Integrative molecular characterization of gallbladder cancer reveals microenvironment-associated subtypes

Abstract

Background & Aims:

Gallbladder cancer (GBC) is the most common type of biliary tract cancer, but the molecular mechanisms involved in gallbladder carcinogenesis remain poorly understood. In this study, we applied integrative genomics approaches to characterize GBC and explore molecular subtypes associated with patient survival.

Methods:

We profiled the mutational landscape of GBC tumors (whole-exome sequencing on 92, targeted sequencing on 98, in total 190 patients). In a subset (n=45), we interrogated the matched transcriptomes, DNA methylomes and somatic copy number alterations. We explored molecular subtypes identified through clustering tumors by genes whose expression were associated with survival in 47 tumors and validated subtypes on 34 publicly available GBC cases.

Results:

Exome analysis revealed TP53 was the most mutated gene. The overall mutation rate was low (median 0.82 Mut/Mb). APOBEC-mediated mutational signatures were more common in tumors with higher mutational burden. Aflatoxin-related signatures tended to be highly clonal (present in ≥50% of cancer cells). Transcriptome-wide survival association analysis revealed a 95-gene signature that stratified all GBC patients into three subtypes that suggested an association with overall survival post-resection. The two poor-survival subtypes were associated with adverse clinicopathologic features (advanced stage, pN1, pM1), immunosuppressive microenvironments (myeloid-derived suppressor cell accumulation, extensive desmoplasia, hypoxia) and T cell dysfunction, while the good-survival subtype showed the opposite features.

Conclusion:

These data suggest that the tumor microenvironment and immune profiles could play an important role in gallbladder carcinogenesis and should be evaluated in future clinical studies, along with mutational profiles.

Graphical Abstract

Lay summary:

Gallbladder cancer is highly fatal, and its causes are poorly understood. We evaluated gallbladder tumors to see if there were differences between tumors in genetic information like DNA and RNA. We found evidence of aflatoxin exposure in these tumors, and immune cells surrounding the tumors were associated with survival.

INTRODUCTION

Gallbladder cancer (GBC) is the most common type of biliary tract cancer (BTC) and particularly more prevalent in women.[1, 2] GBC incidence varies geographically; for instance, it is high in Chile and Northern India, intermediate in Shanghai, China, and rare in Western, industrialized countries.[3] Five-year survival rates across all tumor stages range from 5% to 20%.[3, 4] The majority of gallbladder carcinomas arise in patients with gallstones.[5] Gallstones are nearly always accompanied by chronic inflammation in the gallbladder.[5] Many addition proposed risk factors are also related to inflammation.[6] For example, GBC is strongly related to obesity, which is increasingly considered an inflammatory disorder.[7] Other risk factors, such as bacterial infections with Salmonella enterica serovar Typhi [8] and aflatoxin B1 (AFB1),[9, 10] could also act through inflammatory pathways.

GBC is an intractable malignancy and designated an orphan disease. Molecular understanding of this cancer is limited and has focused on somatic mutational studies.[11–15] The most common features of GBC include mutations in TP53, enrichment of APOBEC-mediated mutational signatures and mutations in the ErbB pathway. However, the results of such studies have had limited impact on understanding the molecular basis of aggressive disease and cancer progression. These facets of GBC pathobiology are particularly important because, given its status as a rare disease, screening of the general population will unlikely be feasible even if adequate biomarkers are identified and, therefore, GBC will perpetually be diagnosed at variable disease stages. Most studies have focused on characterizing differences between biliary tract cancers (BTCs), including intrahepatic cholangiocarcinoma (iCCA), extrahepatic cholangiocarcinoma (eCCA) and GBC, and not on GBC, specifically. The relatively low mutation rates in GBC [12, 13] creates an important challenge for understanding GBC, which could be the foundation for stratification in clinical decision-making.

In this study, we investigated molecular determinants of poor-survival GBC using an integrative approach. Accordingly, we analyzed 190 GBC tumors to characterize the mutational landscape and, in a subset of tumors, integrated matched transcriptome, DNA methylation and copy number alteration data to explore GBC prognostic subtypes based on their molecular profiles. We observed three survival-associated subtypes of GBC with distinct clinical features, tumor microenvironments and immune phenotypes.

MATERIALS & METHODS

Study population

We obtained fresh frozen tumor tissues and paired samples for germline DNA analysis from two NCI studies: the Shanghai Biliary Tract Cancer Study and the Chile Gallbladder Cancer Study. As previously described,[16] the Shanghai Biliary Tract Cancer Study recruited newly diagnosed patients with primary biliary tract cancer (ICD 9 code: 156) aged 35 to 74 years who were permanent residents of urban Shanghai. These patients were identified through a rapid-reporting system established between the Shanghai Cancer Institute and 42 collaborating hospitals in 10 urban districts of Shanghai. Between June 1997 and May 2001, a total of 368 patients diagnosed with GBC were recruited, representing over 90% of all eligible GBC cases. GBC diagnoses were confirmed by an international panel of board-certified pathologists. In addition, H&E FFPE sections were separately reviewed for the current analysis by board-certified pathologists and classified according to WHO IV guidelines. Snap frozen tissue samples were collected from 72 GBC cases, of whom 61 had paired germline DNA or blood clot available. Whole-exome sequencing (WES) was successfully completed for 52 of these cases.

