Multi-omics analysis of adamantinomatous craniopharyngiomas reveals distinct molecular subgroups with prognostic and treatment response significance

Abstract

Background:

Adamantinomatous craniopharyngioma (ACP) is the commonest pediatric sellar tumor. No effective drug is available and interpatient heterogeneity is prominent. This study aimed to identify distinct molecular subgroups of ACP based on the multi-omics profiles, imaging findings, and histological features, in order to predict the response to anti-inflammatory treatment and immunotherapies.

Methods:

Totally 142 Chinese cases diagnosed with craniopharyngiomas were profiled, including 119 ACPs and 23 papillary craniopharyngiomas. Whole-exome sequencing (151 tumors, including recurrent ones), RNA sequencing (84 tumors), and DNA methylome profiling (95 tumors) were performed. Consensus clustering and non-negative matrix factorization were used for subgrouping, and Cox regression were utilized for prognostic evaluation, respectively.

Results:

Three distinct molecular subgroups were identified: WNT, ImA, and ImB. The WNT subgroup showed higher Wnt/β-catenin pathway activity, with a greater number of epithelial cells and more predominantly solid tumors. The ImA and ImB subgroups had activated inflammatory and interferon response pathways, with enhanced immune cell infiltration and more predominantly cystic tumors. Mitogen-activated protein kinases (MEK/MAPK) signaling was activated only in ImA samples, while IL-6 and epithelial–mesenchymal transition biomarkers were highly expressed in the ImB group, mostly consisting of children. The degree of astrogliosis was significantly elevated in the ImA group, with severe finger-like protrusions at the invasive front of the tumor. The molecular subgrouping was an independent prognostic factor, with the WNT group having longer event-free survival than ImB (Cox, P = 0.04). ImA/ImB cases were more likely to respond to immune checkpoint blockade (ICB) therapy than the WNT group (P <0.01). In the preliminary screening of subtyping markers, CD38 was significantly downregulated in WNT compared with ImA and ImB (P = 0.01).

Conclusions:

ACP comprises three molecular subtypes with distinct imaging and histological features. The prognosis of the WNT type is better than that of the ImB group, which is more likely to benefit from the ICB treatment.

Introduction

Craniopharyngioma comprises two histopathological subtypes, including adamantinomatous craniopharyngioma (ACP) and papillary craniopharyngioma (PCP). ^[1] However, because they differ in etiology and age distribution, the recent fifth edition of the WHO Classification of Tumors of the Central Nervous System (WHO CNS5) updated them as two tumor types instead of subtypes of craniopharyngiomas (CPs). ^[2] PCPs, occurring almost exclusively in adults with occasional diagnoses in children, are driven by somatic B-Raf proto-oncogene, serine/threonine kinase gene (BRAF) p.V600E mutation that activates the mitogen-activated protein kinases (MEK/MAPK) signaling pathway. ^[1] Previous reports have demonstrated ACPs are driven by somatic mutations in CTNNB1 (encoding β-catenin) exon 3 that block β-catenin phosphorylation, which induces its nuclear accumulation and Wnt/β-catenin pathway activation. ^[3,4] However, many studies have reported no CTNNB1 mutations in a portion of ACPs, with mutation frequencies ranging from 16% to 100%. ^[5–15] No further recurrent somatic mutations have been identified.

Currently, efforts for targeting CTNNB1 mutations have not been successful so far. ^[16] Various anti-inflammatory tools have been examined to treat ACP. ^[17] Clinical studies have evaluated the response of ACPs to intralesional and systemic administration of interferon-α (IFNα). ^[17–20] Although a response to intracystic IFNα treatment was reported by a few case studies, a recent phase II trial of systemic treatment failed. ^[18] In the retrospective analysis of clinical trials assessing IFNα, imaging response was only observed in predominantly cystic tumors in both intracystic and systemic treatments. ^[19] Transcriptional analysis indicated inflammasome activation in both solid and cystic components, and targeting IL-6 with the monoclonal antibody tocilizumab resulted in a significant response in cystic ACPs. ^[21–24] These findings suggest there are distinct ACP subgroups who respond differently to IFNα or anti-inflammatory treatment.

Since CTNNB1 mutation was identified as the driver event for ACPs, wildtype tumors have been reported in multiple studies. ^[5–13] The effort to divide ACPs into the wild and mutant types has obtained no support from clinical data. ^[25,26] Additionally, transcriptome and DNA methylation profiling revealed no differences between mutated and wildtype ACPs, though a large intertumoral heterogeneity was recognizable in a limited sample size (n = 18) RNA-seq study. ^[24]

The failure of molecular typing with the mutation status of the driver gene has prompted investigators to explore other typing methods. Yuan et al^[27] utilized public datasets (GSE60815 and GSE94349) and divided them into the immune resistance and immunogenic subtypes to evaluate the response to immune checkpoint blockade (ICB) therapy. ^[28] However, this classification was solely based on a bioinformatic analysis with no validation from independent datasets or clinical data, and may not reasonably explain response differences to IFN-α and IL-6 treatments. Furthermore, the latter classification does not clarify the relationship between different subtypes and dysregulated pathways. ^[18–20] Although heterogeneity is also significant among ACPs' transcriptome profiles in Prince et al's ^[29] study, these authors failed to identify any differences between children and adults. However, a study of pediatric brain tumors identified two subtypes, "epithelial" and "immune hot," in ACPs by proteomics and phosphorylated proteomics data. ^[30] An interesting finding is that the MEK/MAPK pathway is activated in one subtype, similar to other tumors driven by the BRAF V600E mutation, and this subtype may benefit from MEK/MAPK inhibitors. ^[24,31] These findings indicate there are different molecular subtypes of ACPs with different responses to anti-inflammatory, immunomodulatory, and/or other therapies.

