Abstract BACKGROUND The analysis and comparison of global DNA methylation patterns has become a gold standard in modern neuropathology. For human central nervous system (CNS) tumors and other cancers, the research community has collected thousands of methylation profiles, leading to the discovery and definition of numerous molecularly distinct tumor types and subtypes. Although reliable preclinical mouse models are essential for the understanding of tumor development and the validation of therapeutic approaches, such DNA methylation analyses have so far not been systematically performed for murine tumors owing to technical limitations. METHODS We employed the Illumina Bead Chip 285k mouse methylation array to profile ~500 samples covering 90 different mouse models representing 33 pediatric solid tumor types, including CNS tumors and sarcomas from 18 laboratories. We compared DNA methylation profiles of these models between each other and to human CNS tumors, performed immune cell deconvolution and conducted an in-depth analysis of copy number variants (CNVs) during tumor evolution. RESULTS Unsupervised clustering showed that mouse models form groups of their presumed tumor types. Using a tailor-made cross-species analysis based on syntenic CpG sites, we found that most tumors also group with their corresponding human counterparts. Immune cell deconvolution revealed distinct tumor microenvironments that showed clear similarities to human tumors. We further identified recurrent, model-specific CNVs that are syntenic to genomic aberrations in respective human tumors, suggesting an essential role of CNVs during tumor initiation or maintenance. CONCLUSION We provide a global resource of well characterized pediatric solid cancer models representing numerous different tumor types. Our comparative analyses reveal novel molecular model characteristics and validate their reliability as human disease models. Our data provide new biological insights into immune composition, tumor cell identities and tumor development by identifying model-specific CNV profiles.
Read full abstract