As previously described,[9, 17] the Chile Gallbladder Cancer Study recruited 52 patients with GBC who were newly diagnosed at cancer referral hospitals in Santiago, Concepción, and Temuco, Chile. The participation rate was 85% for patients with GBC in the Chile study. Snap frozen tumor tissue samples were collected from 9 patients with GBC who had paired germline DNA available, of whom 8 had adequate material for WES. In total, WES was performed on 60 patients (paired tumor tissue and germline DNA; 52 Chinese and 8 Chilean) from the Shanghai and Chile studies combined. Of those, RNA-seq was successfully performed on RNA from 47 fresh-frozen tumor tissues and 5 adjacent non-tumor (ANT) tissues, DNA methylation (Infinium EPIC Chip) on tumor DNA from 59 tumors (46 with matched RNA-seq), and copy number evaluation (Infinium GSA-24v1 array) on tumor DNA from 58 samples (45 with matched RNA-seq). All participants in both studies provided written consent for biospecimen and questionnaire data collection. Institutional review boards at the NCI and the Shanghai Cancer Institute approved the Shanghai study. Institutional review boards at the NCI, Pontificia Universidad Católica de Chile, and Chilean Ministry of Health approved the Chile study.

Whole-exome and targeted sequencing

WES sequencing was conducted at the NCI Cancer Genomics Research Laboratory (CGR) with 100X/v3+UTR for tumor tissue samples and 40X/v3+UTR for the matched normal control samples. All samples were sequenced using the Illumina HiSeq 2500 with v4 chemistry at 2×125bp reads. Exome capture was carried out using the Nimblegen SeqCap EZ Exome + UTR capture kit (Roche). Library templates were prepared using the Kapa HyperPlus kit (Roche). The prevalence of top driver genes was validated in a set of 100 GBC cases from Temuco, Chile [79% female (79/100), median age: 66 (37–91)]. DNA was extracted from FFPE tissue sections using Omega Mag-bind FFPE kit and quantified using Qubit. Targeted DNA sequencing was conducted for a panel of 477 genes using Nimblegen capture at an average coverage of 267 (range 138–352) and Illumina sequencing in 98 patients with adequate DNA extracted.

Integration of publicly available data

Publicly-available data from previous studies [12, 13] (n=66) were included where appropriate. Matched tumor/normal tissue WES data from 32 Chinese patients [13] were downloaded in the mapped BAM format and combined with WES data from the 60 patients from NCI studies described above. We downloaded a total of 34 RNA-seq data of GBC tumors (without matched normal) from three different studies [12, 18, 19]. Raw data were downloaded in FASTQ format and quantified the genes expression levels using Kallisto [20]. We used the ‘normalizeQuantiles’ function from the limma package to correct for batch effects [21]. The corrected gene expression levels were used for independent evaluation of a survival-associated transcriptional signature.

Genomic profiling and analysis

For further information on additional genomics (targeted DNA sequencing, RNA-seq, copy number array, DNA methylation array; Figure 1), patient characteristics and downstream analyses, please refer to Supplementary Materials & Methods.

Figure 1.

Overview of GBC study, genomics platforms and analytical pipelines.

RESULTS

Mutational landscape of GBC

To characterize the landscape of genomic aberrations in GBC, we performed WES on matched tumor and normal samples from 60 patients in the National Cancer Institute (NCI) Chile and Shanghai case-controlled studies [9, 16, 17] and included publicly-available data on 32 Chinese matched tumor and normal WES samples [13] for a total of 92 patients with GBC. The GBC mutational landscape was further evaluated by targeted sequencing (TS) of 477 genes in an additional 98 GBC patients enrolled from Chile (Figure 1 and S1A). The WES and TS cohorts did not significantly differ in age (P=0.1825, Wilcoxon test) or sex (P=1, Fisher test). For the NCI study patients, we also performed RNA-seq, DNA methylation (Infinium EPIC arrays), copy number (Infinium Global Screening array (GSA)) and APOBEC genotyping of the tumor samples, as well as circulating lipopolysaccharide (LPS) and aflatoxin B1-lysine adduct (AFB1 adduct) detection (Figure 2A; Supplementary Table S1). Of the 92 patients, 8.7% (8/92) patients were from Chile, and the remaining 91.3% (84/92) patients were Chinese (Shanghai) (Figure 2B). Seventy-one percent (65/92) of the patients were female. The median age at diagnosis was 64 years (range 36–86), and the median overall survival was 8.0 months (range 0.4–62.4). Among patients with data on stage available, 23% (20/86) of patients were early stage (AJCC stage 1–2) and 77% (66/86) were late stage (AJCC stage 3–4).

Figure 2.