In this study, whole-exome sequencing, RNA-sequencing, and DNA methylation array were utilized to profile a large cohort of Chinese craniopharyngioma cases, to obtain a full-spectrum characterization of tumors, providing deeper insights into the molecular pathogenesis of tumors for further development of precision medicine.

Methods

Ethical approval

Ethical approval was granted by the Ethics Committee of Sanbo Brain Hospital, Capital Medical University (No. SBNK-YJ-2020-014-01) and the Biological and Medical Research Ethics Committee of Fujian Medical University (No. FMU-2019-43). Written informed consent was obtained from the patients and/or their guardians.

Patients and sample collection

Histological diagnosis was confirmed for all tumors by a neuropathologist, and representative fresh-frozen (FF) and formalin-fixed paraffin-embedded (FFPE) blocks were assessed. Tumor and/or blood specimens were collected from 143 patients following surgical resection of CPs (119 ACPs and 24 PCPs) between 2008 and 2020. Whole-exome sequencing (WES) was carried out on 119 tumors (99 ACPs and 20 PCPs) with patient-matched blood specimens [Supplementary Table 1, http://links.lww.com/CM9/B630], as well as 31 tumors (23 ACPs and 8 PCPs) without matched normal DNA samples. The quality of the WES profiles of one PCP patient with matched blood specimen was low and no further analysis was performed for this patient. Therefore, the case was not included in Table 1. The WES profiles of the primary tumors and the recurrent tumors were available for four ACP cases without matched blood specimens, one PCP case with matched blood specimen, and three PCPs without matched blood specimen. Transcriptome data were obtained by RNA sequencing for 84 FF tumor specimens (74 ACPs and 10 PCPs). Library preparation and next-generation sequencing were carried out with standard protocols on a NovSeq 6000 platform (Illumina Inc., San Diego, CA, USA) using a paired-end sequencing mode with a 150-bp read length at Berry Genomics (Beijing, China). Illumina Infinium MethylationEPIC array data were generated for the DNA methylome (DNAm) profiles of 96 tumors (89 ACPs and 7 PCPs) using standard protocols at Shanghai Biotechnology Company (Shanghai, China).

Table 1.

Summary of demographic, clinical and experimental data for the craniopharyngioma patients in this study.

Factors	Groups
	ACP	PCP
WES	119	23
With matched blood	99	18
Tumor only	20	5
RNAseq	74	10
DNAm	89	7
Sex
Female	43	11
Male	76	12
Age group
Pediatric (≤14 years)	61	4
Adult (>14 years)	58	19
Status
Primary	44	8
Recurrent	75	15

The number of cases is given in each cell. For a small number of patients, both the primary and recurrent tumors were available and they were subject to WES profiling twice. Data are shown as number of cases. ACP: Adamantinomatous craniopharyngioma; Ages: Ages at diagnosis; DNAm: Whole-genome DNA methylome profiling with EPIC array; PCP: Papillary craniopharyngioma; RNAseq: Whole-transcriptome RNA sequencing; WES: Whole exome sequencing.

Preprocessing and data analysis

Raw reads obtained by WES were cleaned, and somatic mutations were called with the Mutect2 algorithm using the GATK4 (v. 4.1.8.1, https://github.com/broadinstitute/gatk) best practice pipeline proposed by Broad Institute (https://gatk.broadinstitute.org; "GATK pipeline" is shown in Figure 1). MutScan (v. 1.14.0, https://github.com/OpenGene/MutScan) analysis of the clean reads and manual inspection of BAM files with Integrative Genomics Viewer (IGV v.2.8.0, Broad Institute, Cambridge, USA; https://igv.org) were also used to check the mutation statues of common somatic mutation genes. MutSigCV (v. 1.41, https://www.genepattern.org/modules/docs/MutSigCV) was utilized to determine the significance of mutated genes. Purity, clonality, and ploidy were determined using ABSOULTE (v. 1.0.6, http://www.broadinstitute.org/cancer/cga/ABSOLUTE) and PureCN (v. 1.18.0, https://github.com/lima1/PureCN). Sanger sequencing was performed to validate the obtained somatic mutations in six patients. The whole-genome sequencing (WGS) profiles of 35 pediatric ACPs from the Pediatric Brain Tumor Atlas (PBTA) project were downloaded from Kids First Resource Portal (https://portal.kidsfirstdrc.org). The proteomic profiles of the PBTA cohort were downloaded from the Clinical Proteomic Tumor Analysis Consortium (CPTAC) project (https://gdc.cancer.gov/about-gdc/contributed-genomic-data-cancer-research/clinical-proteomic-tumor-analysis-consortium-cptac). Mutational landscapes of the current cohort and the PBTA cohort are presented in Supplementary Figures 1 and 2, http://links.lww.com/CM9/B629, respectively. Data used from these databases are all publicly available.

Figure 1.

ACPs and PCPs are driven by CTNNB1 and BRAF somatic mutations, respectively. (A) Mutation status and VAFs of CTNNB1 and BRAF in all WES-profiled tumors (122 ACPs and 26 PCPs, from 119 ACP patients and 22 PCP patients, respectively) and copy number variation results. The "mutation status" indicates the final result through all means and the "GATK pipeline" indicates the original Mutect2 result. The samples enclosed by the red square were "wildtype" in the GATK pipeline but found to harbor CTNNB1 mutations by RNAseq; the samples inside the black square were "wildtype" in both the initial WES and RNAseq data and subject to a second round of WES with newly cut FFPE specimens. Somatic copy number variants include both the results of WES and DNAm profiles (EPIC). (B) Distribution of mutant amino acids in CTNNB1 variants. The color indicates the property of the side chain residue. (C) VAFs of CTNNB1 point mutations in RNAseq correlated with WES data (Spearman R = 0.40, P <0.01). ACPs: Adamantinomatous craniopharyngiomas; FFPE: Formalin-fixed paraffin-embedded; PCP: Papillary craniopharyngioma; RNAseq: RNA-sequencing; VAF: Variant allele frequencies; WES: Whole exome sequencing; DNAm: DNA methylome.

