisition of clinical data; drafting of the manuscript. T Ikemoto: Study concept and design; specimen provider; acquisition of clinical data; drafting of the manuscript. Y Saito: Study concept and design; specimen provider; acquisition of clinical data; drafting of the manuscript. H Baba: Study concept and design; specimen provider; acquisition of clinical data; drafting of the manuscript. M Mori: Study concept and design; specimen provider; acquisition of clinical data; drafting of the manuscript. A Goel: Study concept and design; analysis and interpretation of data and statistical analysis; drafting of the manuscript

Funding information

This work was supported by CA72851, CA181572, CA184792, CA202797, and CA227602 grants from the National Cancer Institute, National Institutes of Health.

Data availability

TCGA miRNA expression profiling data were downloaded from the University of California, Santa Cruz Xena Browser (https://xenabrowser.net). Normalised non-coding RNA profiling and clinical data from the GSE53870 dataset were downloaded from the Gene Expression Omnibus database.

Ethics approval and consent to participate

The study was conducted in accordance with the Declaration of Helsinki. A written informed consent was obtained from all patients, and the study was approved by the institutional review boards of all participating institutions.

Consent to publish

No individual person’s data were used in this paper, and thus no consent for publication was required for these studies.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41416-022-01710-z.