Genomic alteration landscape of gallbladder cancer (GBC) patients revealed by muti-omic profiling of 92 patients. (A) The status of different types of omic profiling (exomeseq, RNA-seq, DNA methylation array (Infinium EPIC) and copy number array (Infinium GSA). (B) Description of sex, ethnicity (84 Chinese and 8 Chilean) and presence or absence of circulating aflatoxin-lysine adducts in GBC patients. (C) Distribution of somatic nonsynonymous mutations. Patients are ordered based upon increasing number of mutations. Y-axis represents mutation per megabase (Mut/Mb) of coding sequences. (D) Different types of mutations among the significantly mutated genes across GBC patients. P-values for each of these genes, as reported by MuSic, are plotted on the left.

We analyzed 92 GBC WES tumor/normal pairs (with an average coverage of 150x) (Figure 2A) and annotated 4,277 high-confidence nonsynonymous mutations (called by three algorithms with QC, details in Supplementary Materials); 3,845 were single nucleotide variants (SNVs), and 432 were small insertion/deletions (indels) (Supplementary Table S2). The number of nonsynonymous mutations ranged from 9 to 566 per patient with a mean of 1.34 and a median of 0.82 mutations per megabase (Mut/Mb) (Figure 2C). We identified only one hypermutated patient (>10 Mut/Mb) with somatic missense mutations in POLD1, a gene that is associated with the DNA mismatch repair (MMR) pathway. No cases of high confidence micro satellite instability (MSI) were detected. Clinically, aflatoxin detection (FDR-P=0.03, likelihood ratio test (LTR)) and tumor stage (FDR-P=0.004, LTR) were associated with higher mutational burden, while gallstones (FDR-P<0.001, LTR) and obesity (FDR-P<0.03, LTR) were associated with lower mutational burden (Supplementary Table S3). However, including all four of these covariates in a multivariate model, only gallstones remained statistically significant (effect size=−1.24, suggesting that patients with gallstones on average have 1.24 less nonsynonymous mutations than ones without gallstones, 95%CI=−1.8,−0.6, P<0.001, LTR).

Our analysis identified a significant burden of mutations in the tumor suppressor p53 (TP53), namely 29.3% (27/92) of patients. TP53 mutations were distributed across the gene, providing clear evidence of highly recurrent somatic mutations (Figure S1B). Out of 27 TP53 mutations, 93% (25/27) of them are predicted to be deleterious mutations (Supplementary Table S2). In addition, 96% (26/27) of TP53 mutations overlaped with previously reported hotspot loci within the gene [22]. Consistent with previous findings, no significant associations were observed between TP53 mutations and circulating AFB1 adduct status (P=0.23, Fisher test) [23]. Targeted sequencing in an independent set of 98 GBC patients from Chile confirmed TP53 as the most recurrent mutation (21.4%; 21/98), and no R249S variants (the risk allele often present in aflatoxin-associated HCC tumors) were detected (Supplementary Fig S1A; Supplementary Table S4). In addition, significantly mutated genes were observed in kinase and cell cycle (ELF3, ERBB2, CDC27, TGFBR2, PIK3CA), immune-related (KIR2DL4, KIR2DL3) and chromatin remodeler (ARID2) genes (Figure 2D). These recurrent mutations suggest closer genetic similarity of GBC to extrahepatic cholangiocarcinoma (eCCA) compared to intrahepatic CCA (iCCA), though all these tumors are often considered collectively as biliary tract cancer.

Somatic mutation-targeted processes in GBC

We investigated whether the mutational landscape in GBC were enriched in genes representing specific pathways and identified multiple dysregulated pathways associated with GBC mutations (defined by KEGG), broadly classified into cell growth- and proliferation-related (cell cycle, apoptosis, WNT signaling, and mitogen-activated protein kinase (MAPK) signaling pathways) and immune system (antigen processing, graft-versus-host disease and intestinal immune network for IgA production) processes (Figure 3A; Supplementary Table S5). Almost a third of patients (29.3%; 27/92) had TP53 mutations within the p53 pathway affecting genes in pathways related to dysregulation of apoptosis, cell cycle and genome instability (Figure 3B). Immune-associated pathways involved “antigen-processing and presentation,” including mutations in major histocompatibility complex (HLA-I and HLA-II) and killer cell immunoglobulin-like receptor (e.g., KIR2DL4, KIR2DL3, as mentioned above). Notably, though ErbB signaling pathway has been reported as recurrently mutated in GBC,[13, 24] we only observed enrichment of this pathway among APOBEC-signature positive tumors (P=0.01) (Figure S2A).

Figure 3.

Significantly mutated pathways in 92 GBC tumors (84 Chinese and 8 Chilean). (A) List of significantly mutated KEGG pathways in gallbladder. Only pathways with a corrected p-value of ≤0.05 are shown. (B) Interactions between individual mutated genes of three major pathways (p53 pathway, WNT signaling pathway and MAPK signaling pathway) related to cell cycle, apoptosis and genomic instability, and other three pathways (antigen-processing and presentation, host versus graft disease and intestinal immune network for IgA production) related to immune response. These pathways included genes, such as CDKN2A and NF1, known to have recurrent somatic copy number alternations. The number inside the box indicates the number of patient with mutation in this gene. The percentage of affected patients with mutations in each pathway compared to all patients analyzed are denoted on top (Supplementary Table S5).

