Screening of differentially enriched circRNAs in bile and serum exosomes to identify potential candidates for CCA-BO diagnosis and prognosis prediction

The recruitment process of all four cohorts and sample acquisition was summarized in Fig. 1a, b. The pilot cohort consisted of ten hospitalized patients with biliary obstruction who were eventually confirmed as CCA-BO (n = 5) and BBO (n = 5) by pathology. Exosomes were isolated by differential ultracentrifugation from bile samples and were confirmed by transmission electron microscope (TEM) (Fig. 2a) and immunoblotting analysis of exosomal markers (Alix, TSG101, and CD63) (Fig. 2c). Nanoparticle tracking analysis (NTA) revealed an average diameter of 142 nm and 145 nm in the extracted exosomes of bile and serum, respectively (Fig. 2b). Microarray analysis revealed 745 differentially enriched (absolute logFC ≥1.5, P < 0.05) exosomal circRNAs (489 upregulated, 256 downregulated, and 12,783 conservatively expressed circRNAs) as shown in Fig. 2d and Supplementary Fig. 2a. These results revealed the active secretion of exosomal circRNAs from CCA into its microenvironment, indicating that biliary exosomal circRNAs have the potential to serve as CCA-specific biomarkers. Among the top 20 upregulated circRNAs (absolute logFC ranged from 4.75 to 3.15), 6 were selected for further investigation according to p value, documentation in circBase, origin (exonic/intronic) and in vitro pre-experiments (CCK8) as shown in Supplementary Fig. 2b. The expression profile of the 6 candidate circRNAs was visualized by heatmap (Fig. 2e).

Fig. 1

Summary of cohort recruitment and sample acquisition. a A pilot cohort and a discovery cohort were prospectively recruited during April 2022 and June 2022 to identify CCA-BO-specific exosomal circRNAs. A training cohort and a validation cohort were retrospectively analyzed for model establishment and verification. b Serum samples were acquired from all recruited patients while bile samples were acquired from patients undergoing biliary drainage, followed by exosome isolation

Fig. 2

The discovery of a triple-circRNA panel in both bile and serum exosomes. The isolated exosomes were confirmed by a TEM, b NTA, and c immunoblotting. The main band of TSG101 was indicated with an arrowhead. d Volcano plot of microarray analysis revealed the pattern of differentially enriched exosomal circRNAs between CCA-BO and BBO patients. e Heatmap of six candidate circRNAs detected by microarray analysis. f Quantification of six candidate circRNAs with divergent primers in the discovery cohort. Three of them (hsa-circ-0000367, hsa-circ-0021647, and hsa-circ-0000288) were found to be upregulated in both bile and serum exosomes. g RNase R digestion, and h Sanger sequencing verified the qPCR products. *P < 0.05; **P < 0.01; ***P < 0.001; \(^}.}.}}\;}\); (Mann–Whitney U test for continuous variables, presented as median with range and quartile)

We then went on for further screening of exosomal circRNA biomarkers. Compared with bile samples, peripheral blood has better accessibility as not all patients with obstructive jaundice undergo preoperative biliary drainage; while theoretically, biliary exosomes better reflect the status of biliary disease, as bile fluid is non-circulatory and is harder to be interfered by exosomes from non-hepatobiliary origins.15 To verify our hypothesis, a discovery cohort of ten CCA-BO patients and ten BBO patients were recruited to prospectively collect bile and serum samples. Divergent primers for each circRNAs were designed for quantitative real-time PCR (qPCR), and standard curve method was applied for absolute quantification. Among the six circRNA candidates, three were found to be simultaneously upregulated in bile and serum exosomes, namely hsa-circ-0000367, hsa-circ-0021647, and hsa-circ-0000288 (Fig. 2f). The circularity of the target circRNAs was verified by RNase R and dactinomycin treatment (Fig. 2g and Supplementary Fig. 2c). Sanger sequencing was also performed to verify qPCR products (Fig. 2h). These findings suggested that the selected three exosomal circRNAs have the potential to serve as CCA biomarkers in both bile and serum samples.

