Abstract 5010: High-resolution multi-omic dissociation of brain tumors with deep multimodal autoencoder

Single cell technology measures multi-dimensional molecules at cellular resolution in cancer biology with challenges of low coverage and cell annotation. The inherent heterogeneity and complexity of central nervous system (CNS) tumors may hinder single-cell multi-omic profiling in large population-based studies. An alternative and efficient approach is digital dissociation, i.e., cell abundance quantification and high-resolution prediction of cell-type-specific (CTS) bulk samples. Many tools were developed to resolve mixture bulk transcriptomes by integrating scRNA-seq reference, not immediately transferable to other omics (e.g., chromatin accessibility, DNA methylation) because of ambiguous cell marker identification. We develop MODE - a novel multimodal autoencoder neural network with parallel decoders - to purify CTS multi-omic profiles and estimate the multimodal cellular compositions. MODE is trained on the large-scale pseudo bulk multiomes constructed by probabilistic data generation with an external scRNA-seq reference and individualized non-RNA reference panel recovered from the target tumors. The CTS differentially expressed genes (ctsDEGs) in purified transcriptomes were detected by Wilcoxon rank-sum test with false discovery rate corrected, while survival analysis was performed by log-rank test. MODE was benchmarked in an extensive simulation study generating realistic multiome from normal and tumor brain tissues, showing outperformance and superior replicability over the competing deconvolution pipelines. The purified CTS multiomes exhibited high resolution and fidelity with enhanced robustness in detecting ctsDEGs. We applied MODE to the bulk methylome-transcriptome in two independent brain tumor studies, recovering modality-specific cellular landscape for pediatric medulloblastoma (MB) and adult glioblastoma multiforme (GBM). The embryonal lineage cells for MB tumor groups were accurately recovered, i.e., Group 3 (G3) MB tumors having the origin of rhombic lip (RL) stem cells (P <10-5) and Group 4 (G4) arising from unipolar brush cells (P <10-5). Notably, hundreds of marker genes for RL or granule neuron progenitor cells identified in human fetal cerebellum were differentially expressed among (n=253) Sonic Hedgehog, G3, G4 MB patients in the corresponding lineage-specific tumor transcriptomes predicted by MODE. The results in TCGA GBM cohort showed higher abundance of myeloid cells in either methylation (P=0.002) or mRNA (P=0.019) being associated with poorer event free survival, while increased tumor purity delayed (P=0.009) the relapse or progression of GBM tumor. Overall, the high-resolution multimodal dissociation by MODE elucidates the origins, evolution, and prognosis in pediatric and adult CNS tumors, without the need for large-scale single cell multimodal profiling. MODE is freely available at https://github.com/jsuncompubio/MODE. Citation Format: Jiao Sun, Yue Pan, Tong Lin, Kyle Smith, Arzu Onar-Thomas, Giles W. Robinson, Wei Zhang, Paul A. Northcott, Qian Li. High-resolution multi-omic dissociation of brain tumors with deep multimodal autoencoder [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 5010.