Investigation of a global mouse methylome atlas reveals subtype-specific copy number alterations in pediatric cancer models

Abstract

Copy number alterations (CNAs) are hallmarks of cancer, yet investigation of their oncogenic role has been hindered by technical limitations and missing model systems. Here we generated a genome-wide DNA methylation and CNA atlas of 106 genetic mouse models across 31 pediatric tumor types, including 18 new models for pediatric glioma. We demonstrated their epigenetic resemblance to human disease counterparts and identified entity-specific patterns of immune infiltration. We discovered that mouse tumors harbor highly recurrent CNA signatures that occur distinctly based on the tumor subgroup and driving oncogene and showed that these CNAs share syntenic regions with the matching human tumor types, thereby revealing a conserved but previously underappreciated role in subgroup-specific tumorigenesis that can be analyzed using the presented models. Our study provides insights into globally available mouse models for pediatric solid cancers and enables access to functional CNA interrogation, with the potential to unlock new translational targets in pediatric cancers.

Fig. 1. Autochthonous cohort of mouse models for pediatric glioma.

a, Mouse models of pediatric glioma were generated by in utero electroporation of CRISPR and Tol2-transposon constructs. b, Eighteen different mouse models for various glioma types were generated, with tumor development in 18 of 22 electroporated constructs. Models were categorized into the indicated four subgroups (BRAF-driven, EGFR/FGFR-driven, PDGFRA-driven or IHG-like) and are color-coded respectively. Hit combinations that did not induce tumor formation are shown in gray. Letters indicate the figure panels (c–t) showing the respective model. c–t, Different hit combinations induced brain tumor formation in mice. The model latency and penetrance depended on the hits involved. Mouse tumors showed histological features of high-grade glioma including necrosis and pleomorphic nuclei, as well as mitoses indicating rapid proliferation. Scale bar, 100 μm. Panels a and b created using BioRender.com. FL, full length.

Fig. 2. DNA-methylation-based clustering meaningfully stratifies mouse tumors.

a, Cohort overview. b, UMAP of all 453 primary samples using the 10,000 most significantly differentially methylated CpG sites. c, Heatmap of significantly differentially methylated CpG sites in promoter regions between all entities. Canonical cancer genes are annotated. d, UMAP of methylation profiles of mouse models for embryonal tumors. e, Comparison of three different mouse models driven by loss of tumor suppressor Smarcb1. The model for MRTs was generated by intramuscular (i.m.) electroporation of adult mice and additional KO of Trp53. The models for AT/RTs were generated by embryonal tamoxifen-driven loss of Smarcb1. These models develop two types of AT/RT depending on tumor location. Methylation profiles of mouse rhabdoid tumor models showed more similarity between MRT and AT/RT MYC compared to AT/RT SHH tumors. The 10,000 most significantly differentially methylated CpG sites are displayed. The corresponding human entities showed the same pattern in their methylation profile, with MRTs and AT/RT MYCs clustering in proximity. f, UMAP of integrated gene expression data of publicly available GEO datasets (for details, see Methods) of embryonal mouse tumor models. g, Comparison of transcriptome and methylome data. The average of each tumor entity was compared to data for healthy cerebellum, and the z-score (for transcriptome) or log fold change (for methylome) was calculated. After applying a cutoff (z-score difference >0.5 or <−0.5 and log fold change >0.25 or <−0.25), XY-plots were constructed, and the number of points per area was calculated. h, XY-plot of gene expression and DNA methylation of ATRT MYC versus ATRT SHH. Genes that were also significantly differentially expressed in human ATRTs are shown in orange (Extended Data Fig. 2i). NB, neuroblastoma; MPNST, malignant peripheral nerve sheath tumor; CPC, choroid plexus carcinoma; FB, fibroblastic tumors; IFS, infantile fibrosarcoma; UPS, undifferentiated pleomorphic sarcoma; SM, skeletal muscle tumor; aRMS, alveolar rhabdomyosarcoma; eRMS, embryonal rhabdomyosarcoma; UD, uncertain differentiation tumors; APSP, alveolar soft part sarcoma; Sy, synovial sarcoma; ETMR, embryonal tumor with multilayered rosettes; aGBM, adult glioblastoma; PXA, pleomorphic xanthoastrocytoma; EPN, ependymoma; WT, wild type; BS, brainstem; CB, cerebellum; CC, cerebral cortex; TH, thalamus; FC, fold change; CNS-NB, central nervous system neuroblastoma; i.u., in utero; G3, grade 3.

