The Immunogenomic Landscape of Resected Intrahepatic Cholangiocarcinoma

Abstract

Cholangiocarcinoma is a deadly and highly therapy-refractory cancer of the bile ducts, with early results from immune checkpoint blockade trials showing limited responses. While recent molecular assessments have made bulk characterizations of immune profiles and their genomic correlates, spatial assessments may reveal novel actionable insights. Here, we have integrated immune checkpoint-directed immunohistochemistry with next-generation sequencing of resected intrahepatic cholangiocarcinoma (iCCA) samples from 96 patients. We find that both T-cell and immune checkpoint markers are enriched at the tumor margins compared to the tumor center. Using two approaches, we identify high PD-1 or LAG3 and low CD3/CD4/ICOS specifically in the tumor center as associated with poor survival. Moreover, loss-of-function BAP1 mutations are associated with and cause elevated expression of the immunosuppressive checkpoint marker B7H4. In conclusion, this study provides a foundation on which to rationally improve and tailor immunotherapy approaches for this difficult-to-treat disease.

Intrahepatic cholangiocarcinoma (CCA) is the second most common primary liver cancer world-wide (1). The low survival rate, few therapeutic options for CCA, and increasing global incidence make it an urgent unmet clinical need (2). Despite recent broad advances in cancer therapy, the treatment of CCA still primarily rests on surgery and chemotherapy, with gemcitabine plus cisplatin as the systemic standard of care, showing limited efficacy (3). FGFR and IDH1 inhibitors have recently been investigated for CCA with encouraging results including approval of the FGFR inhibitor pemigatinib, but these alterations occur in <10% of all CCA patients. Immune checkpoint inhibitors have led to unprecedented clinical benefit in high mutation burden cancers such as melanoma, lung, and bladder, but responses have been limited in low mutation burden cancers including CCA: phase II clinical trials have indicated a 6–17% response rate with anti-PD-1 agents for CCA (4, 5).

The immune landscape of CCA also remains poorly defined and this has limited the optimization of immunotherapeutic approaches. Recent transcriptomic studies have shed some early light on CCA immune profiles (6, 7), but these have been limited to bulk tumor/stroma analyses without immune cell localization data. These studies are also hindered by the high degree of clinical, histological, and genomic heterogeneity among intrahepatic, perihilar, and extrahepatic CCA; for example, FGFR2 fusions and IDH1/2 mutations are almost entirely exclusive to intrahepatic while ERBB2, SMAD4 and KRAS mutations are much more prevalent in perihilar and extrahepatic (6–10). Moreover, the prevalence of CCA is markedly higher in Asia, including China, Japan, Korea, and Thailand (2), likely due to a different distribution of risk factors including hepatitis B and liver fluke infection compared to western countries. How these and other risk factors as well as the heterogeneous somatic mutations impact the immune tumor microenvironment is unknown at this time. In the present study, we minimized heterogeneity by investigating the genetic and immune biomarker profiles of surgically resected iCCA from a single institution in the U.S.

Experimental Procedures

Human subjects

100 formalin fixed paraffin embedded (FFPE) tissue blocks were selected from primary surgical iCCA samples from patients diagnosed between 2004 and 2016. All patients were treated and operated on at the University of Texas MD Anderson Cancer Center, Houston, TX. Informed consent in writing was obtained from each patient and the study protocol conformed to the ethical guidelines of the 1975 Declaration of Helsinki, as reflected in approval by the MD Anderson Cancer Center Institutional Review Board.