Raw reads obtained by RNAseq were cleaned, aligned with Hisat2 (v.2.1.0, https://daehwankimlab.github.io/hisat2/), and quantitated with Subread (v. 1.6.4, https://subread.sourceforge.net/) with the GENCODE (v. 32, https://www.gencodegenes.org) annotation. Salmon (v. 0.8.2, https://combine-lab.github.io/salmon/) was also employed to generate gene and transcript expression profiles. STAR-Fusion (v. 1.8.1, https://github.com/STAR-Fusion/STAR-Fusion/wiki) and GeneFuse (v. 0.6.1, https://github.com/OpenGene/GeneFuse) were applied to call gene fusion events for clean reads.

DNAm analyses were carried out with the ChAMP (v. 2.18.3, https://bioconductor.org/packages/release/bioc/html/ChAMP.html) pipeline, and copy number variation analysis utilized the "champ.CNA" function, and the results are shown in Supplementary Figure 3, http://links.lww.com/CM9/B629.

Dimensionality-reduction was carried out by t-distributed stochastic neighbor embedding (t-SNE) analysis (Rtsne package, v. 0.15, https://github.com/jkrijthe/Rtsne). Consensus clustering was performed using the Monte Carlo reference-based consensus clustering algorithm (M3C package, v. 1.10.0, https://bioconductor.org/packages/release/bioc/html/M3C.html). Consensus non-negative matrix factorization (NMF) analysis utilized the NMF package (v. 0.23.0, nrun = 100, http://renozao.github.io/NMF/). Results for RNAseq profiles are shown in Supplementary Figures 4 and 5, http://links.lww.com/CM9/B629, and results for DNAm profiles are shown in Supplementary Figures 6 and 7, http://links.lww.com/CM9/B629. CIBERSORT (https://cibersort.stanford.edu/) and xCell (https://xcell.ucsf.edu) were employed to deconvolute RNAseq profiles, and EpiDISH (v. 2.4.0, https://bioconductor.org/packages/release/bioc/html/EpiDISH.html) was employed to deconvolute DNAm profiles (see Supplementary Figures 8 and 9, http://links.lww.com/CM9/B629). Differential expression analyses employed edgeR (v. 3.30.3), while enrichment analyses were carried out with the gprofiler webservice, Gen set enrichment analysis (GSEA), and single-sample GSEA algorithms. [Supplementary Figures 10 and 11, http://links.lww.com/CM9/B629]. Pathway signatures were downloaded from the Molecular Signatures Database (MsigDB, https://www.gsea-msigdb.org/gsea/msigdb/). Response to ICB therapy was predicted with the Tumor Immune Dysfunction and Exclusion (TIDE) algorithm (http://tide.dfci.harvard.edu) using the RNAseq profiles as the input.

Hematoxylin and eosin (H&E) and immunohistochemistry (IHC) staining

H&E and IHC staining were performed routinely. Tumor tissue structures, the degree of infiltration of inflammatory cells in the stroma, and the existence of multinuclear macrophages were observed under a microscope. The presence of β-catenin (ZM-0442, ZSGB-BIO, China) and BRAF V600E (G05510, Roche, USA) proteins, as well as various inflammatory markers, including cluster of differentiation 3 (CD3), 20 (CD20) and 38 (CD38) were studied by IHC. Immunostaining slides were evaluated by observation under a high-resolution microscope.

Survival analysis

The follow-up of 116 ACP patients was completed on March 2023 and 18 cases were lost to follow-up before recurrence was observed. The mean follow-up time was 5.31 ± 4.29 (standard deviation) years. Tumor recurrence was first assessed by computed tomography (CT) and magnetic resonance imaging (MRI) during the postoperative review, and confirmed by a further resection as well as pathological analysis. Event-free survival (EFS) was the time elapsed from the initial surgical resection to the first relapse, progression, or surgical resection of the recurrent tumors. Kaplan–Meier curves were utilized to estimate the event-free probability, and the differences between the two survival curves were compared with the log-rank test in the survival (v. 3.2.7, Therneau T, Rochester, USA; https://github.com/therneau/survival) package. A multivariate Cox proportional hazards model was employed to assess the statistical significance of different factors.

The R language (4.0.0, R Core Team, Vienna, Austria; https://www.r-project.org) was used for statistical analysis, with two-tailed P-values reported. A P value of ≤0.05 was considered to be significant.

Results

Apparent "wildtype" ACPs also harbor CTNNB1 mutations

Patient-matched tumor and blood whole-exome DNA sequences were collected from 117 individuals diagnosed with CPs, including 99 ACPs (40 primary and 59 relapsed) and 19 PCPs (9 primary and 10 relapsed in 18 patients, 1 patient with both primary and recurrent specimens; Table 1 and Supplementary Table 1, http://links.lww.com/CM9/B630). At the time of tumor resection, patients' ages were 2–69 years (median = 17 years), and the initial ages at diagnosis were 0–66 years (median = 14 years). Somatic single-nucleotide variants (SNVs) and short insertions and deletions (indels) were called using Mutect2 [Figure 1 and Supplementary Table 2, http://links.lww.com/CM9/B631]. The pipeline identified oncogenic CTNNB1 exon 3 mutations in 59% (57/97) of ACPs and BRAF p.V600E mutations in 82% (14/17) of PCPs.