Fig. 3. Epigenetic resemblance of mouse models to their human counterparts.

a, Schematic of the workflow to compare mouse and human DNA methylation profiles. The human reference set was derived from Capper et al. and included 1,554 samples (a maximum of 20 samples per entity). After preprocessing, the 15,000 most significantly differentially methylated CpG sites were extracted. Of these, the 675 ortholog CpG sites on the mouse array were identified and used for analyses. b, UMAP of human tumor samples indicated robust clustering using only the identified 675 orthologue CpG sites. c, Each mouse sample was separately compared to a randomly chosen 90% of the human reference set, and this process was iterated ten times. The closest human match was calculated. d, A random forest (RF) classifier was trained on the same 675 CpG sites, and the entity of all mouse samples was predicted. e, Comparison of matching between UMAP and RF for all models that matched the expected corresponding human entity. f, Matching to the expected human tumor entity is displayed. Samples displayed a match in both RF and UMAP, only UMAP or only RF, or not at all. g, The mutational load of modeled human entities was derived from Sturm et al. and compared to the percentage of matching per model. Only entities with more than three samples in the human reference set and with more than one mouse model were included. The trend line represents a regression curve of all entities with at least one match (full circles).

Fig. 4. Deconvolution reveals model-specific immune infiltration.

a, Mouse immune cells were isolated from the spleens of C57BL/6 mice, followed by FACS. Methylation profiles of B cells, NK cells, CD8 T cells, CD4 T cells, T_reg cells, monocytes, neutrophils and eosinophils, as well as average methylation profiles of healthy mouse brain and representative mouse brain tumors, were used to build a reference matrix of 1,000 CpG sites. b, Heatmap showing the reference matrix. c, Validation of reference-matrix-based deconvolution using FACS-sorted immune cells of CD1 mice. d, Deconvolution of mouse brain tumors and control brain samples revealed distinct populations of immune cells. e, Average percentages of T cells and B cells as well as monocytes after deconvolution. n = 24 in all graphs, representing 24 different tumors. Data are presented as mean values, with error bars representing the standard deviation of a two-sided t-test. f, IHC validation of the deconvolution results showing low B cell and T cell levels, as well as various monocyte infiltrations. All stainings were performed under standard conditions in at least three biological replicates. One representative picture is provided per tumor type. Scale bar, 100 μm. g, Comparison of immune cell compositions of mouse and human tumor samples. Deconvolution of human tumors was performed by applying the reference matrix described by Grabovska et al. to the human reference set derived from Capper et al.. h, Comparison of immune cell ratios in human and mouse tumor samples. i–k, Correlation of mouse and human immune cell fractions in MB SHH (i), DMG K27M (j) and RTK/IHG (k). Sig. diff., significantly different; chr., chromosome; NS, not significant.

Fig. 5. Highly recurrent mouse CNA profiles recapitulate human counterparts.

a, Overview of identified CNAs in all mouse models. b–j, CNA profiles of three individual mouse SHH MB (b), MYC-driven Gr3 MB (c), MYCN-driven Gr3 MB (d), MYCN-driven pHGG (e), G34R/PDGFRA-driven pHGG (f), NF1-driven pHGG (g), ALK-driven pHGG (h), NTRK2-driven pHGG (i) or ROS1-driven pHGG (j). Recurrent changes are color-coded. Middle sections of panels display the frequency of the color-coded CNAs in all analyzed mice of that model, as well as the respective syntenic regions on the human genome. Lower sections show human CNA summary profiles of the indicated human tumor entity. Sections that did not overlap with the indicated mouse CNAs are shown in black. Syntenic overlap of each color-coded mouse CNA with human summary plots is depicted in red and was statistically analyzed as indicated by the P values next to the respective CNA. The P values were derived by one-sided t-tests without adjustment for multiple testing. Enrichment scores were calculated by comparing the overlap of the indicated CNA with the related human tumor entity to the overlap of that CNA with a pediatric reference set.