Target circRNAs promote proliferation and metastasis of CCA in vitro and in vivo

Before verifying the abundance of the three selected circRNAs in a larger population, we wanted to explore their functional role in the development of CCA. We used two human primary CCA cell lines (CCLP-1 and Huh-28) for in vitro verification of the three target circRNAs. Loss-of-function experiment was carried out using circRNA-specific anti-sense oligonucleotides (ASO) targeting their back-splicing sequence. The knock-down efficiency of target circRNAs was verified by qPCR, without affecting mRNA expression of their host genes (Supplementary Fig. 2d). Remarkably, ASO-mediated depletion of hsa-circ-0000367, hsa-circ-0021647, and hsa-circ-0000288 significantly reduced the viability of CCA cells in both short-term and long-term experiments (Fig. 3a, b). Edu assay revealed that the apoptotic rate was significantly elevated following depletion of each target circRNA (Fig. 3c). Annexin V/PI flow cytometry also revealed an increased proportion of cell apoptosis following ASO-mediated depletion of target circRNAs (Fig. 3d). These results suggested that silencing hsa-circ-0000367, hsa-circ-0021647, and hsa-0000288, respectively, leads to impaired CCA proliferation in vitro. We next performed a transwell assay to determine the role of target circRNAs in CCA metastasis. The results showed that the invasion and migration potential of CCA was suppressed by respectively knocking down each circRNA (Fig. 3e).

Fig. 3

Loss-of-function assay of target circRNAs in vitro. hsa-circ-0000367, hsa-circ-0021647, and hsa-circ-0000288 were respectively knocked down by circRNA-specific ASOs (ASO-367, ASO-21647, and ASO-288) in two human primary CCA cell lines. a CCK8 and b clone formation assay and c Edu assay revealed short and long-term attenuation of proliferative activity in CCA cells following ASO transfection. Scale bar: 200 μm. d Cell apoptosis was detected by Annexin V/PI flow cytometry following ASO treatment. e Loss of target circRNAs attenuates CCA migration and invasion, evaluated by transwell assays. Scale bar: 500 μm. *P < 0.05; **P < 0.01; ***P < 0.001; \(^}.}.}}\;}\); (two-way ANOVA with Turkey’s multiple-comparison test for curve comparison; two-tailed unpaired Student’s t test for continuous variables, presented as mean ± SD)

In addition, we intended to look for underlying mechanisms of these circRNA-induced malignant phenotypes. To find out whether the three circRNAs function as ceRNAs to affect downstream miRNAs and mRNAs, we constructed a circRNA–miRNA–mRNA network (Supplementary Fig. 3a). Cell adhesion molecules (CAMs) were revealed as the most significantly enriched pathway downstream, and the regulation of key molecules was verified by immunoblotting following treatment of circRNA–ASO complex in vitro (Supplementary Fig. 3b).

We then wanted to know whether these in vitro effects of the three circRNAs can be verified in vivo. CCLP-1 cells were implanted subcutaneously into BALB/c nude mice and then treated with circRNA-specific ASO by injection around tumor planting site (Fig. 4a). Based on similar initial tumor volume 1 week following implantation, circRNA-specific ASO treatment led to significant lower tumor volume compared with ASO-NC control (Fig. 4b). Tumors were harvested to investigate the expression of key biomarkers for tumor proliferation (Ki-67) and epithelial-to-mesenchymal transition (EMT) (markers: E-cadherin, β-catenin, and Vimentin), since these markers are known to be associated with biliary oncogenesis.16,17 Immunochemistry revealed decreased expression of Ki-67, β-catenin, Vimentin, and increased expression of E-cadherin in circRNA-specific ASO-treated groups, which was further validated by immunoblotting (Fig. 4c, d).

Fig. 4

Loss of target circRNAs attenuates CCA proliferation and EMT in vivo. a Schematic representation of circRNA-specific ASO treatment in balb/c nude mice. CCLP-1 cells were subcutaneously injected into the right flank to establish a tumor-bearing model and randomized into ASO-NC group (n = 5), ASO-367 group (n = 5), ASO-21647 group (n = 5) and ASO-288 group (n = 5) 7 days following tumor implantation. b ASO treatment of each target circRNA leads to reduced tumor volume 22 days following tumor implantation. Expression of Ki-67, Vimentin, β-catenin, and E-cadherin were evaluated in harvested tumor by c immunohistochemistry and d immunoblotting. Scale bar: 100 μm. *P < 0.05; **P < 0.01; ***P < 0.001; \(^}.}.}}\;}\); (two-way ANOVA with Turkey’s multiple-comparison test for curve comparison; two tail unpaired Student’s t test for continuous variables, presented as mean ± SD)

These in vitro and in vivo findings indicated that these three circRNAs may support the progression of CCA, thereby potentially influencing the prognosis of patients who have undergone treatment.