Mutational signatures in GBC

To evaluate underlying mutational processes, we analyzed patterns of somatic substitutions and observed C>T substitutions as the most dominant (40.5%), followed by C>A (20.6%) and C>G (20.5%) (Figure 4A–B). Trinucleotide matrix analysis revealed that T[C>G]W and T[C>T]W (W = A or T) trinucleotides associated with the APOBEC-mediated mutational signatures (signature 2 and signature 13 from COSMIC) were generally present in tumors with higher mutation burden (Figure 4C). Notably, tumors with any evidence of APOBEC signatures had a higher mutational load compared to tumors without APOBEC signatures (P<0.002) (Figure S2B). Aflatoxin signature (signature 24), which has been associated with aflatoxin exposure in HCC,[25] was observed in 39 patients with no significant enrichment in GBC patients with detectable levels of circulating AFB1 adducts (Figure 4A and C). Overall, the most common COSMIC signatures [26] were clock-like signatures 1 and 5 (mean proportions 36.0% and 26.4%, respectively), while other signatures were less common (APOBEC signatures 2 and 13, mean proportions 4.2% and 6.6%; smoking and tobacco chewing signatures 4 and 29, mean proportions 6.6% and 8.5%; aflatoxin signature 24, 8.8% and aristolochic acid exposure 22, 2.9%) (Figure 4D, S2C). Stratification of mutational signatures by clonality, defined by cancer cell fraction (CCF) >50% (high) or <50% (low), revealed signature 13 (APOBEC; P=0.001, by two-sided Wilcoxon signed-rank test; outliers as larger than 1.5*the interquartile range were removed) and signature 24 (aflatoxin; P=0.047) to be highly clonal among GBC tumors (Figure S2D).

Figure 4.

Somatic mutational signatures in 92 GBC tumors (84 Chinese and 8 Chilean). (A) GBC patients are classified into three groups based on circulating aflatoxin-lysine adduct information (detectable: ≥0.5 pg aflatoxin–lysine/mg albumin, undetectable, or no information) where patients in each group are ordered based on decreasing mutational load. (B) Contribution of six types of single base substitutions (SBS) for each patient. X-axis represents the fraction (scaled between 0–100) of each SBS type. (C) Heatmaps represent the distribution of 96 mutation types defined by the mutated nucleotide and its adjacent 5’ and 3’ nucleotide. Frequencies of trinucleotide substitution patterns are scaled between 0–1 for each patient. (D) Assignment of trinucleotide matrix to known mutational signatures from COSMIC database.

Recurrent gene fusions and survival-associated transcript clustering highlights GBC survival molecular subtypes

Though mutation-targeted pathways reflect important pathobiological processes in GBC, the relatively low mutational burden and extensive inter-tumor heterogeneity underscore the difficulty in interpreting mutations alone. Since transcriptional levels appear to be affected by mutations, as indicated by the the enrichment of mutations for “Transcriptional misregulation in cancer” pathway (Figure 3A), we performed RNA-seq on 47 primary patient tumors to identify recurrent gene fusions and GBC survival molecular subtypes.

We observed a novel recurrent fusion between the cytoskeletal gene, tubulin polymerization-promoting protein (TPPP), and the chromatin remodeler, bromodomain-containing protein 9 (BRD9) (Supplementary Figure S3A–B), a fusion previously identified in lung cancer.[27] This intrachromosomal fusion was detected in 19.1% (9/47) patients, with the breakpoint occurring in exon 1 of TPPP and intron 9 of BRD9. We further analyzed 34 RNA-seq data from 3 previous studies [12, 18, 19] and identified an additional fusion-positive case with the exact breakpoint. We validated the TPPP:BRD9 fusion using unique PCR primers and Sanger sequencing of the product in Chilean fusion-positive cases compared to tumors found negative for the fusion by all algorithms (Supplementary Figure S3C) Neither TPPP or BRD9 were differentially expressed between fusion-positive and fusion-negative tumors (Supplementary Figure S3D). However, the fusion eliminated the DUF3512 domain of BRD9, which is vital for the integrity of the SWItch/Sucrose Non-Fermentable (SWI/SNF) complex, and could have consequences for the corresponding protein interactome which includes the oncoproteins cyclin-dependent kinase 11B (CDK11B) and BRD4 (Supplementary Figure S3E).

