Common DNA methylation changes in biliary tract cancers identify subtypes with different immune characteristics and clinical outcomes

The included samples of the Eastern Hepatobiliary Surgery Hospital (EHSH) cohort consist of three parts: cancerous, adjacent, and precancerous tissues. The 105 tumor samples (24 intrahepatic CCAs [iCCAs], 20 perihilar CCAs [pCCAs], 13 distal CCAs, and 48 GBCs) and 50 adjacent non-tumor samples (gallbladder [n = 28] and bile duct [n = 22]) were obtained immediately following surgery at EHSH from October 2017 to September 2019. The diagnosis and tumor purity were confirmed by two independent pathologists based on the resected samples. Characteristics of the cancerous and adjacent normal tissues are shown in Additional file 1: Table S1-2, respectively. All participants had not received anti-tumor treatment before surgery. Participants with carcinoma or benign disease of other organs were excluded from this study. In addition, eight samples of precancerous disease (e.g., gallbladder polyps) were collected for exploratory analysis (Additional file 1: Table S3).

The Cancer Genome Atlas-cholangiocarcinoma (TCGA-CHOL) cohort of 36 patients with epigenomic, transcriptomic, mutational, immune-related, and survival (OS and progression-free interval [PFI]) data were analyzed in the present study [10, 16]. The immune characteristics of this cohort (e.g., signatures and immune cell sorting) were retrieved from the study of Thorsson et al. [16].

Patients or the public were not involved in the design, conduct, reporting, or dissemination plans of our research. All collection and usage of human samples and clinical data were in accordance with the principles of the Declaration of Helsinki and approved by the Ethics Committee of Eastern Hepatobiliary Surgery Hospital (EHBHKY2018-02-014). The written consents were received from all the participated patients. This report follows the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline.

All sequencing experiments were implemented in a College of American Pathologists (CAP)- and Clinical Laboratory Improvement Amendments (CLIA)-certified laboratory (Burning Rock Biotech, Guangzhou, China) before May 2020. The procedure for DNA extraction was as previously described [17]. In brief, DNA was extracted with a QIAamp DNA formalin-fixed paraffin-embedded (FFPE) tissue kit according to the manufacturer’s instructions. DNA concentration was measured by the Qubit double-stranded DNA assay (Life Technologies, Carlsbad, CA, USA).

As for methylation sequencing, a capture-based method, SeqCap Epi CpGiant Probes (Roche Sequencing Solutions, Madison, WI, USA) was performed to detect > 5.5 million CpG sites (capture size, 80.5 Mb) [18]. We generated a bisulfite sequencing library with the brELSA™ method (Burning Rock Biotech, Guangzhou, China) [19]. The target libraries were quantified by real-time PCR and sequenced on NovaSeq 6000 with 50× target depth on average.

Since differentially methylated regions consisting of multiple CpG sites played more important roles than a single CpG site in cancer detection as reported [20], we defined 319,133 methylation regions of CpG sites with close genomic distance and highly correlation in methylation level [19], which were analyzed in the present study.

As for mutation sequencing, a capture-based targeted deep sequencing was performed using a 520-gene panel, spanning 1.64 Mb of the human genome (included genes are shown in Additional file 1: Table S4). Detailed descriptions of sequencing and capturing single nucleotide variant and copy number variation are shown in Additional file 2: Method S1.

Mutated genes included in our analysis were restricted to non-silent mutations consisting of non-sense mutation, missense mutation, frameshift mutation, inframe mutation, splice site mutation, translation start site mutation, and non-stop mutation. Truncating mutations of oncogene were excluded because most of these are passenger mutations with limited cancer-promoting function.

The signaling pathways and their members we analyzed are shown in Additional file 1: Table S5. The definition of pathways drew upon previous genomic studies [3, 5].