Among 30 "WES-wildtype" ACPs with RNAseq profiles, 21 (70%, marked with the red square in Figure 1) carried somatic mutations in CTNNB1, which, as a housekeeping gene, has very high expression levels. MutScan analysis and manual IGV inspection confirmed mutations in 10 samples, though variant allele frequencies (VAFs) were very low in these samples. However, in eight ACPs neither WES nor RNAseq could detect CTNNB1 mutations and these samples were subjected to another round of deep-coverage WES with newly cut FFPE specimens (marked with the black square in Figure 1). Majority of the samples (5/8) were CTNNB1 mutated. In total, 89% (109/122 in terms of tumor numbers) of all ACPs were definitely CTNNB1-mutated. The distribution of mutated amino acids of CTNNB1 is depicted in Figure 1. The VAFs inferred from WES were strongly correlated with those inferred from RNAseq, suggesting the tumor purity was a common latent factor determining the VAFs [Figure 1]. This comprehensive analysis revealed all ACPs harbor CTNNB1 mutations, and the presence of wildtype samples is due to the low proportion of tumor cells in specimens and/or the low sensitivity of sequencing techniques. No other driver events were detected in ACPs through comprehensive analysis in either our cohort or the PBTA cohort (Supplementary Figure 1 and 2, http://links.lww.com/CM9/B629). Further analysis detected 25 out of 26 PCPs (22 cases, excluding 1 case carrying both CTNNB1 and BRAF mutations, and both primary and recurrent tumors specimens were available for 4 cases) carrying the BRAF p.V600E mutation, and immunostaining using BRAF VE-1 antibody revealed the remaining wildtype sample was positive as well.

More mutations (54%, 57/106) occurred at codon S33 or adjacent positions, and reduced mutation frequencies were found at S37 (25%, 26/106) and T41 (17%, 18/106). Tumors with mutations at codon positions 32–34 showed a trend with improved EFS than those at 35–37 or 41 (log-rank P = 0.35). The property types of mutant residues also showed position dependence. All the mutant residues detected were non-polar at T41 and S45. However, there were some pseudo-phosphorylation mutant residues (cystine and tyrosine, with a thiol or hydroxyl group on the side chain) at S33 and S37. There was a significant number of polar mutant residues at D32 and G34, and these mutations hindered the phosphorylation of S33 via steric effects. In addition to point mutations, three samples showed in-frame deletions covering either phosphorylation site T41 and/or S45 or the three amino acids upstream of S41, suggesting similar oncogenic mechanisms. The mutant codon position demonstrated a weak association with the age group, with a higher proportion of mutations at 31–34 or 35–37 in the pediatric group (≤14 years) compared with adult cases [Supplementary Table 3, http://links.lww.com/CM9/B632].

Among ACPs with CTNNB1 point mutations, tumors with pseudo-phosphorylation residues (cystine and tyrosine) or polar mutant residues had better EFS than those with non-polar mutant residues (log-rank P = 0.01). Cytosine and phenylalanine were the two major mutant residues at S33, and cytosine mutants also showed a trend of improved EFS compared with phenylalanine mutants (log-rank P <0.01). Transversion mutants had a slightly better EFS than transition mutants. Mutations at D32 and G34 had approximately the same EFS as S33 mutants (log-rank P = 0.35).

ACP comprises three distinct molecular subgroups

While in a portion of ACPs, no CTNNB1 mutations were detected by WES or WGS, their RNAseq profiles (71 ACPs and 9 PCPs, excluding the low-quality profiles of 3 ACPs and 1 PCP) did not cluster based on the called CTNNB1 mutation status, and few differentially expressed genes were detected between the mutated and "wildtype" ACPs. However, consensus clustering analysis with M3C and consensus NMF identified three subgroups in ACPs, which were named WNT, ImA, and ImB, while PCPs were homogeneous [Figure 2]. Consensus clustering and NMF analysis of DNAm profiles (88 ACPs and 7 PCPs, excluding the low-quality profile of 1 ACP) confirmed this finding [Figure 2]. Compared with clustering results obtained by RNAseq profiles, DNAm-based clustering showed a strong concordance for 55 tumors with both data available (Rand index =0.779, Supplementary Table 4, http://links.lww.com/CM9/B633). Analysis of RNAseq profiles in the PRJEB18056 (https://www.ncbi.nlm.nih.gov/bioproject/?term=PRJEB18056, n = 18) and PBTA (n = 36) cohorts also confirmed the existence of distinct ACP subgroups [Supplementary Figure 12, http://links.lww.com/CM9/B629], with similar molecular characteristics to subgroups in the current cohort. Furthermore, 3 distinct subgroups were also identified in 16 proteomic profiles of CPTAC [Supplementary Figure 13, http://links.lww.com/CM9/B629].

Figure 2.

RNAseq and DNAm profiles revealed that ACPs comprise three distinct molecular subgroups. (A) Dimensionality reduction plots with t-SNE of RNAseq profiles; (B) The consenus matrix heatmap in the M3C analysis of RNAseq; (C) Dimensionality reduction plot with t-SNE of DNAm profles. (D) VAFs of CTNNB1 mutations are significantly elevated in the WNT subgroup compared with the immune subgroups (ImA and ImB), which also consist of wildtype tumors (Kruskal–Wallis χ² = 36.8, P <0.01, n = 24, 15 and 32 for the WNT, ImA, and ImB subgroups, respectively). (E) xCell immune scores estimated from RNAseq are strongly correlated with the proportions of immune cells inferred from DNAm analysis (Spearman R = 0.57, P <0.01). (F) Proteomic profiles showed that CD38 is expressed in both the ImA and ImB subgroups, but not in the WNT subgroup (Kruskal–Wallis χ² = 9.19, P = 0.01); CD14 expression was elevated in the ImB subgroup (Kruskal–Wallis χ² = 7.81, P = 0.02, n = 6, 6, and 4 for the WNT, ImA, and ImB subgroups, respectively). (G) Heatmap of signature genes in various subgroups, xCell scores for epithelial, immune, and stromal cells, and single-sample GSEA scores for representative tumor hallmark pathways. Samples in each subgroup are sorted in a decreasing order of VAF of CTNNB1 mutations. The colors in A, C, D, E, F follow the same convention and they are indicated in the legend of G. ACPs: Adamantinomatous craniopharyngiomas; CD14: cluster of differentiation 14; CD38: cluster of differentiation 38; t-SNE: t-distributed stochastic neighbor embedding; VAFs: Variant allele frequencies; DNAm: DNA methylome.