Next, we conducted supervised clustering to explore the molecular basis of aggressive disease, as post-resection survival time can be highly variable (Supplementary Table S1). We performed transcriptome-wide univariate Cox proportional hazards modeling with Wald statistics. In total, we identified 95 transcripts whose expression was each associated with survival time following tumor resection (Supplementary Table S6). Hierarchical clustering of tumors using expression of these survival-related transcripts led to the identification of 3 clusters (Figure 5A), with significant difference in survival time between these clusters (P<0.0001, log-rank test; Figure 5B). Subtype 1 and subtype 3 tumors collectively comprised GBC associated with poor survival and subtype 2 tumors included patients with longer survival time post-surgery. Since our cohort includes patients of Chinese and Chilean ethnicities, we conducted a sensitivity analysis, excluding the Chilean cases, and observed no difference in survival subsets.

Figure 5.

Transcriptomic identification of GBC survival subtypes among 47 patients (40 Chinese and 7 Chilean). (A) Hierarchical clustering of 95 transcripts associated with overall survival of GBC patients. Survival-associated transcripts were identified by univariate Cox-proportional hazards modeling with Wald statistics. (B) Kaplan-Meier analysis of overall survival differences between GBC subtypes. (C) Clinicopathologic, epidemiologic and genomic parameter associations with GBC survival subtypes. *: P<0.05, **: P<0.01, ***: P<0.001. (D) Representative H&E for each of the histomorphological subtypes of GBC (biliary, gastric foveolar-like, and intestinal histology). The biliary morphology section shows poorly differentiated (arrows) and moderately differentiated (arrowheads) adenocarcinoma intermixed zones, whereas intestinal morphology sections exhibit well-developed glands (arrows) lined by medium-high or high neoplastic epithelium, respectively. (E) Comparison of mutation count and copy number alterations (CNA) count across GBC survival subtypes (Kruskal-Wallis test). ns: not significant.

We evaluated the association between survival subtypes with key clinicopathologic, epidemiologic and genomic features (Figure 5C, Supplementary Table S7). When comparing tumors from patients with good (subtype 2) versus poor survival (subtypes 1 and 3), the poor-survival subtypes were associated with advanced AJCC disease stage (P=0.0002) whereas the good-survival subtype was associated with gastric foveolar histomorphology (P<0.04, Fisher test) (Figure 5C–D). Histomorphology was confirmed by immunohistochemical staining for a subset of cases (Supplementary Table S8). Further, the TPPP:BRD9 fusion trended towards enrichment in survival subtype 2 (66.7%, (6/9) fusions; P<0.1) (Figure 5C). Subtype 3 tumors were associated with smoking history (P=0.03) and more advanced tumor stage (pTNM, P=0.001) compared to the other subtypes. GBC survival subtypes were not associated with sex, ethnicity, alcohol consumption, aflatoxin status, gallstones, chronic hepatitis B virus (HBV) infection (as measured by HBV surface antigen seropositivity) or obesity. Further, these subtypes were not associated with aflatoxin exposure (as measured by the intensity of signature 24), APOBEC mutational signatures, mutations in chromatin remodelers and DNA repair genes, hypermutation and TP53 mutations. Globally, GBC molecular subtypes did not significantly differ by mutation burden, tumor allele frequency or copy number alterations (Figure 5E, Supplementary Figure S4, Supplementary Table S9).

Molecular pathways associated with GBC survival subtypes

To investigate the underlying biological differences between different GBC survival subtypes, we performed gene set enrichment analysis (GSEA) between the good- and poor-survival molecular subtypes (Figure 6A, Supplementary Table S10). GBC tumors of the good-survival subtype had higher expression of diverse metabolic pathways (defined by KEGG). In contrast, the poor-survival subtype tumors displayed high expression of MAPK and TGF-β pathway members and immune pathways, such as B cell receptor signaling and cytokine-cytokine receptor interaction (Figure 6A), as well as key microenvironmental processes (defined by Hallmarks gene sets of MSigDB) (Figure 6B), specifically epithelialmesenchymal transition (EMT; P<0.001), angiogenesis (P<0.001), and hypoxia (P=0.04). In an independent sample of 34 GBC cases with publicly-available data[12, 18, 19], hierarchical clustering of the same 95-gene transcripts identified a similar ‘poor survival-like’ subgroup (65%; 22/34) (Supplementary Figure S5A–B, Supplementary Table S11). This poor survival-like subgroup recapitulated many cancer hallmarks observed in our own patients, such as ‘IL6 JAK STAT3 signaling’, ‘inflammatory response’, ‘KRAS signaling’, ‘epithelial mesenchymal transition’, ‘angiogenesis’ and ‘TNFA signaling via NKFb’.

Figure 6.

Distinct microenvironment properties differentiate GBC survival subtypes among 47 patients (40 Chinese and 7 Chilean). (A) Expression heatmap of significant leading edge genes among KEGG processes significantly enriched in good- or poor-survival tumors, as determined by GSEA. (B) Enrichment plots of Hallmark processes significantly enriched in poor-survival tumors, as determined by GSEA. NES: normalized enrichment score. (C) Relative abundance of 36 distinct cell types according to xCELL in good- versus poor-survival GBC. (D) Inter-GBC subtype differences in cancer-associated fibroblast (CAF) content as computed by TIDE algorithm. (E) Inter-GBC subtype differences in T cell exclusion potential as computed by TIDE algorithm. (F) Inter-GBC subtype differences in T cell dysfunction potential as computed by TIDE algorithm. (G) Inter-GBC subtype differences in T cell dysfunction potential as computed by TIDE algorithm. *: P<0.05, **: P<0.01, ***: P<0.001, ****: P<0.0001.