Immunohistochemistry in whole slide and image analysis

All FFPE samples were subjected to singleplex immunohistochemistry with 14 antibodies (Supplemental Table 3) using an automated IHC stainer system (Leica Bond Max, from Leica Biosystems™). Slides were then digitally scanned in the Aperio AT system (Aperio™, Leica Micrrosystems™). Image analysis was performed in 5 regions of interest (ROI) measuring 1 mm², classified as either invasive margin (IM) or central tumor (CT) areas by two experienced pathologists (WCF and LS), following standardized guidelines set forth by the International Immuno-Oncology Biomarkers Working Group (11). IM is defined as tumoral projections into the normal structure; technically, IM ROIs were centered on the border of the malignant cells and normal liver (but only the malignant portion was scored), while CT ROIs assessed malignant cells that were fully bordered by other malignant cells and/or IM areas. Previous studies have demonstrated that IM and CT can have different predictive abilities (12). Scoring was performed by pathologists blinded to the clinical data, using image analysis software (Aperio™). IHC expression was quantitated in their proper subcellular location (i.e. membrane, cytoplasm or nucleus). PD-L1, B7-H3, and B7-H4 were evaluated in all cells and were scored by an H-score, calculated as the percentage of positive cells (0–100%) multiplied by its intensity (0–3). All other singleplex markers were quantitated as positive cells per mm².

Tissue Microarray

TMA blocks utilized 2 tumor areas per sample, reviewed and selected by two experienced pathologists (WCF and FC). 1 millimeter diameter cores from each area were embedded together per sample. A total of 2 paraffin blocks were generated, with cores from 45 and 38 samples, respectively.

Multiplex immunofluorescence (mIF) and multispectral analysis

TMA slides were sectioned at 5 microns and used to perform mIF staining with 2 panels of 6 multiplexed antibodies each (Supplemental Table 4, including pan-cytokeratin which is used to identify tumor cells), then scanned with the Vectra imaging system at high resolution (20x). Histologic assessment of each analysis area was performed to quantify the percent tumor purity (Supplemental Table 1). InForm software was used to individualize and unmix the multiplex stains. The spectral library information was used to associate each fluorochrome component with a mIF marker. All immune cell populations from each panel were quantified as positive cells per mm² using the cell segmentation and phenotype cell tool from the InForm 2.2 image analysis software (PerkinElmer) under pathologist supervision (FC and WCF).

FGFR2 FISH

FISH was carried out using the breakapart FGFR2 translocation Vysis probe (Cat # 08N42–020) Briefly, CitroSolv was used to remove paraffin, then samples were immersed in pretreatment buffer (Abbott Molecular Cat # 32–801210) at 80°C for 10min, followed by denaturation with protease buffer at 37°C for 30 minutes. Samples were then washed into 3 series of ethanol, then placed in a hybridization buffer water bath at 73°C for 5 minutes. The FGFR2 probe was added, followed by a cycle of 80°C for 5 minutes and 37°C for 16 hours in a hybridization machine. Next day the samples were washed with 2x SSC/0.3% NP-40 at 73°C for 30 minutes. Slides were air-dried for 30 minutes and counterstained with 1 drop of DAPI. Scoring was performed by examining 200 nuclei with at least one 3’-centromeric (labeled with Spectrum Red) or 5’-telomeric (labeled with Spectrum Green) signal. Samples with 20% or more cells with overlapping red and green signals were considered as harboring FGFR2 fusions.

Genomic analysis of iCCA using next generation sequencing (NGS)

For T200 sequencing, DNA was extracted from unstained sections from 56 FFPE blocks (Qiagen) and quantified using a double-stranded DNA PicoGreen assay, with a QC cutoff of >50 ng/ml, OD 260/280 between 1.8–2.0, and high molecular weight. The T200 NGS targeted sequencing panel was performed by the MD Anderson Sequencing Core. MAF files were generated through a dedicated pipeline and all mutations were manually assessed to ensure accuracy before being assembled into a matrix. For Foundation Medicine sequencing, whole blocks were sent to Foundation Medicine for sequencing using their proprietary panel. Only tumor tissue was sequenced. As matching normal tissue was not sequenced, we cannot formally rule out that some pathogenic mutations may have been germline or in normal somatic cells. One example is rare germline BAP1 mutations which have a known cancer predisposition, although we note that CCA only occurs in ~1% of BAP1 germline-mutant probands (13). To minimize identifying non-tumor mutations, we have focused on genes with a known functional role in iCCA etiology.