ACP subgroups show different molecular and histopathological characteristics

A comparison of demographic features revealed a notable difference in age distribution between the subgroups (Fisher's exact test, P <0.01, excluding PCPs; Supplementary Table 3, http://links.lww.com/CM9/B632). The ImB subgroup mostly consisted of children, whereas the WNT subgroup had a larger proportion of adult cases. Differential expression and pathway enrichment analysis were performed between subgroups and the results are presented in Supplementary Tables 5–7, http://links.lww.com/CM9/B648.

All ACPs in the WNT subgroup were CTNNB1 mutated, whereas the apparent wildtype tumors were distributed in the ImA and ImB subgroups. VAFs were significantly higher in WNT compared with the other two subgroups, indicating a higher proportion of tumor cells in WNT specimens [Figure 2 and Supplementary Tables 8 and 9, http://links.lww.com/CM9/B649], whereas there was still a large proportion of ImB samples with high VAF values. From RNAseq profiling, the estimated immune cell scores were significantly higher in the ImA and ImB subgroups compared with the WNT subgroup [Supplementary Figure 9, http://links.lww.com/CM9/B629]. The xCell immune scores estimated from RNAseq profiles were strongly correlated with the proportions of immune cells inferred from DNAm profiles (Figure 2, Spearman's R = 0.57, P <0.01).

Immunohistochemical (IHC) staining revealed nuclear accumulation of β-catenin in most ACPs, whereas the rate of β-catenin-nuclear positive cells was the smallest in the ImB subgroup [Figure 3 and Supplementary Table 10, http://links.lww.com/CM9/B634]. H&E staining showed higher degree of immune infiltration in the ImA and ImB subgroups, which was confirmed by IHC staining [Figure 3]. Reactive gliosis was only observed in the ImA subgroup [Figure 3]. The presence of whorl-like cell clusters is a histological hallmark of ACP, and the density of these cell clusters also followed a trend similar to that of β-catenin-nuclear positive cells [Figure 3].

Figure 3.

ACP subgroups show distinct histological and imaging characteristics. (A) Typical IHC staining of β-catenin in the three ACP subgroups (arrow heads, whorl-like cell cluster; scale bar = 50 μm). (B) Typical H&E images demonstrate immune infiltration and gliosis characteristics in the ACP subgroups (left for immune infiltration, scale bar = 20 μm; right for gliosis, scale bar = 50 μm). The degree of immune infiltration was significantly more severe in the ImA and ImB subgroups than in the WNT subgroup (Kruskal–Wallis χ² = 7.90, P = 0.02, Supplementary Table 11, http://links.lww.com/CM9/B635). Tumors with a significant amount of multinucleated giant cells (indicated by arrows in ImB) were mainly detected in the ImB subgroup. Gliosis (indicated by arrows in the WNT and ImA subgroups) was more pronounced in the ImA subgroup. (C) Typical IHC staining images for CD3 and CD38 detection, respectively (scale bar = 50 μm). Additional IHC images for other markers are depicted in Supplementary Figure 14, http://links.lww.com/CM9/B629. (D) Both the proportions of cells with β-catenin nuclear accumulation (ImB vs. WNT, Welch t = 2.38, P = 0.02) and numbers of whorl-like clusters (ImB vs. WNT, Welch t = 3.68, P <0.01), indicated by arrows in (A) were significantly lower in ImB (n = 23, 15, and 32 for the WNT, ImA, and ImB subgroups, respectively). ImA samples had higher infiltration rates of CD3+ cells (Kruskal–Wallis χ² = 2.66, P = 0.26), and ImA and ImB had more CD38+ cells infiltrated (Kruskal–Wallis χ² = 6.21, P = 0.04, n = 5 for each subgroup). (E) Enrichment with xCell showed that macrophages were enriched in ImA and ImB samples (Kruskal–Wallis χ² = 47.06, P <0.01), while NKT cells were enriched in ImA samples only (Kruskal–Wallis χ² = 44.92, P <0.01, n = 24, 15, 32, and 9 for WNT, ImA, ImB, and PCP, respectively). (F) Typical CT and MRI images for predominantly solid and predominantly cystic ACPs, respectively. (G) The WNT subgroup consisted of more predominantly solid tumors, while the other two subgroups consisted of more predominantly cystic tumors (Fisher's exact test, P <0.01). (H) Receiver operating characteristic curve analysis to classify 35 ACPs based on the proportion of non-tumor area in H&E images (area under the receiver operating curve, AUC = 0.866). ACP: Adamantinomatous craniopharyngioma; CD3: cluster of differentiation 3; CD38: cluster of differentiation 38; CT: computed tomography; H&E: Hematoxylin and eosin; IHC: Immunohistochemical; Immune inf.: immune infiltration; NKT: Natural killer T cell; PCP: Papillary craniopharyngioma.