Comparison of immunogenicity, cellular composition and immune checkpoint regulators across GBC survival subtypes

Based on microenvironmental pathway enrichment in poor-survival GBC (subtype 1 and 3), we evaluated the relative immunogenicity of GBC across survival subtypes. We found no significant difference in neoantigens generated by somatic mutations and mRNA expression of cancer germline antigens (details in Supplementary Material). We used xCELL to infer the abundance of diverse cell types of GBC tumors (Figure 6C), and no significant differences were observed in B cells, T cells or most immune cells. However, the good-survival GBC subtype had a greater abundance of epithelial cells (P=0.02, Wilcoxon test), which is consistent with the Tumor Immune Dysfunction and Exclusion (TIDE)-based prediction[28] of decreased cancer-associated fibroblasts (CAFs) in good-survival tumors compared to poor-survival tumors (Subtype 1: P=0.0002; Subtype 3: P<0.0001; Figure 6D). In contrast, poor-survival GBC had increased endothelial cell content, as well as elevated numbers of myeloid-derived suppressor cells (MDSCs), which have highly potent immunosuppressive activity. We further compared gene expression of immune checkpoint inhibitors and found significant upregulation of PDCD1 (PD1), CD274 (PDL1), CTLA4 and CD80 in subtype 3 versus subtype 2 (P=0.01, 0.01, 0.007 and 0.002, respectively, Supplementary Figure S6A). This elevated expression of immune checkpoint inhibitors suggest that subtype 3 gallbladder tumors could be immune-suppressed.

Immune-related biomarkers in serum and bile associated with GBC survival subtypes

Since survival subtypes displayed possible differential immunity, we assessed associations between immune-related biomarkers and survival subtypes. We analyzed 68 immune-associated proteins in serum and bile of GBC patients (Supplementary Table S12) using two multiplex platforms: the Milliplex assay (EMD Millipore, Billerica, MA, USA; 61 markers) and the Meso Scale Discovery (MSD) Human Vascular Injury II kit (Meso Scale Diagnostics LLC, Rockville, MD, USA). Uniquely, bile profiles of subtype 1 tumors displayed lower soluble interleukin 4 receptor (sIL4R; P=0.02, Wilcoxon test) and sIL6R (P=0.047), as well as higher chemokine (C-X-C motif) ligand 6 (CXCL6; P=0.048), when compared with other subtypes, while bile levels of serum amyloid A (SAA) were elevated in the good-survival subtype (P=0.03) and lowest in the most aggressive subtype (subtype 3) patients when compared with all others (P=0.04). Among 64 serum proteins tested, the most aggressive subtype (subtype 3) displayed elevated systemic expression of the soluble CD40 ligand (sCD40L/sCD154, P=0.04), a platelet activation marker associated with endothelial interaction, which is consistent with the higher endothelial cell content we observed in poor survival tumors (Figure 6C), and matrix metalloproteinase activity. sCD40L likely comes from activated T cells since CD40L is predominantly expressed on activated T cells and is the ligand that binds to CD40 on antigen-presenting cells. Subtype 3 also had decreased expression of the obesity-related cytokine RESISTIN (P=0.048) and epidermal growth factor (EGF; P=0.02).

Functionality of the immune cells differs across GBC survival subtypes

Because we observed between-GBC-subtype alterations of immune-related biomarker levels in serum and bile and of immune checkpoint inhibitor gene expression in tumors but generally did not observe overt alterations in immune cell populations, we considered whether the functionality of the immune cells present around the tumor could underpin GBC survival subtypes. We computed the TIDE score[28], a multi-metric transcriptional signature demonstrated to outperform mutational burden and checkpoint inhibitor expression in predicting immune checkpoint blockade response in clinical trials across multiple cancer types.

One such functional factor is tumor T cell exclusion, which TIDE predicted to be significantly higher in subtype 1 (P<0.0001, Wilcoxon test) and 3 (P=0.0005) tumors compared to subtype 2 tumors (Figure 6E, Supplementary Figure S6B). To confirm the prediction of T-cell exclusion differences by survival subtype, we annotated regions of tumor versus non-tumor on digital H&E images from cases with adequately stained material available. We then ran a tumor infiltrating lymphocyte algorithm on the images and found that subtype 2 had higher levels of infiltrates than subtype 3 (P=0.02, Supplementary Figure S6C–D). Subtype 1 was not significantly different than the other two groups but only included 4 cases. These findings suggest T-cell exclusion in GBC tumors could be associated with subtypes with poor survival.

Another possible tumor-extrinsic factor is T cell dysfunction. Subtype 3 tumors exhibited significantly higher T cell dysfunction scores, as predicted by TIDE, when compared to subtype 1 (P=0.005) and subtype 2 (P=0.03) tumors (Figure 6G, Supplementary Figure S6E).