K-means clustering and correlation analyses

K-means clustering was performed in R using the Bioconductor’s ComplexHeatmap package (14). The distance calculation was Euclidian and the clustering method was ward. The correlation matrix of singleplex markers was generated in R using the corrplot.mixed function from the Corrplot package.

Gene expression analysis

Bulk tumor tissue was scraped from four 5uM unstained slides per sample, macrodissecting to avoid normal tissue, thus enriching for tumor. RNA was extracted following the manufacturer’s protocol (Roche High Pure miRNA isolation kit, Cat. #05 080 576 0001). Analytes were run using the Nanostring nCounter Immune Exhaustion Panel (NanoString, USA) according to the manufacturer’s recommendations. Briefly, 500 ng of total RNA was loaded per sample, and probes for each gene in the panel bind to and detect their cognate mRNA molecules. The number of mRNA molecules was counted by the NanoString nCounter, and the raw count data was normalized with the nSolver Analysis Software version 4.0 using the included housekeeping genes.

BAP1 knockout and cell lines

The iCCA cell lines CCSW1 and SSP25 were maintained in DMEM + 10% FBS + 1% penicillin/streptomycin. 3 sgRNAs targeting human BAP1 in the pLentiCRISPRv2 backbone were purchased from GenScript, along with a non-targeting control sgRNA. The sgBAP1 sequences are: #1 CCTGATCGTAGGTGTCAAAG, #4 TCTACCCCATTGACCATGGT, and #5 ACCCACCCTGAGTCGCATGA. Both cell lines were transduced with lentivirus and selected with puromycin. After 2 weeks, cells were lysed and protein run on a western blot for BAP1 and B7H4 (Cell Signaling Technologies #13271 and #14572, respectively).

Results

Patient cohort.

A total of 96 patients diagnosed with primary iCCA between 2004 and 2016 were included (Fig. 1A): 58% were female, 100% fluke-negative, 84.7% HBV and/or HCV negative, and with a median age of 61 years (range 21–83 years) (Supplemental Tables 1 and 2). All patients underwent surgical resection of iCCA, and all pathological blocks used were from the surgery. Four patients had multiple surgeries, resulting in 2 samples each at different time points, for a total of 100 samples for analysis. 95% of samples were mass forming, 61% invaded the lymph-vascular space, and 95% were primary tumors. 22% of the patients had multiple tumor nodules at the time of the surgery.

Figure 1. The immunogenomic landscape of resected iCCA.

A) Patient selection schematic. B) Mutation, immune IHC, and selected clinical data of 100 resected IHC samples from 96 patients. IHC data represent z-scores except for B7H4 and B7H3 which were measured by H-score. Dark grey = not done or failed IHC. IM = invasive margin; CT = central tumor.

Mutation status.

95 of the samples successfully underwent next generation sequencing for 300+ cancer genes, either with the Foundation Medicine™ or the UT MD Anderson T200 platform (15) (Fig. 1B and Supplemental Table 1). Both assays contain all known major iCCA driver genes, with the exception of FGFR2 fusions not being detected by T200. For these samples, FGFR2 fusions were called using FGFR2 break-apart FISH. One sample was assayed by both NGS platforms, which were in agreement.

Counting only unique patients, known pathogenic oncogene mutations were seen in IDH1 (24%), IDH2 (6%), FGFR2 (10% fusions), KRAS (6%), NRAS (5%), BRAF (2%) and PIK3CA (1%); loss-of-function (nonsense, frameshift, or known loss-of-function missense) tumor suppressor gene mutations were seen in ARID1A (19%), BAP1 (17%), PBRM1 (7%), TP53 (7% including gain-of-function), CDKN2A (3%), and PTEN (1%). Focal copy number aberrations were seen in CDKN2A (deletion 4%) and CCND1 (amplification 1%). Overall, the prevalence of these mutations were within the ranges seen in other NGS studies.

Singleplex and multiplex immune marker immunohistochemistry.