H&E staining confirmed that immune cell infiltration was significantly more severe in the ImA and ImB subgroups than in the WNT subgroup [Supplementary Table 11, http://links.lww.com/CM9/B635 and Supplementary Figure 14, http://links.lww.com/CM9/B629]. IHC also revealed more CD3+ cells in the ImA subgroup than in the WNT subgroup, and more CD38+ cells in the ImB subgroup compared with the WNT subgroup [Figure 3 and Supplementary Table 12, http://links.lww.com/CM9/B636]. Additionally, multinucleated giant cells were detected in 9/72 samples and appeared only in the ImB (6 cases) or ImA (2 cases) subgroups [Supplementary Table 11, http://links.lww.com/CM9/B635]. The proteomic profiles of CPTAC showed CD38 expression in both the ImA and ImB subgroups, but not in the WNT subgroup [Figure 2], and CD14, a macrophage marker, was expressed at the highest level in the ImB subgroup. Enrichment analysis with xCell confirmed that ImB specimens had the highest macrophage scores among the three subgroups while the natural killer T (NKT) cell score was significantly higher in the ImA subgroup compared with the ImB and WNT subgroups [Figure 3].

Differential expression analysis identified a large number of differentially expressed genes (DEGs) in each ACP subgroup [Supplementary Figure 10, http://links.lww.com/CM9/B629]. A few marker genes are shown in Figure 2. The membrane receptors EGFR and FGFR2, the sonic hedgehog marker gene GLI3 and the chromatin remodeling gene CHD3 were upregulated in WNT, and the neurosystem-specific marker genes GFAP, CLU, and CTNND2 were mainly expressed in ImA, with a high enrichment score for astrocytes. The epithelial–mesenchymal transition (EMT) marker gene COL1A2 was upregulated in the ImB subgroup. The CpG probes on these genes also showed distinct methylation patterns [Supplementary Figure 8, http://links.lww.com/CM9/B629]. Among the 16 signature genes in Figure 2, 8 were detected by proteomics, of which 7 except CLU, displayed similar differential expression patterns at the protein level.

Enrichment analysis revealed that Wnt/β-catenin, Notch, E2F targets, and other signaling pathways were significantly enriched in the WNT subgroup, the myogenesis pathway was enriched in ImA, and the EMT pathway was enriched in ImB, whereas inflammatory and IFNα/γ response pathways were significantly enriched in both the ImA and ImB subgroups [Figure 2]. Enrichment with ssGSEA also revealed macrophage/microglial and core T-cell gene sets were significantly enriched in the ImB and ImA subgroups, and microglia were mostly in the M1 (pro-inflammatory) polarization phenotype.

ACP subgroups differ in imaging features

ACPs are highly solid and/or cystic tumors that are often characterized by calcification, which is rarely detected in PCPs. Based on CT and MRI findings and surgical records, ACPs were classified into predominantly solid and predominantly cystic tumors [Supplementary Table 13, http://links.lww.com/CM9/B637]. Predominantly cystic tumors often show a circular calcification pattern, while solid-dominant tumors have a scattered calcification pattern instead. Molecular subgrouping was significantly associated with the above solid-vs.-cystic classification (Fisher's exact test, P <0.01, Figure 3). WNT specimens had a larger proportion of predominantly solid tumors, whereas predominantly cystic tumors were more frequently found in the ImA and ImB subgroups.

Dysregulated DNA methylation may lead to the development of ACP subgroups

An integrated analysis of DNAm, RNAseq, and proteomic profiles was carried out to investigate factors driving ACP tumors to develop into distinct molecular subgroups. From 51 ACPs with paired DNAm and RNAseq profiles available, 259 genes had expression levels with a significant negative correlation with the DNA methylation level of the promoter CpG sites (Figure 4, Spearman's correlation <-0.50). From 12 ACPs with paired RNAseq and proteomic profiles available, 713 genes had protein expression levels with a strong positive correlation with mRNA levels (Spearman's correlation >0.55). These two sets shared 19 genes [Figure 4], most of which had distinct mRNA expression and DNA methylation levels among the three ACP subgroups [Figure 4]. For example, probes in the promoter region of SMAD3 (cg20353227 and cg25547520) were hypomethylated, and gene expression was significantly upregulated in WNT compared with the ImA and ImB subgroups [Figure 4].

Figure 4.

Integrated correlation analysis of DNAm, RNAseq, and proteomics profiles. (A) Distribution of Spearman correlation coefficients between mRNA expression and beta values for the corresponding promoter CpG sites and between mRNA and protein expression levels. Genes with a coefficient below -0.50 in the former or above 0.55 in the latter were considered strongly correlated genes. (B) These two sets shared 19 common genes. (C) Most of the common genes showed distinct expression or methylation pattern among ACP subgroups; the mRNA expression levels of nine genes and the beta values of the corresponding probes are plotted in the heatmap. (D) Detailed RNA-seq transcription and DNA methylation data are shown for SMAD3, an important regulator of β-catenin signaling. (E) Plot of the normalized protein level of SMAD3 versus its transcription level measured by RNAseq. Red lines are the best linear fits, and the gray shade shows 95% confidence interval. ACP: Adamantinomatous craniopharyngioma; DNAm: DNA methylome; RNAseq: RNA-sequencing; SMAD3: SMAD family member 3.

The WNT subgroup has a better EFS than the ImB subgroup

Overall survival in patients with CPs was relatively good. In our cohort, three patients died during the perioperative period, six died from other complications, and one died from unrelated causes [Supplementary Table 14, http://links.lww.com/CM9/B638]. Therefore, we focused on EFS in ACP cases with follow-up data (n = 100), defined as the time elapsed from the surgical resection of primary tumors to the first relapse, which was detected by CT and/or MRI. Ten patients had radiation therapy, including eight whose treatment occurred between the resection of primary tumors and the first relapse; this factor was not considered in the Cox proportional-hazards model due to a limited sample size.