Considering both T-cell exclusion and T cell dysfunction, subtype 2 tumors displayed significantly lower TIDE scores when compared with subtype 1 (P<0.0001) or subtype 3 (P=0.0002) (Figure 6F), suggesting that subtype 2 patients have higher likelihood of response to immune checkpoint blockade in spite of expressing lower levels of checkpoint inhibitors.

Microenvironment-associated DNA methylome remodeling in GBC subtypes

To determine the extent to which epigenomic differences characterize GBC subtypes, we performed DNA methylation profiling on 46 tumors (Figure 1). Differentially methylated probe (DMP) calling (based on the RnBeads-integrated limma method followed by empirical Bayes-guided model fitting) revealed 73867 DMPs between all subtypes (Supplementary Table S13, Supplementary Figure S7A). For each subtype, DMPs displayed unique patterns in overall hypermethylation:hypomethylation ratio (Supplementary Figure S7B), overlap with CpG islands (Supplementary Figure S7C) and enhancers (Supplementary Figure S7D). As a large fraction of DMPs overlapped with enhancers, we predicted putative target genes of enhancers by correlating expression levels of genes and methylation levels of enhancers (see methods). At the defined significance level (FDR-adjusted P < 0.05; r ≤ −0.3), we identified 788 enhancer:gene pairs where genes expression level were negatively correlated with enhancer methylation (Supplementary Figure S7E). Interestingly some putative driver genes (ERBB2, ERBB3, ELF3, TGFBR2) appear to be under enhancer regulation, thus highlighting another layer of regulation in these key driver genes. Pathway analysis of the predicted target genes of these aberrantly methylated enhancers revealed over-representation of metabolic, immune and structural pathways and processes (Supplementary Figure 7F), reflecting inter-subtype differences observed at the mRNA level.

Tumors in subtype 3 were most distinct by epigenetic analyses (Supplementary Figure S7A); displayed higher methylation at DMPs (P<0.01, Wilcoxon test) with the lowest inter-tumor variability (Figure 7A–B). These data indicate that subtype 3 tumors are characterized by relatively homogenous hypermethylated profiles at these key loci. To infer whether such association between hypermethylation and subtype was influenced by differential expression of epigenetic writer and/or eraser genes, we compared mRNA expression of DNA methyltransferases (DNMTs) and Ten-eleven translocation methylcytosine dioxygenases (TETs) (Figure 7C–D). DNMT1 was increased in subtype 3 compared to subtype 2 (P<0.02); DNMT3A was increased in subtype 1 compared to subtype 2 (P<0.03); and DNMT3B was increased in subtype 1 (P<0.02) and subtype 3 (P<0.02) compared to subtype 2. In contrast, neither TET1, TET2 or TET3 were differentially expressed between subtypes. Further, no DNMTs or TETs were significantly correlated with methylation levels at DMPs in GBC patients (Supplementary Figure S8), indicating that epigenetic writer and/or eraser expression are not the predominant causes of such epigenomic profiles.

Figure 7.

DNA methylome remodelling in 46 GBC tumors with good (subtype 2) or poor survival (subtypes 1 and 3). (A) Comparison of intra-subtype mean methylation and variance of 73,867 differentially methylated positions (DMPs) in GBC patients. (B) Comparison of median methylation levels of 73,867 DMPs between GBC subtypes. (C) Differential mRNA expression of DNA methyltransferases (DNMTs) between GBC subtypes. (D) Differential mRNA expression of Ten-eleven translocation methylcytosine dioxygenases (TETs) between GBC subtypes. (E) Spearman correlation of hypoxia score versus mean methylation among 73,867 DMPs. Hypoxia score was defined by the sum of the leading edge genes within hypoxia gene set, as defined by Hallmarks. Black represents the correlation between the hypoxia score and mean methylation across the three subtypes. Red represents the correlation between hypoxia score and mean methylation within subtype 1, green within subtype 2, and purple within subtype 3. ns: not significant.

We previously observed that the poor-survival tumors (subtype 1 and subtype 3) are associated with hypoxia (Figure 6B), which correlates with DNA hypermethylation, presumably through impeding the oxygen-dependent enzymatic activity of TETs.[29] Accordingly, we computed a hypoxia score in GBC tumors and observed that the degree of hypoxia significantly correlated with the methylation level of DMPs between subtypes (r=0.5, P=0.0007, Spearman test), but not within subtypes (Figure 7E). Collectively, these data implicate hypoxia as a feature of the most aggressive subtype 3 tumors, which could contribute to impaired epigenome homeostasis and altered expression of key microenvironment processes.