For each tumor FFPE block, a 5-μm section was cut and stained with the Dako 22C3 mouse monoclonal antibody with Dako Autostainer Link-48 platform according to the manufacturer’s instructions. Cores showing a neoplastic component ≥ 30% were considered as adequate. PD-L1 positivity was determined by any expression in tumor cells or immune cells evaluated independently by two pathologists.

To assess the between-group difference, we performed the (i) Fisher exact test, chi-square test, and Cochran-Armitage test for trend for categorical variables; (ii) Mann-Whitney test for continuous variables; and (iii) Kaplan-Meier (KM) method, log-rank method, and Cox regression (hazard ratio [HR] and 95% confidence interval [CI]) for survival variables (OS and PFI). The covariates with P value below 0.05 in the univariable analysis were included in the following multivariable model. The differentiated methylation regions (DMRs) between tumors and adjacent tissues were selected by the Mann-Whitney test for further clustering. Detailed description of single sample gene set enrichment analysis is shown in Additional file 2: Method S2.

The non-supervised clustering of methylation data in both the EHSH and TCGA cohorts was performed by the K-means method (R package, ConsensusClusterPlus). Gene Ontology (GO) enrichment analysis was performed by R (R package, clusterProfiler) [21‐23]. The 12-DMR prognostic model was built via least absolute shrinkage and selection operator (LASSO).

All statistical analyses mentioned above were performed using IBM SPSS Statistics 22 and R 3.4.2, and the graphs were drawn by GraphPad Prism 9 and R 3.4.2. The nominal level of significance was set as 5%, and all 95% CIs were 2-sided.

By comparing the methylomes of cancerous and normal adjacent tissues separately in CCAs (57 treatment-naïve CCAs vs. 22 adjacent bile ducts) and GBCs (48 treatment-naïve GBCs vs. 28 adjacent gallbladders), 5279 significant DMRs with β value difference above 0.15 in both CCAs and GBCs were identified (Fig. 1A). Moreover, we assessed the consistency of these 5279 DMRs among iCCAs, pCCAs, and dCCAs, and 3369 DMRs had minimal β value differences of 0.15 in all three subgroups (Fig. 1A). Of these, 835 DMRs showed lower β values in tumor samples compared to adjacent non-tumor samples (referred to as hypomethylation DMRs), and the other 2534 DMRs exhibiting higher β values were referred to as hypermethylation DMRs. All of the 3369 DMRs showed great significance by comparing BTCs and adjacent non-tumor samples (maximal P value = 5.66 × 10⁻⁸, FDR P value< 0.05, Fig. 1B). In addition, we found significant positive correlations of the β value differences of the 835 hypomethylation DMRs among iCCAs, pCCAs, dCCAs, and GBCs (Fig. 1C). Similarly, significant positive correlations were found in the 2534 hypermethylation DMRs (Fig. 1D). These correlations further indicate the consistency of these DMRs among the BTCs at different anatomic locations.

The hypermethylation DMRs were enriched in the regions of CpG islands, promoters, and exons of coding sequence, but not CpG shores, CpG shelves, open sea, introns, and intergenic regions (Fig. 1E and Additional file 1: Table S6). The top 50 hypermethylation and hypomethylation DMRs are displayed in Additional file 1: Table S7 and S8, respectively. GO analysis revealed that hypomethylation mainly affected the genes related to Rho GTPase binding, transmembrane transporter activity, and lyase activity (Fig. 1F), and hypermethylation targeted DNA binding and transcription (Fig. 1G).