All 100 samples underwent singleplex, whole-slide immunohistochemistry (IHC) for 14 immune cell and immune checkpoint markers (Table S1 and S3). In parallel, 68 of the samples were successfully embedded as a tissue microarray (TMA), with 2 cores per sample. These TMA slides then underwent multiplex immunofluorescence (mIF) for 2 panels of 5 immune markers each (Table S1 and S4). We found a high correlation across the 2 mIF panels for CD3+ and for CD3+/CD8+ cell counts (Supplemental Fig. 3), indicating replicability between TMA sections. Despite differences in sectioning and tissue sizes between singleplex and mIF IHC, we also observed a good concordance between the two platforms for both CD3 and CD8 counts (Supplemental Fig. 1). Examples of tumor, adjacent normal liver, and lymph node staining are shown in Supplemental Fig. 1 and 2 for all markers.

For singleplex slides, we first pathologically defined the tumor center and the invasive margin, then assessed marker density for each region separately. We found that for nearly all markers, the density was significantly higher at the tumor margins than at the tumor center (Fig. 2A). In a subset of samples with adjacent normal liver, CD3/4/8 showed a marked similarity between normal liver and IM counts, suggesting a gradient of decreasing T cell infiltration toward the tumor center (Supplemental Fig. 4). We next assessed correlations between markers and identified expected correlations, such as among CD3, CD4, CD8, and ICOS (Fig. 2B). Among immune checkpoint markers, LAG3, TIM3, PD1, VISTA, and OX40, there was limited correlation, with the strongest noted between central LAG3 and PD1. As expected, correlations between central and marginal counts within the same marker tended to be stronger than correlations to other markers. PD-L1 staining was conducted on both singleplex and mIF platforms using clone 22C3, and we noted that PD-L1 was completely absent in the vast majority of samples (n=93/97); the few positive samples showed only sporadic staining in the tumor parenchyma. We note that 22C3 is the companion clone used for patient selection for pembrolizumab.

Figure 2. Characterization of singleplex immune markers.

A) Box and whiskers plot of log2 cell counts per mm² for each singleplex immune marker, stratified by location at the tumor center or invasive margin. Paired Wilcoxon test, corrected: *** p<0.0001, ** p<0.001, * p<0.01. B) Correlation matrix for all singleplex markers. C) Single-marker associations with Kaplan-Meier survival, Mantel-Cox test, selected based on being statistically significant (p<0.05) or near-significant.

Patient subgroups stratified by immune profiles are associated with clinical outcomes.

We first conducted an analysis of patient survival associated with single markers and identified higher levels of central CD4 as significantly associated with improved survival in our cohort, and marginal TIM3 and central CD3 as borderline associated (Fig. 2C). By contrast, no single or pathway-grouped mutations reached significance in stratifying patients by overall survival.

We then sought to take advantage of the multi-marker nature of the data to identify combinations of markers that can more discretely identify patient subgroups associated with clinical outcomes. We first applied unsupervised hierarchical clustering to identify 4 patient subgroups among 74 patients with complete immune marker data (Fig. 3A). Group 2 was consistent with an immune “hot” phenotype, while Groups 3 and 4 were relatively “cold”. Notably, Group 3 stood out as having a particularly low overall survival (Fig. 3B) and differed from Group 4 primarily by higher levels of PD-1 and LAG3 (Fig. 3A).

Figure 3. Patient subgroups associated with poor survival.

A) Unsupervised hierarchical clustering of 74 patients who had complete singleplex IHC marker data (z-score). * p<0.05, ** p<0.005, *** p<0.0005, student’s t-test. B) Kaplan-Meier survival curves for the 4 cluster groups, Mantel-Cox test. C) Nanostring RNA data from 10 samples from Group 2 and 10 from Group 3, showing differentially expressed genes between the two groups, log2 values. D) Metascape protein-protein interaction map of the differentially expressed genes from C). Expression data are available in Supplemental Table 5. E) Kaplan-Meier survival curves for all 96 patients stratified by the relative expression of the listed pairwise markers, Mantel-Cox test. F) The percentage of PD1+ cells per sample that are also CD3+, for samples with >20 PD1+ cells. Data are calculated from the mIF TMA data in which co-localization can be determined from overlapping fluorescent signals. The TMA cores are primarily from CT regions.