Construction and verification of target circRNA-based diagnostic model

To establish a diagnostic and prognostic model intended for clinical use, we then investigated the expression profile of the 3 circRNAs in both bile and serum exosomes based on a larger population. A training cohort (71 BBOs and 113 CCA-BOs) was retrospectively analyzed to collect bile and serum samples along with the clinicopathological features of the patients. Before verifying the abundance of target circRNAs, we examined the performance of lab tests currently used in CCA-BO diagnosis. We first conducted a baseline comparison between the benign and malignant groups (Table 1). The results showed that there was a significant difference between the BBO group and CCA-BO group in terms of CA19-9, CEA, CA125, TB, DB, DB/TB ratio, ALT, AST, and ALB. Next, the diagnostic potential of these biomarkers was evaluated by area under the receiver operating characteristic curves (AUROC). We used previously described criteria to estimate the diagnostic power of these biomarkers according to AUROC: an AUROC of 0.5 suggests no discrimination, 0.7–0.8 is considered acceptable, 0.8–0.9 is considered excellent, and more than 0.9 is considered outstanding.18 As expected, CA19-9 was revealed as the most eligible biomarker for CCA-BO diagnosis (AUROC = 0.759), which is merely considered acceptable. Other potentially useful biomarkers were DB (AUROC = 0.714), DB/TB ratio (AUROC = 0.711), CEA (AUROC = 0.707), TB (AUROC = 0.706), and CA125 (AUROC = 0.699) (Supplementary Fig. 4). These findings suggested that the currently used biomarkers for CCA-BO diagnosis is still not satisfactory.

Table 1 Baseline characteristics of patients in the training cohort and validation cohort to establish a diagnostic model

Next, we quantified the three circRNAs in the training cohort and found that, consistent with the discovery cohort, hsa-circ-0000367, hsa-circ-0021647, and hsa-0000288 were significantly upregulated in CCA-BO group in both bile and serum exosomes (Fig. 5a, b). Based on the qPCR data measured by absolute quantification, we first evaluated the diagnostic potential of each circRNA, and then established two pooled models via logistic regression:

$$\begin\,\,\,(-)=-4.829+219.749\times }_}}0000367}+210.340 \\\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\times_}-}0021647}}+254.645\times }_}}0000288}\end$$

$$\begin\,(-)=-2.945+35.448\times }_0000367}+290.995\\\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\times }_0021647}+72.636\times }_0000288}\end$$

Fig. 5

Establishment and validation of diagnostic models. The abundance of hsa-circ-0000367, hsa-circ-0021647, and hsa-circ-0000288 in a bile and b serum exosomes were measured by qPCR in the training cohort. ROC curves of c bile and d serum circRNA signatures (Bile-DS and Serum-DS) were illustrated. e The diagnostic power of Bile-DS, Serum-DS and CA19-9 was compared. f Summary of the established diagnostic models and their verification in the validation cohort. Cutoff values were determined by optimal Youden’s index. g The abundance of circRNAs in bile exosomes and serum counterparts showed correlation with each other. h Patients in the training cohort were grouped by high/low diagnostic scores and compared of clinical outcomes. *P < 0.05; **P < 0.01; ***P < 0.001; \(^}.}.}}\;}\); (Mann–Whitney U test for continuous variables, presented as median with range and quartile; Pearson’s chi-squared test for categorical variables)