Using the top 1000 most variable DMRs of the 3369 DMRs identified above, 163 tissues (105 treatment-naïve BTC samples, 50 adjacent non-precancerous bile duct or gallbladder tissues, and 8 precancerous lesions) were classified into 6 groups by non-supervised consensus clustering (Fig. S1). Baseline characteristics of the abovementioned samples and the comparison among iCCAs, d/pCCAs, and GBCs are shown in Additional file 1: Table S1-3. As delineated in the heatmap (Fig. 2A), ranging from the methylation cluster 1 to cluster 6, the degree of global methylation changes gradually decreased, and the proportion of non-precancerous adjacent tissues gradually increased. Despite that the GBCs in our cohort exhibited poorer histological grade and later TNM stage compared to iCCAs and d/pCCAs (Additional file 1: Table S1), similar proportions of GBCs, dCCAs, pCCAs, small-duct iCCAs, and large-duct iCCAs were observed in different methylation clusters (P = 0.72), indicating that the methylation clusters based on the 3369 common DMRs were independent of anatomic sites and potentially reflected a shared characteristic among the cancers along biliary tract.

As non-precancerous adjacent tissues were enriched in the methylation clusters 4–6 with minimal methylation changes (Fig. 2A), we speculated that the BTCs in clusters 4–6 might exhibit lower invasiveness, and therefore, clusters 4–6 might be associated with longer survival. Among the 80 BTC patients with OS data (characteristics are shown in Additional file 1: Table S1), clusters 4–6 showed better prognosis compared to clusters 1–3 (HR = 0.07, 95% CI 0.01–0.65, P = 0.017, Fig. 2B). Similar trends were observed in the subgroups classified by anatomic location, i.e., iCCAs, d/pCCAs, and GBCs (Fig. 2C, D), and stage, i.e., stages I–II and stage III–IV (Fig. 2E). We hereby defined the cluster-based risk according to our methylation clusters (high cluster-based risk, clusters 1–3; low cluster-based risk, clusters 4–6). We further use the top 500 and 250 most variable DMRs to perform non-supervised clustering. Compared to the one using top 1000 DMRs as described above, 95.7% of the samples had the same results of the cluster-based risk (Additional file 1: Table S9), indicating the robustness of the classification by common methylation DMRs among BTCs.

Tumor purity (also termed as tumor cellularity) might somewhat influence the sequencing data of methylome and therefore the clustering outcome. Similar to the result of Goeppert et al.’s study in iCCAs, we observed slightly lower tumor cellularity in the samples with lower methylation changes (average, 40.8% in clusters 4–6 compared to 64.1%, 56.5%, and 47.7% in clusters 1, 2, and 3, respectively, Fig. 2F). Lower tumor purity might indicate a higher proportion of stromal area and/or higher density of tumor-infiltrating immune cells, which was partially suggested by the lower CD8A methylation in clusters 4–6 which might reflect higher infiltration of CD8⁺ T cells. Despite the slight difference of tumor purity in different clusters, tumor purity was not linked with OS (continuous variable: HR = 1.00, 95% CI 0.99–1.02, P = 0.933; categorical variable [≥ 50% vs. < 50%]: HR = 1.00, 95% 0.99–1.02, P = 0.672). The cluster-based risk showed an undifferentiated prognostic effect in the patients with higher tumor purity (≥ 50%, P = 0.077) and lower tumor purity (< 50%, P = 0.022, Fig. 2G). These results rule out the influence of tumor purity on the prognostic effect of the cluster-based risk.

To investigate whether the prognostic effect of the cluster-based risk is independent of other variables, we firstly analyzed the association between clusters and key clinicopathological variables, including sex, age, smoking history, anatomic location, TNM stage, histological grade, resection margin, carcinoembryonic antigen (CEA), carbohydrate antigen 19-9 (CA19-9), alpha-fetoprotein (AFP), and multiple immunohistochemical staining markers (e.g., Hep-1, Mucin 1 [MUC-1], P63, and S100). No significant association was revealed except the lower frequency of CEA abnormality in clusters 4–6 (P = 0.007, Additional file 1: Table S10). Moreover, univariable analyses discovered that anatomic location, TNM stage, histological grade, resection margin, and CEA abnormality were significantly associated with OS in addition to the cluster-based risk (Table 1). In the multivariable model, OS was significantly associated with the cluster-based risk (HR = 0.13, 95% CI 0.02–0.96, P = 0.045), rather than anatomic site (Table 1). These results suggest that the methylation-based clusters might be an independent biomarker predicting prognosis in BTCs.