To determine whether differences between “hot” and “cold” tumors are conserved at the RNA level, we conducted bulk Nanostring RNA analysis using the 773-gene immune exhaustion codeset for 20 samples: 10 from immune “hot” Group 2 and 10 from “cold” Group 3. 19/20 samples reflected their hot/cold IHC status, indicating strong agreement between RNA and protein (Fig. 3C and Supplemental Table 4). Assessment of differentially expressed genes identified the IHC markers CD3E, CD4, CD8A, ICOS, PDCD1 (encoding PD-1), and LAG3 as increased in group 2, as well as LCK, FYN, and STAT1, known regulators of tumor immune infiltrate signaling. Metascape analysis of protein-protein interactions among genes upregulated in Group 2 uncovered a dense chemokine signaling network including the ligands CCL13, CCL21, CCL22, CCL4, CCL5, CXCL9, CXCL11, and the receptors CCR1, CCR2, CCR5, CCR7, CXCR3, and CXCR6, all connected to a T cell module (Fig. 3D), suggesting that chemokines may be candidate biomarkers or therapeutic targets for surgically resected iCCA. Few genes were upregulated in Group 3 (Supplemental Table 4) and did not show significant enrichment for particular pathways or modules, although we note that a probe for CXCL1/2/3 was among them.

We then took a second approach to multi-marker analysis that included all 96 unique samples, incorporating those with incomplete immune marker data: we considered all pairwise immune markers. This approach identified patients with high central PD-1 or LAG3 expression and low central CD3, or to a lesser extent CD4 or ICOS, as significantly associated with low overall survival (Fig. 3E, Supplemental Fig. 5), consistent with the features of Group 3 and indicating that the two approaches converged on a similar set of markers. We note that central PD1 and LAG3 are only moderately correlated (ρ = 0.48, Supplemental Fig. 5). Interestingly, there was no significant enrichment for specific mutations in either Group 3 or low-survival-associated markers pairs. Group 3 also did not show any clear association with other clinical parameters including time to tumor relapse, multi-nodularity, presence of metastases, right/left lobe location, differentiation status, lymph-vascular invasion, steatosis, or CA19–9 levels, nor with any putative etiological factors including smoking, HBV/HCV status, alcohol consumption, hypertension, hyperlipidemia, or the etiological inflammatory factors primary sclerosing cholangitis, gallstones, cholecystitis, or fatty liver (Supplemental Table 1).

We note that “low” values of CD3+ cells are relative and that the absolute counts always outnumbered PD1+ cells within a sample, even when PD1+ was “high” (Supplementary Figure 2), To more deeply explore their relationship, we analyzed the available PD-1 co-localization markers CD3 and CD68 in the mIF TMA data. This indicated that when present, PD1 is primarily expressed on CD3+ cells (mean = 67%, 95% CI [53%, 80%], Fig. 3F), with nearly none on CD68+ cells (mean = 0.06%). The latter contrasts with findings in gastric cancer, for example, that find high levels of tumor-associated PD1-positive macrophages (16), suggesting that iCCA may be deficient in this immune cell subtype. Together, these findings suggest that patients with relatively low tumor lymphocyte infiltration and an exhausted immune phenotype at the time of surgery may be at higher risk for shorter overall survival.

BAP1 mutations are associated with higher B7H4 and B7H3 expression.