Molecular subgrouping was the second most significant prognostic factor for EFS after surgical resection, and the ImB subgroup had a worse prognosis than the WNT subgroup (Hazard Ratio [HR] = 2.11, 95% Confidence Interval [CI]: 1.04–4.30, P = 0.04, Figure 5). Incomplete resection resulted in higher risk of recurrence than complete resection (HR = 4.69, 95% CI: 2.16–10.16, P <0.01, Figure 5, Supplementary Figure 15, http://links.lww.com/CM9/B629 and Supplementary Table 15, http://links.lww.com/CM9/B639), and CTNNB1 mutation status was not a prognostic factor.

Figure 5.

Clinical relevance of ACP subgrouping. (A,B) Comparison of EFS curves among three ACP subgroups and the degree of resection. The WNT subgroup had a better prognosis compared with the ImB subgroup (log-rank P = 0.10). Patients with complete resection had a significantly longer EFS (log-rank P <0.01). Additional survival curves are shown in Supplementary Figure 15, http://links.lww.com/CM9/B629 (C) Multivariate Cox regression found that ImB had a significantly higher recurrence hazard ratio than WNT as an independent prognostic factor other than the degree of resection. (D) Frequencies of predicted responders and non-responders to ICB therapy in ACP subgroups and PCP. (E) Enrichment analysis by ssGSEA showed that WNT samples had significantly lower scores for INFα response than other subgroups (Kruskal–Wallis χ² = 35.78, P <0.01, n = 24, 15, 32, and 9 for WNT, ImA, ImB, and PCP, respectively). (F) SOX10 expression was higher in the ImA group compared with the other subgroups (Kruskal–Wallis χ² = 39.73, P <0.01), same sample sizes as in (E). ACP: Adamantinomatous craniopharyngioma; CR: Complete resection; EFS: Event-free survival; ICB: Immune checkpoint blockade; IR: Incomplete resection; PCP: Papillary craniopharyngioma; IFNα: interferon alpha; SOX10: sex determining region Y-box 10.

The WNT subgroup is less likely to benefit from immune therapies

The TIDE algorithm was utilized to predict response to ICB therapy in CPs [Figure 5]. All PCPs except one were predicted to be responders to ICB, corroborating the observation that programmed death-ligand 1 (PD-L1, also known as CD274), a primary target of ICB, was highly expressed only in PCPs. Among ACPs, all WNT samples were classified as non-responders, whereas 10/15 ImA and 8/32 ImB samples were classified as responders. The low probability of response in the WNT subgroup was due to high T cell exclusion score, which is consistent with low immune infiltration caused by Wnt/β-catenin pathway hyperactivation. In the ImA and ImB subgroups, non-response was often due to high T cell dysfunction score. Enrichment analysis with ssGSEA showed that the ImA and ImB subgroups had higher IFNα response scores compared with the WNT subgroup [Figure 5], consistent with clinical trial data demonstrating that predominantly cystic tumors are more likely to benefit from either intracystic or systemic IFNα treatment. ^[18,19] SOX10 expression was significantly higher in the ImA subgroup compared with the WNT and ImB subgroups [Figure 5]. This gene has been shown to promote T cell-mediated tumor killing.

Discussion

In this study, we characterized the genomic landscape and heterogeneity of craniopharyngiomas in a large cohort based on multi-omics techniques. We identified three subgroups in ACPs with distinct molecular and imaging features though CTNNB1 mutations were a common driving event, and we confirmed that all PCPs, both adult and pediatric cases, were driven by BRAF p.V600E mutation and had homogeneous transcriptional and DNA methylation profiles.

In the current cohort, the initial CTNNB1 mutation rate was 59%, as detected from the WES calling pipeline, similar to the rate (62%) detected in the PBTA cohort [Supplementary Figure 2, http://links.lww.com/CM9/B629] and 69% reported in another cohort of 16 ACPs examined by WGS. ^[32] However, combining RNAseq and WES of re-sampled specimens, at least 89% ACPs were found to harbor CTNNB1 mutations. Additionally, there were no other driving mutations in wildtype ACPs, while IHC and H&E images revealed similar whorl-like clusters and nuclear accumulation of β-catenin in these samples. This suggests that failure to detect CTNNB1 mutations is due to a low proportion of tumor cells and the limited sensitivity of current sequencing techniques. A recent study applying targeted amplicon deep sequencing on tumor cells extracted by laser capture microdissection also confirmed that all ACPs harbored CTNNB1 mutations. ^[15] Unsurprisingly, efforts were fruitless to find differences between "wildtype" ACPs and mutated samples for which both were driven by the same Wnt/β-catenin pathway.

Despite the above common driver mutations of CTNNB1, previous RNAseq studies by Apps et al^[24] indicated a high degree of heterogeneity among 18 ACPs, and three clusters were recognizable. Hierarchical clustering revealed two major clusters with distinct tumor purity, VAF and abundance levels for the reactive glial tissue. Through a gene co-expression network analysis, three major components were identified and attributed to epithelial differentiation, nervous system development, and inflammation, respectively. These three gene signatures were strongly associated with the molecular characteristics of these three ACP subgroups.

Previous DNA methylation studies by Hölsken et al^[33] also reported a much larger degree of heterogeneity in ACPs compared with PCPs and the CTNNB1 mutation status and age were not associated with the clustering result of ACPs. These findings confirm that heterogeneity in ACPs is authentic, and the molecular classification proposed here is due to underlying biological differences. An integrated analysis suggested that dysregulated DNA methylation of Wnt/β-catenin pathway-related genes, including SMAD3, might contribute to the formation of subgroups. SMAD3 plays a critical role in the nuclear translocation of β-catenin preventing its degradation and also acts as a co-transcription factor of β-catenin. ^[34–36] Nuclear factor one X-type (NFIX) regulates GFAP in astrocytes and plays an important role in the induction of quiescence of neural stem cells, on the one hand, ^[37] and induces Wnt/β-catenin signaling, on the other hand. These molecules might lead ACPs to develop different subgroups despite the identical oncogenic event.