To specifically discover the prognosis-related DMRs, 80 BTC patients with OS data were randomly assigned to the training and validation sets with a 2:1 ratio (Fig. 3A), and univariable analysis revealed 54 prognosis-related DMRs in the training set. Based on these 54 DMRs, LASSO regression was performed and constructed a LASSO score based on 12 DMRs (optimal lambda selection and LASSO coefficient profiles are shown in Fig. S2 and Additional file 1: Table S11, respectively). Using this score and the cutoff (median value) derived from the training set, we observed a consistent prognostic effect in both the training set (P < 0.001, Fig. 3B) and the validation set (P = 0.047, Fig. 3C). In the total 80 BTC patients (methylation data are shown in Fig. 3D), a lower LASSO-based risk was associated with better OS (HR = 0.21, 95% CI 0.10–0.43, P < 0.001, Table 1). Combining the LASSO-based risk and the cluster-based risk could differ the BTC patients into subgroups with distinct survivals (P < 0.001, Fig. 3E). The prognostic effect of the LASSO-based risk was undifferentiated in the subgroups classified by anatomic location (GBC: P = 0.009; CCA: P = 0.041, Fig. 3F), TNM stage (stages I–II: P = 0.070; stages III–IV: P < 0.001, Fig. 3G), or tumor purity (≥ 50%: P = 0.001; < 50%: P = 0.004). Furthermore, in the multivariable model, the LASSO-based risk (HR = 0.26, 95% CI 0.11–0.59, P = 0.001), rather than the anatomic site (d/pCCA vs. iCCA vs. GBC), was associated with OS (Table 1). These results indicate the good robustness of LASSO-based risk that may be effective regardless of other key variables including TNM stage and anatomic location.

Of the 105 tumor samples, 99 received targeted deep sequencing of mutational events of the coding sequence and splice sites, large genomic rearrangement, copy number variation (CNV), fusion, and microsatellite (Fig. 4A). Tumor mutational burden (TMB) was slightly higher in cluster 1 (Fig. 4B), and CNV events were enriched in clusters 1–2 (Fig. 4C), indicating the association between methylation changes and chromatin instability.

The mutational frequencies of TP53, BRCA1/2, ATM, and PI3K; WNT; homologous recombination repair (HRR); and cell cycle pathway gradually decreased with the order of the anatomical position (GBC, dCCA, pCCA, iCCA-large duct, and iCCA-small duct), while the changing trend of BRAF mutational frequency was the opposite (Fig. 4D). Similar differences were observed between the GBCs and the CCAs of the public cohorts in cBioPortal [3‐7, 24, 25]. In addition, CCND1/E1 and APC mutations were enriched in metastatic BTCs (Fig. S3).

Ranging from clusters 1, 2, and 3 to 4–6, the mutational frequency of cell cycle pathway (P = 0.022), CCND/E1 (P = 0.005), FGFR family (P = 0.049), and NOTCH pathway (P = 0.029) decreased gradually (Fig. 4E), suggesting the potential association between these mutational events and DNA methylation changes. In addition, mutations in the MMR pathway, BRAF, TGFBR1/2, and ARID1A were associated with worse OS in both univariable and multivariable analyses (Fig. 4E and Additional file 1: Table S12).