Finally, we sought associations between mutation status and immune profiles. Samples harboring BAP1 mutations were found to have significantly higher counts for B7H4 and B7H3 signals compared to BAP1 wild type samples (Fig. 4A; p=0.009 and p=0.016, respectively) and much higher than normal liver (Supplemental Fig. 4). B7H4 and B7H3 are immunosuppressive checkpoint markers. Of the 18 BAP1 mutations, 13 are frameshift or nonsense mutations resulting in a truncated protein, 1 is a focal copy number deletion, 1 is an in-frame deletion, and 3 are missense (Fig. 4B). All 4 in-frame deletion and missense mutations are in the highly conserved UCH domain responsible for carrying out Bap1’s deubiquitinase activity. L230Q and H141R have been found in renal carcinoma (17) and uveal melanoma (18) samples, respectively, two cancers with a high incidence of pathogenic BAP1 mutations. E198K and 211_214del have not previously been described. Thus, at least 16/18 of the mutations are predicted to be loss-of-function.

Figure 4. BAP1 mutants are associated with elevated B7H4 and B7H3.

A) Quantification of singleplex IHC data for B7H4 and B7H3, H-score. B) Lollipop diagram of BAP1 mutations in this study. Each mutation was seen once. C) Western blot showing the effect of CRISPR-mediated BAP1 knockout on BAP1 and B7H4 protein levels in two human iCCA cell lines, CCSW1 and SSP25. sgNT = non-targeting sgRNA.

To cross-validate, we assessed the similar-sized Japanese ICGC RNAseq dataset of 135 iCCA (6), and found a trend of >2-fold higher average VTCN1 (encoding B7H4) and CD276 (encoding B7H3) expression in BAP1 mutant versus BAP1 wild type samples, but neither reached significance (p=0.11 and p=0.054, respectively; Supplemental Fig. 6). In our 20-sample Nanostring data, the 3 BAP1 mutant samples all showed high expression of VTCN1 (Supplemental Fig. 6). To determine whether the increase in B7H4 is a direct result of BAP1 loss, we used CRISPR to knock out BAP1 in 2 human iCCA cell lines, CCSW1 and SSP25. B7H4 showed a strong increase using 3 independent BAP1 sgRNAs in both cell lines, but not in the non-targeting sgRNA control (Fig. 4C), indicating that the effect is tumor cell-intrinsic. No other association between mutations and immune or clinical data reached significance in this cohort.

Discussion

To our knowledge, this is the first study of a wide immune marker immunohistochemistry panel in CCA. Moreover, our selective assessment of surgically-resected, fluke-negative iCCA from a single institution minimizes the high degree of molecular and pathologic heterogeneity otherwise seen across different CCA anatomical and geographic subtypes. Given the low percentage of CCA responses to immune checkpoint therapies in early clinical trials, we focused specifically on TILs and immune checkpoint markers as a first step to understanding the immune tumor microenvironment in iCCA. We found patterns that trend towards immune “hot” and “cold” profiles, but as expected, not all markers are so broadly correlated. We took advantage of the nuanced differences to identify multi-marker patient subgroups, and highlight a tumor-center PD1- or LAG3-high, CD3/CD4/ICOS-low profile as being associated with poor overall survival, suggesting that anti-PD1 and anti-LAG-3 therapies may benefit specific patient subgroups in an adjuvant/neoadjuvant setting. Our assessment of the additional checkpoint markers TIM3, OX40, and VISTA afford the first immunohistochemical look at their correlation and dynamic range in iCCA. As targeted immune agonists and antagonists for these molecules are showing early promise in clinical trials for other cancers (19), further assessment of these markers in CCA will help delineate appropriate clinical inroads to testing and patient stratification.

A key though controversial element in stratifying the responsiveness to PD1/PD-L1-based checkpoint therapies is sufficient levels of PD-L1 expression, which is inadequate as a single biomarker (20). In our cohort, we noted a broad lack of PD-L1 staining (4% positive), consistent with the relatively low efficacy of ant-PD1 therapies in CCA in early clinical trials. Our low results using clone 22C3 are consistent with CCA studies conducted in the U.S. and China using clone SP142 (21, 22), while other studies in the U.S. and Japan show higher values using clones 5H1 (23) or E1L3 (24). The degree to which these discrepancies are dependent on the antibody clone versus geographic location or hepatitis infection status (25) will require larger studies directly comparing different clones. Our results are also potentially consistent with a low prevalence of microsatellite instability in CCA (~1–3%), which is associated with improved immune checkpoint therapy responses and an inflamed tumor microenvironment across cancers (26).