The goodness-of-fit was measured by component plus residual plots (Supplementary Fig. 7b). We also verified the independency of Bile-DS and Serum-DS by performing Spearman’s correlation analysis with currently used clinical indices. The results showed that Bile-DS and Serum-DS were independent from serum CA19-9, CEA, CA125, TB, DB, ALT, AST, and ALB levels, at the same time had moderate to strong correlation (Spearman’s ρ = 0.536, P < 0.001) with each other (Supplementary Figs. 6 and 7a). The diagnostic accuracy of each circRNA and their pooled models were evaluated by AUROC. The results showed that, compared with individual circRNAs, the pooled model had better diagnostic accuracy in bile (AUROC 0.947 vs. 0.850 or 0.801 or 0.902) and in serum (AUROC 0.861 vs. 0.731 or 0.822 or 0.733) (Fig. 5c, d). Meanwhile, we compared the diagnostic accuracy of Bile-DS, Serum-DS and CA19-9, revealing Bile-DS as the most accurate index (AUROC = 0.947; DeLong test: Bile-DS vs. CA19-9 p < 0.01, Bile-DS vs. Serum-DS P = 0.030), followed by Serum-DS (AUROC = 0.861; DeLong test: Serum-DS vs. CA19-9 P = 0.045) and CA19-9 (AUROC = 0.759) (Fig. 5e). These results suggested that this bile/serum exosomal circRNA panel may serve as a novel diagnostic tool in patients with CCA combined with biliary obstruction. For further verification, a validation cohort (54 BBOs and 51 CCA-BOs) was retrospectively analyzed and assayed using the same methods. Using the cutoff value determined by the training cohort, we verified the diagnostic performance of our established models (Fig. 5f). The accuracy index (ACC) of Bile-DS reached 0.897, followed by Serum-DS (ACC = 0.810) and CA19-9 (ACC = 0.705). Furthermore, we performed a propensity score match analysis to adjust the baseline of the CCA-BO group and the BBO group, in order to simulate more challenging application scenarios. The results showed that, Bile-DS (AUROC = 0.921, RR = 8.40 [95% CI, 2.78–30.30]) and Serum-DS (AUROC = 0.826, RR = 2.33 [95% CI, 1.63–3.63]) remained to be powerful diagnostic indices for CCA-BO in the training cohort after adjustment of clinicopathologic features including Sex, Age, CA19-9, CEA, CA125, TB, DB, ALT, AST, and ALB (Supplementary Figs. 8 and 9). Similar results were noted in the validation cohort (Bile-DS: ACC = 0.852; Serum-DS: ACC = 0.792) (Supplementary Figs. 10 and 11). These results suggested that our diagnostic model based on the triple-circRNA combination could robustly distinguish CCA-BO from benign origins.

Few studies involved detection of target ncRNAs in different types of bodily fluids, and their correlation remained unnoticed. Therefore, we performed a correlation analysis between the abundance of circRNAs in bile exosomes and their serum counterparts, revealing a moderate correlation (Fig. 5g and Supplementary Fig. 5). Based on these findings, we proposed that CCA-derived exosomal circRNAs can be detected simultaneously by bile and serum liquid biopsy.

Given that the three circRNAs had positive in vitro and in vivo findings in terms of CCA progression, we assumed that they might be related with some clinicopathological features of CCA-BO. To test this hypothesis, we divided CCA-BO patients in the training cohort into two groups based on their Bile-DS and Serum-DS levels, and compared postoperative outcomes between the two groups. The results showed significant differences in histologic grading, R0 resection rate, lymph node status and AJCC TNM staging (Fig. 5h and Supplementary Table 1). These findings suggested that the three circRNAs may also serve as prognostic indicators in CCA-BO patients undergoing surgical treatment.

Construction and verification of target circRNA-based prognostic model

To establish a prognostic model, CCA-BO patients who underwent curative-intent surgery in the training cohort (n = 83) and the validation cohort (n = 42) were followed up (Fig. 6a). Baseline characteristics of these surgical patients were summarized (Supplementary Table 2). We first performed a survival analysis and found indistinctive correlations between bile/serum exosomal level of each circRNA and overall survival of CCA-BO patients in both training and validation cohorts (Supplementary Figs. 12 and 13). Similarly, the correlation between each circRNA and recurrence-free survival (RFS) of CCA-BO patients was noted to be significant in only a few instances (Fig. 6c and Supplementary Fig. 14a). Considering the high invasiveness of CCA, we hypothesized that our target circRNAs might have better prediction value within a certain postoperative period. Previous research found that CCA tend to have early recurrence following curative-intent surgery, defined by a cutoff of 12 or 24 months following resection.19 Using a retrospective dataset from our medical center, we compared the recurrence patterns of patients with HCC and iCCA (without involvement of the hepatic hilus) along with patients from our cohorts (Fig. 6b). The results showed that, compared with HCC patients undergoing surgical resection, patients with CCA tend to have poorer RFS (Log-Rank P < 0.001). By a cutoff of 1 year, the RFS proportion in the three groups were 79.3%, 66.1%, and 53.3%, respectively. We then compared the abundance of each circRNA by dividing the patients using an early recurrence cutoff of 1 year. Remarkably, hsa-circ-0021647 and hsa-0000288 were significantly elevated in bile exosomes of patients with early recurrence, while similar outcome was noticed in all three serum exosomal circRNAs (Fig. 6d). The predictive effect of individual circRNAs was unremarkable (Supplementary Fig. 15). We then conducted a multivariate stepwise logistic regression analysis to determine target circRNAs to be included in the early recurrence model (Supplementary Fig. 16a). Early recurrence scores (ERS) were calculated by logistic regression analysis:

$$\,(-)=-2.522+116.399\times }_0021647}+61.133\times }_0000288}$$

$$\,(-)=-2.945+35.448\times }_0000367}+290.995\times }_0021647}+72.636\times }_0000288}$$