In melanoma and other malignancies, reports have reported an exclusive effect of Wnt/β-catenin pathway activation on lymphocyte infiltration, where active β-catenin signaling upregulates ATF3 that in turn suppresses CCL4 transcription and leads to insufficient dendritic cell recruitment. However, most ACPs harbored CTNNB1 mutations, ^[38–40] but immune infiltration was repressed only in the WNT subgroup. Further studies are required to determine the roles of epigenetic factors in tumorigenesis for developing effective immunotherapeutic approaches for ACPs.

Previously, clustering of proteomic profiling revealed a subset of ACPs similar to pediatric low-grade gliomas (LGGs) with the BRAF V600E mutation. ^[30] GFAP and other astrogliosis markers were expressed in this cluster only. A comparative analysis led us to conclude that this cluster corresponds to the ImA subgroup discussed above. The major link between the ImA subgroup of ACPs and BRAF V600E LGGs is the unique invasion niche present in both of them, which creates a tumor-specific cellular environment, although ACP is a non-neuroepithelial tumor in the CNS. An immune infiltration study of the PBTA cohort classified most samples in this subset into "hot" tumors and the other cluster ("C8") into "epithelial" tumors. These results are consistent with our findings, and anti-MAPK therapy was shown to be useful for this subset of ACPs. ^[24]

Clinical trials of intracystic or systemic treatment with IFNα resulted in mixed responses, and a retrospective analysis found that only predominantly cystic ACPs could benefit from immune modulation therapy. ^[18,19] Tocilizumab, an IL-6 receptor inhibitor, was successfully applied to reduce huge cystic ACPs safely and effectively. ^[21] CT and MRI demonstrated that the ImA and ImB subgroups were enriched with predominantly cystic tumors. Many inflammatory molecules have been detected in ACP cyst fluids, including α-defensins, β-thymosins, IL-6, IL-8, C-X-C motif chemokine ligand 1 (CXCL1), and interleukin 10 (IL-10). ^[17,23,41] Molecular profiling studies also reported that cluster cells can generate a pro-tumorigenic microenvironment by expressing inflammatory cytokines and chemokines. ^[39] Intracystic interferon therapy decreased α-defensin levels, suggesting innate immune responses may play a critical role in cyst generation, and immune modulation may exert therapeutic effects. ^[41] Our classification results suggested that ACPs responding to IFNα or anti-inflammatory treatment should belong to the immune subgroup. In addition, this study found that ICB therapy is suitable only for the ImA and ImB subgroups.

While the degree of resection was the most significant factor, Cox analysis revealed the WNT subgroup had a longer relapse-free survival compared with the ImB subgroup; in addition, it was an independent prognostic factor of relapse, which reflects the fact that it is easier to achieve complete resection in the WNT subgroup. GFAP and other reactive astrogliosis markers were expressed in the ImA subgroup only, suggesting the presence of severe finger-like protrusions at the invasive front of the tumor, which renders complete resection particularly challenging. The EMT marker gene COL1A2 was expressed in the ImB group. ^[42]

Notably, CD38+ cells were widespread in the ImA and ImB subgroups. CD38 was found to play an important immunosuppressive role in solid tumors. ^[43] In tumor immunotherapy, CD38-targeted treatment is expected to relieve its inhibitory effect on immune cells and may be employed as a novel therapeutic target in ACPs.

Although this study revealed that there are three subgroups of ACPs, the differences in their tumorigenesis remain to be elucidated. Additionally, simple and effective biomarkers are needed to distinguish these subgroups in clinical settings. The response of ImA/ImB subgroups to ICB treatment was based on bioinformatic prediction, and clinical trials are needed to validate these results.

In conclusion, this study unequivocally revealed three distinct subgroups of ACPs, providing a guidance for the development of precision treatment in ACPs. The WNT subgroup is suitable for surgical resection, while immune subgroups are suitable for immunotherapy. The ImA subgroup may benefit from MEK/MAPK pathway inhibitor treatment, and anti-inflammatory treatment may be more suitable for the ImB subgroup. CD38 can be employed not only as a marker for subtyping but also as a novel therapeutic target.

Acknowledgments

We thank the patients in our study and their families.

Funding

This study was supported by the Fujian Medical University (No. XRCZX2017001 to XW), the Natural Science Foundation of Fujian Province (No. 2019J01294 to XW), the Sanbo Brain Hospital Management Group (No. SBJT-KY-2020-002 to ZL), and the Capital Health Research and Development Special Fund (No. 2022-2-8013 to ZL).

Conflicts of interest

None.

Supplementary Material

cm9-137-859-s001.pdf ^{(5.8MB, pdf)}

cm9-137-859-s002.xlsx ^{(21.8KB, xlsx)}

cm9-137-859-s003.xlsx ^{(883.4KB, xlsx)}

cm9-137-859-s004.xlsx ^{(10.2KB, xlsx)}

cm9-137-859-s005.xlsx ^{(9.1KB, xlsx)}

cm9-137-859-s006.xlsx ^{(15.2KB, xlsx)}

cm9-137-859-s007.xlsx ^{(15.5KB, xlsx)}

cm9-137-859-s008.xlsx ^{(13.7KB, xlsx)}

cm9-137-859-s009.xlsx ^{(12.8KB, xlsx)}

cm9-137-859-s010.xlsx ^{(25.5KB, xlsx)}

cm9-137-859-s011.xlsx ^{(10.4KB, xlsx)}

cm9-137-859-s012.xlsx ^{(3.4MB, xlsx)}

cm9-137-859-s013.xlsx ^{(21.7KB, xlsx)}