Our IHC approach also allowed us to divide markers between tumor-central and -marginal localizations, providing higher resolution than bulk assays such as RNAseq, and positional information unavailable from single cell RNAseq. We found that nearly all markers tested had higher expression in the tumor margin, suggesting that localization is an important distinguishing parameter when assessing immune markers. Moreover, it was central PD1/LAG3/CD3/CD4/ICOS and not marginal that most strongly associated with overall survival, consistent with tumor immune infiltration being critical for understanding iCCA prognoses.

We also identified an association of clinical BAP1 mutations with high B7H4/3 expression, and validated it as a causal relationship between CRISPR-mediated BAP1 loss and B7H4 upregulation in iCCA cell line models. Bap1 possesses a deubiquitinase domain that primarily targets H2AK119ub1, causing derepression of a wide variety of genes in a tissue-dependent manner, but also stabilizes the repression of specific PRC1/2-regulated genes (27, 28). BAP1 mutations have previously been associated with increased immune suppression in uveal melanoma (29), which has a high prevalence of BAP1 mutations. Interestingly, however, BAP1 mutations are not associated with B7H4/3 expression changes in public datasets of uveal melanoma or clear cell renal carcinoma (Supplemental Fig. 6), suggesting that our observation may be unique to iCCA. Notably, BAP1 mutations have also been associated with an adverse prognosis in several cancer types (30) and specific inhibitors targeting B7H4/3 are in clinical trials at this time (31); further exploration of these inhibitors may be warranted in BAP1-altered iCCA.

Interestingly, no significant immune marker association was found with clinically actionable IDH1 mutations or FGFR2 fusions, suggesting that deeper analyses such as longitudinal biopsies from patients undergoing IDH or FGFR inhibitor treatment will be needed to further nominate optimal immunotherapy co-targets. Moreover, the lack of enrichment for specific mutations in the low-survival groups suggests that immune biomarkers may be superior to mutation analyses in identifying the most at-risk patients after surgery. However, we note here that our analysis is limited to cases of surgical resection, and may or may not extend to non-operable cases. We hope that our study stimulates similar assessments, particularly in metastatic cohorts.

As CCA treatments have lagged behind many other cancers in testing and adopting targeted and immune therapies, a continually diversifying array of multi-platforms analyses will help stabilize the field as it is finally poised to adapt newer-generation therapies such as IDH and FGFR inhibitors and immune checkpoint therapies. Our overall approach provides a first look at detailed immune profiles associated with survival and mutation status, with potential relevance to developing neoadjuvant and adjuvant approaches given the surgical resection nature of the cohort. Moreover, our data creates a foundation upon which to build a more detailed understanding of the immune tumor microenvironment in larger CCA cohorts. With sufficient statistical robustness, such large combined immunohistochemical and omics studies will create the basis on which to accurately personalize effective therapies for this devastating disease.

List of Abbreviations:

BAP1

BRCA1 associated protein-1

B7H3

B7 Homolog 3

B7H4

B7 Homolog 4

CCA

cholangiocarcinoma

CD3

cluster of differentiation 3

CD4

cluster of differentiation 4

FGFR

fibroblast growth factor receptor

FFPE

formalix-fixed, paraffin-embedded

iCCA

intrahepatic cholangiocarcinoma

HBV

hepatitis B virus

HCV

hepatitis C virus

ICOS

inducible T Cell costimulator

IDH

isocitrate dehydrogenase

IHC

immunohistochemistry

LAG3

lymphocyte-activation gene 3

mIF

multiplex immunofluorescence

PD-1

programmed cell death protein 1

PD-L1

programmed-death ligand 1

TMA

tissue microarray