Fig. 6

Establishment and verification of early recurrence monitoring models. a Patients who underwent curative-intent surgery in the training cohort and validation cohort were followed up. b Recurrence patterns were compared in patients with HCC (n = 237), iCCA without involving the hepatic hilus (n = 80) and the training and validation cohorts (n = 118) from a single center to determine the best cutoff for early recurrence of CCA-BO patients. c Patients were grouped by their circRNA abundance to compare the recurrence-free survival. d Patients were grouped by the presence of early recurrence (1 year) to compare circRNA abundance. e ROC curves of bile and serum circRNA signatures (Bile-ERS and Serum-ERS) were illustrated. Patients were grouped by their Bile-ERS/Serum-ERS levels to compare the recurrence-free survival in (f) the training cohort and g the validation cohort. *P < 0.05; **P < 0.01; ***P < 0.001; \(^}.}.}}\;}\); (Mann–Whitney U test for continuous variables, presented as median with range and quartile; Kaplan–Meier analysis for survival data, Log-Rank test for curve comparison)

Goodness-of-fit of the models was measured by component plus residual plots (Supplementary Fig. 16b). The prognostic power of the models was verified by AUROC (training cohort, Bile-ERS: AUROC = 0.767; Serum-ERS: AUROC = 0.782; validation cohort, Bile-ERS: AUROC = 0.851; Serum-ERS: AUROC = 0.759) (Fig. 6e and Supplementary Fig. 16c). We then divided the patients in the training cohort into a high ERS group and a low ERS group, and found that both Bile-ERS and Serum-ERS were closely related to RFS (Fig. 6f). This was followed by verification in the validation cohort (Fig. 6g and Supplementary Fig. 14b). These results suggested that this recurrence monitoring model based on ERS is predictive of early recurrence of CCA and has high translational potential.

Based on this circRNA model aimed at predicting early recurrence of CCA, we sought for practical tools for the calculation of overall recurrence risk. A multivariate COX regression analysis was conducted to identify potential indicators for postoperative CCA recurrence. Bile-ERS and Serum-ERS were respectively analyzed with other covariates, including histologic grade, AJCC TNM staging, CCA subtype, tumor margin, preoperative biliary drainage, tumor diameter, lymph node status, macrovascular invasion, portal vein invasion and microvascular invasion. Schoenfeld’s global test was applied to test the proportional hazards assumption (Supplementary Figs. 17 and 18). The number of positive lymph nodes, portal vein invasion and macrovascular invasion were revealed as risk factors along with Bile-ERS; while number of positive lymph nodes and portal vein invasion were revealed as risk factors along with Serum-ERS (Supplementary Fig. 19). Nomograms based on the COX-PH model were subsequently established (Fig. 7a, b). The C-index of Bile-ERS/Serum-ERS-based models were 0.739 (95% CI, 0.723–0.755) and 0.726 (95% CI, 0.709–0.743), respectively. Calibration plots showed the close agreement between the predicted RFS and the actual status (Fig. 7c). The established recurrence monitoring models were verified by the validation cohort: the AUROC of Bile-ERS and Serum-ERS were 0.851 and 0.759, respectively, and the prognostic accuracy using cutoff values defined by the training cohort were 0.720 and 0.762, respectively (Fig. 7d). In summary, these findings strongly suggested that our circRNA-based prognostic models are robust tools for potential clinical application, and a schematic workflow on how to choose between detection strategies based on bile and serum was provided (Supplementary Fig. 20).

Fig. 7

Establishment of nomograms for CCA recurrence monitoring. a Bile-DS and b Serum-DS were trained into nomograms along with other prognostic predictors revealed by COX-PH analysis. c The prediction accuracy of the nomograms was verified by calibration curves. The 45-degree reference line represents an ideal nomogram. d Summary of the establishment and verification of CCA recurrence monitoring signatures