Fig. 6. Continuous specific CNA evolution in mouse models.

a, Scheme displaying propagation through allografting. b, UMAP of the 10,000 most significantly differentially methylated CpG sites of primary tumors, as well as allografts and/or cultured cells of the respective models. Most propagated samples showed a close similarity with their respective primary tumor. c–e, CNA profiles of ALK/CDKN2A-driven pHGG (c), FGFR1/CDKN2A-driven pHGG (d) and MET/TRP53-driven pHGG (e) revealed continuous acquisition of CNAs during passaging, including recurrent gain of chromosome 11. f, Three independently propagated allograft lines of the ZFTA-RELA ependymoma model ultimately all acquired a gain of chromosome 1 and loss of chromosome 4, indicating highly specific tumor evolution during in vivo passaging. Recurrent changes are color-coded. Lower section shows human CNA summary profiles of the indicated human tumor entity. The overlapping changes between mice and human are depicted in red. Syntenic overlap of each color-coded mouse CNA with human summary plots was statistically analyzed as indicated by P values in the same color. Enrichment scores were calculated by comparing the overlap of the indicated CNA with the corresponding human tumor entity to the overlap of that CNA with a pediatric reference set. Illustration in a created using BioRender.com.

Extended Data Fig. 1. Histological overview of novel glioma mouse model cohort.

a–r, Mouse models of pediatric high-grade glioma (pHGG) were generated by in utero electroporation of the indicated CRISPR and Tol2-transposon constructs. H&E overview pictures are demonstrated on the left, immunohistochemical staining of an HA-tag is displayed on the right. The staining is positive for all models with an HA-tag present in the vector. All tumors include tumor cells expressing the glial cell markers Gfap and/or Olig2. The scale bars in (a) represents 500 μm in all overview pictures and 100 μm in all high-magnification IHC pictures. s, Combined survival of mouse models and closest corresponding human entity. IHG = Infantile hemispheric glioma. FL = full length, KO = knockout, IHG = Infant High-Grade Glioma, RTK = Receptor Tyrosine Kinase.

Extended Data Fig. 2. Composition of the sample cohort.

a–c, Samples from three different material types, >5 different methods for mouse model generation and 20 different laboratories were included. d–f, Quality control assessment leads to the exclusion of 19 samples. Exclusion criteria are mean intensity, number of missing values (NAs) and probe success. For all criteria, the 95th percentile was calculated, and all samples in the lowest 5% in all three filter criteria were excluded for all further analyses. g–i, UMAPs of the primary samples show no material-, method- or lab-specific methylation pattern. j, UMAP-analysis of control samples reveals a tissue, age and strain dependent methylation profile. k, UMAP-analysis of glial tumor samples shows oncogene-dependent methylation profiles. l, Volcano plot of differential gene expression between human ATRT SHH and human ATRT MYC. In red are genes which have a log2-fold change >1 and are expressed mainly in ATRT SHH, in green are the genes significantly higher expressed in ATRT MYC. Labeled are genes also found in mouse samples showing high correlation between RNA and methylome (Fig. 2h). Significantly expressed genes were determined by a two-sided t-test and values were corrected for multiple testing by the Benjamini-Hochberg procedure.

Extended Data Fig. 3. The 675 CpG sites employed for human mouse comparisons are sufficient to distinguish human CNS tumor entities.

a, Position of each human sample in a UMAP upon mutiple iterations. Each sample is depicted as a line representing the coordinates. The resulting plot indicates that samples of one entity group together consistently validating the stability of the UMAPs. b, The position of each sample in 500 UMAP iterations was used to calculate the correlation of different samples. The heatmap of pearson correlation coefficients shows stable classes of human tumors representing the different entities. The correlation within classes is 0.897. This analysis shows that the selected 675 CpGs are sufficient for distinguishing human CNS tumors.