Due to the limited publicly available datasets of GBCs with all epigenomic, transcriptomic, and survival data, we sought to firstly investigate the association between methylation and immune characteristics in a cholangiocarcinoma dataset from TCGA (TCGA-CHOL). Thirty-six cancerous tissues and nine matched normal tissues were classified into six groups via non-supervised consensus clustering (Fig. S4). Similar to the result in the EHSH cohort, gradual changes of methylation were observed ranging from cluster 1 to cluster 6 (Fig. 5A). All normal samples were in cluster 6. Clusters 1–2 and clusters 3–5 were defined as the methyl-risk high and the methyl-risk low groups, respectively. Compared to the methyl-risk high group, OS (HR = 0.47, 95% CI 0.15–1.46, P = 0.19, Fig. 5B) and PFI (HR = 0.14, 95% CI 0.04–0.50, P = 0.002, Fig. 5C) were longer in the methyl-risk low group. In multivariable analysis, consistent results were revealed (OS: multivariable HR = 0.25, 95% CI 0.05–1.31, P = 0.100; PFI: multivariable HR = 0.06, 95% CI 0.01–0.52, P = 0.011, Additional file 1: Table S13-14). Fraction altered (percentage of copy number altered chromosome regions out of measured regions) was higher in the methyl-risk high group, indicating chromatin instability (Fig. 5D).

Thorsson et al. presented immunogenomics analyses of more than 10,000 tumors in TCGA, identifying 6 immune subtypes (C1: wound healing, C2: IFN-γ dominant, C3: inflammatory, C4: lymphocyte depleted, C5: immunologically quiet, C6: TGF-β dominant) that encompass 33 cancer types based on 6 key signatures [16]. Higher frequencies of C1 and C2 and lower frequencies of C4 and C6 were observed in the methyl-risk low group (P = 0.045, Fig. 5E), and a lower methyl-risk was associated with higher scores of macrophage and lymphocyte signatures (Fig. 5F). As described by Thorsson et al., both C1 and C2 subtypes had low Th1/Th2 ratio, high proliferation rate, and high intratumoral heterogeneity, and C2 had the highest M1/M2 macrophage polarization, a strong CD8 signal, and the greatest TCR diversity. On the contrary, C4 displayed a more prominent macrophage signature with suppressed Th1 and high M2 response, and C6 had the highest TGF-β signature, demonstrating their immunosuppressive tumor microenvironments [16].

We further assess the association between the clusters and immune features. The samples with a lower methyl-risk exhibited higher (i) richness and Shannon index of B/T cell receptor (Fig. 5G); (ii) expression of CD274 (PD-L1) and cytotoxic T-lymphocyte antigen 4 (CTLA-4, Fig. 5H); (iii) fraction of infiltrated leukocytes (Fig. 5I), including naïve B cell, plasma cell, monocyte, macrophage, CD8⁺ T cell, follicular helper T cell, and regulatory T cell (Fig. 5J); and (iv) multiple immune-related and angiogenesis signatures (Fig. 5K, L). Previous studies have demonstrated the pivotal role of vascular endothelial growth factor (VEGF) in interfering (i) the migration of antigen-specific T cells from the vessel into tumor and (ii) the recognition of cancer cells by cytotoxic T cells [26‐28]. Generally, the angiogenesis score is negatively correlated with immune infiltration; however, we observed an overlap of greater infiltration of CD8⁺ T cell and higher angiogenesis signature in the methyl-risk low group (Fig. S5). We further calculated the signatures concerning the comparisons between naïve, effector, and exhausted CD8⁺ T cells (Fig. S6) and observed that the scores of (i) naïve vs. effector (P = 0.089) and (2) naïve vs. exhausted (P = 0.089) trended higher in the methyl-risk high subgroup, and the effector vs. exhausted score was similar in the methyl-risk high and the methyl-risk low subgroups (P = 0.58). These results indicate that although the samples with lower methyl-risk had more effector and exhausted CD8⁺ T cells compared to naïve CD8⁺ T cells, the quantities of effector and exhausted CD8⁺ T cells were comparable, reflecting an adaptive resistance to immune attack, which might be partially due to the higher level of angiogenesis and potentially benefit from the regimens inhibiting VEGF and immune checkpoints.