DISCUSSION

To advance our understanding of the molecular basis of GBC, we performed multi-omics characterization among GBC patients. We identified significantly mutated genes, somatic mutation-targeted pathways and recurrent gene fusions. In addition, we found that clock-like signatures (COSMIC signature 1 and 5) contribute to about half of the point mutations. Although aflatoxin (signature 24) contributed a relatively small proportion of mutations, these mutations are likely to be clonal, defined by a prevalent CCF higher than 50% within tumors positive for this signature. Highly recurrent TPPP:BRD9 fusion was jointly identified by three gene fusion methods. This fusion has been reported previously for non-small cell lung cancer [30] but not in other cancer types or normal tissues. Future investigion is warranted to rule out the possibility of technical errors and verify its frequency in GBC population. Based on 95 transcripts, we identified molecular subgroups associated with survival. Our results suggest that microenvironment-associated features and changes in the local immune function could be key contributors to GBC patient survival.

Notable strengths of this study include matched genomic, serum, bile and clinical data. The findings are intriguing and warrant future investigation. We supported our genomics-based findings through review of digital H&E images, finding higher levels of immune infiltrates and gastric foveolar histomorphology in subtype 2, which was associated with good survival. In addition, we investigated whether immune-related markers in blood and bile supported our observations in the tissue; we found higher expression of the inflammatory marker and soluble variant of the T cell ligand sCD40L in serum of poor-survival subtype 3 compared to other GBC subtypes, suggesting that local tumor immunosuppression could have important implications. A previous study observed associations between some additional serum markers and survival [31] that we did not find to be associated with subtype, suggesting that additional research in larger studies is needed. Although the sample size is limiting, subclustering and multiple comparisons require interpretation with caution. Additionally, this integrative genomics study was not designed for investigating the utility of immunotherapy and thus, our observations warrant subsequent clinical evaluation, including incorporating immunohistochemistry and genomics into current and future exploratory trials.

Numerous risk factors are known to be associated with GBC, but their specific contribution to somatic mutations and corresponding cancer genome profiles remains unknown. Here, we observed that mutational load was negatively associated with gallstone-associated GBC. Further, though we previously identified a 3-fold increased risk of GBC following aflatoxin exposure,[23] we did not observe major contribution of aflatoxin on mutation signatures, rates or profiles. Although the TP53-R249S hotspot mutation is detected in approximately 45% of aflatoxin-associated HCC tumors,[25] we did not observe this mutation in GBC tumors from patients with detectable circulating AFB1 adducts. The lack of R249S mutations in these GBC patients may not be surprising since few cases had chronic HBV infection, and the R249S mutation seems to arise from interaction with HBx protein.[32] High mutational load and enrichment of G[C>A]N trinucleotides (the aflatoxin signature) were previously reported as dominant features of aflatoxin-associated HCC.[25] We did not observe higher mutational loads or enrichment of the aflatoxin signature in GBC patients with detectable circulating AFB1 adducts, potentially be due to exposure misclassification as circulating AFB1 adduct levels only indicate recent (~3 weeks) exposure. Alternatively, differences in expression of AFB1-metabolizing enzymes between GBC and HCC could result in differences in genomic exposure to the adduct-forming metabolite, AFB1 exo-8,9-epoxide, thereby explaining the absence of the previously reported HCC-associated AFB1 mutational signature in GBC.

Though increased tumor infiltration of specific T cell subpopulations has previously been linked to extended survival in GBC patients,[33] here we observed no significant association between T cell infiltrates and survival-based subtype, suggesting other cell populations may be important. For example, we observed higher levels of MDSCs in tumors with poor survival. MDSCs are potent immunosuppressive cells that are low in healthy individuals and support Treg production and the conversion of fibroblasts to cancer-associated fibroblasts,[34] consistent with our findings of higher CAF content in the poor-survival GBC. Taken together, these results suggest that immunosuppression and microenvironment remodeling could contribute to poor patient survival and possibly poor response to immune checkpoint blockade.[35, 36]

Immunotherapy has improved survival for select cancers, though only a small proportion of patients respond to immunotherapies.[37] Our results suggest that altered microenvironment, including immune signaling and functionality, appears to be a defining feature of poor survival following surgery for GBC. Although results must be interpreted with caution given the small sample size, these findings build further upon a pathobiological subgroup tentatively alluded to by Nakamura and colleagues, who reported a poor prognosis subgroup to have distinct immune gene expression.[12] However, given the economic costs of immunotherapy, as well as the narrow survival windows for patients with advanced GBC, robust selection tools to predict responder patients are needed for GBC. In this study, we carried out integrative molecular profiling of GBC and identified survival subgroups associated with tumor microenvironment features. Future biomarker-guided combination therapies evaluating the potential of epigenetic agents to reinvigorate T cells and enhance response to immune checkpoint blockade,[38] as well as therapies directed against immune checkpoint blockade -confounding microenvironment features, could be critical to adapt immunotherapy for advanced GBC.

Highlights:

• We profiled the mutational landscape of GBC tumors.

• Aflatoxin-related signatures tended to be highly clonal.

• We identified a novel TPPP-BRD9 fusion.

• Poor-survival subtypes were associated with immunosuppressive microenvironments.

• Tumor microenvironment could play an important role in gallbladder carcinogenesis.