Extended Data Fig. 4. Comparing mouse methylation patterns with human tumors reveals high resemblance of mouse models to their human counterparts.

a–f, Example UMAPs of DNA methylation patterns of the human reference set and one mouse sample per UMAP. Samples are color-coded based on their entity, the black arrow points at the mouse sample. The inset is a zoom into the region of the mouse sample. The abbreviations are adopted from Capper et al., 2018.

Extended Data Fig. 5. Mouse models match corresponding human entities with statistical significance.

a, Heatmap of the average distance of each mouse model to the different human entities. b, Oncoplot showing the predicted entity as well as the expected human entity to each mouse model. The matches in UMAP and random forest are statistically validated using a binomial test. c, Dotplot showing no correlation between matching success and distinctiveness of human entities in UMAP depictions. The color legend of Familiy, Class and Type in a) also applies to b) and c).

Extended Data Fig. 6. Technical validation of the deconvolution approach and comparison to human tumor samples.

a, Gating strategies used for sorting B cells (CD45 + CD19+ TCRb-), CD4 + T cells (CD45 + CD19+ TCRb+ CD4 + CD8-), CD4+ Treg cells (CD45 + CD19- TCRb+ CD4 + CD8- CD127- CD25+), CD8 + T cells (CD45 + CD19- TCRb+ CD4- CD8+), NK cells (CD45 + CD19- TCRb- CD11bdim NKp46 + NK1.1+), Granulocytes (CD45 + CD19- TCRb- CD11b + Ly6G + Ly6C-), Monocytes (CD45 + CD19- TCRb- CD11b + Ly6G- Ly6C+) and Eosinophils (CD45 + CD19- TCRb- CD11b+ SiglecF+ SSChigh) from C57BL6 and CD1 mice. b, Deconvolution of the sorted mouse immune cell populations used for the newly developed reference matrix using the PRmeth algorithm39 does not show a better performance than the employed EpiDish-algorithm. c, To test the newly generated reference matrix for deconvolution, a published reference matrix for mouse immune cells40 was used for deconvolution of the samples of the here described reference matrix. The published reference does not include brain tissue and is not able to outperform the here described new reference matrix. d, Comparison of immune cell populations derived from DNA methylation deconvolution of murine and human tumors.

Extended Data Fig. 7. Immunhistochemistry of immune cell markers in mouse CNS tumors.

a, T cells (CD3), monocytes (Iba-1) as well as B cells (CD20) were immunohistochemically stained in represenative embryonal tumors (AT/RT SHH and MB SHH) with available FFPE-tissue. b, The same immune cell markers were used to identify immune cell populations in all new glioma models presented in this study. Most samples show low to no T cell and B cell infiltration and the amount of Iba-1 positive monocytes varies among models. Scale bar represents 100 μm. All stainings in a and b were performed under standard conditions and all performed stainings are depicted.

Extended Data Fig. 8. Additional CNA profiles.

a, b, Cumulative CNA plots of human entities which are used for comparison to mouse tumors. c–e, CNA profiles of three individual mice per mouse model indicated. Recurrent changes are color coded. Middle sections of panels display the frequency of the color-coded CNAs in all analyzed mice of that model as well as the respective syntenic regions on the human genome. Lower sections show the average human CNA profile of the indicated human individual mice of the indicated models. Overlapping changes between mice and human in the same direction are depicted in red. f, T-SNE analysis of all human RTK-driven HGGs shown in Extended Data Fig. 8b, e. chr. – chromosome, CNA - copy number aberration, CPC - choroid plexus carcinoma.

Extended Data Fig. 9. Characteristics of allografts and in vitro cultivated tumors.

a–c, Hierarchical clustering of global DNA methylation profiles (10,000 most significantly differentially methylated CpGs) of indicated samples. In the PPP1CB-ALKand FGFR K656E-driven model, the primary tumors that were used for in vitro and in vivo propagation (P) cluster closer to allografts than to the intermediate in vitro culture. d–f, The immune cell infiltration is comparably low in cultivated as well as in allografted mouse tumors, allowing comparison of methylation profiles.