Given the above findings from the TCGA-CHOL cohort, we sought to partially verify them in the EHSH cohort in terms of PD-L1 protein expression and tumor-infiltrating CD8⁺ T cells. A lower methyl-risk was associated with PD-L1 positivity (P = 0.040, Fig. 6A). The quantity of tumor-infiltrating CD8⁺ T cells could be reflected by CD8A methylation, as significant correlations were revealed between higher CD8A methylation, lower CD8A mRNA expression, and fewer tumor-infiltrating CD8⁺ T cells in the TCGA-CHOL cohort (Fig. 6B). In the EHSH cohort, we observed higher methylation levels of the five DMRs related to the CD8A gene in clusters 1/2/3 (P < 0.001, Fig. 6C), indicating poorer infiltration of CD8⁺ T cells in the clusters with larger methylation changes. Similar trends of the above results were shown in both the CCAs and the GBCs in the EHSH cohort, indicating the associations of methylation changes with PD-L1 expression and tumor-infiltrating CD8⁺ T cells might be a common feature in all BTCs. Given all the above immune-related results from two independent cohorts, fewer methylation changes might potentially be a novel predictor of better response to immunotherapy.

Distinguished by the density of tumor-infiltrating lymphocytes (TILs), there are two causes of PD-L1 expression: (i) prior attack by immunity (adaptive resistance) and (ii) activation of the oncogenic pathway (intrinsic induction) [29] respectively predicting better and worse benefit from immune checkpoint inhibitors (ICIs) in non-small cell lung cancer [30]. In our cohort, all PD-L1⁺ samples in clusters 1/2 had a high level of CD8A methylation (above median), and all PD-L1⁺ samples in clusters 3–6 had a low level of CD8A methylation (below median, P < 0.001, Fig. 6D). In addition, a large difference of OS with borderline significance was uncovered between the PD-L1⁺/CD8A^methyl-high and the PD-L1⁺/CD8A^methyl-low groups (HR = 4.96, 95% CI 0.59–41.75, P = 0.104, Fig. 6E), indicating the differed prognostic effects of the PD-L1 elicited by immune attack and the one induced by the oncogenic pathway. To identify the possible “PD-L1-inducing oncogenic pathway,” we revealed several enrichments of mutational event in the eleven PD-L1⁺/CD8A^methyl-high samples, including the mutations in BRAF (proportion = 4/11, OR = 13.1, 95% CI 2.43–71.0, P = 0.005), MMR pathway (proportion = 4/11, OR = 7.66, 95% CI 1.66–35.3, P = 0.016), Fanconi pathway (proportion = 4/11, OR = 5.31, 95% CI 1.24–22.7, P = 0.035), NOTCH pathway (proportion = 6/11, OR = 3.60, 95% CI 0.98–13.2, P = 0.053), Hippo pathway (proportion = 8/11, OR = 6.06, 95% CI 1.47–25.0, P = 0.010), and SWI/SNF pathway (9/11, OR = 9.59, 95% CI 1.92–48.0, P = 0.002). Within the SWI/SNF pathway, the mutations of BAF complex (ARID1A: 5/11, OR = 5.17, P = 0.024, and ARID1B: 4/11, OR = 13.1, P < 0.001) rather than those of PBAF complex (ARID2: 3/11, OR = 2.63, P = 0.194; PBRM1: 2/11, OR = 1.78, P = 0.395) were enriched in the PD-L1⁺/CD8A^methyl-high samples. Our results might provide novel insights into the molecular basis underlying the intrinsic induction of PD-L1 in BTCs.

Getting back to the prognostic effect of methylation changes, in order to determine whether the effect of methylation on OS is mediated by immune characteristics, we performed multivariable analysis to adjust for CD8A methylation and PD-L1 positivity. The association of the cluster-based risk and the LASSO-based risk with OS remained significant (Fig. 6G), indicating that the prognostic effect of DNA methylation might be partially but not completely on account of the infiltration of CD8⁺ T cells and the immune escape via PD-L1.