High-resolution digital dissociation of brain tumors with deep multimodal autoencoder

Single-cell technologies enable high-resolution, multi-dimensional analysis of molecular profiles in cancer biology but face challenges related to low coverage and cell annotation. The inherent hetero-geneity and complexity of brain tumors may hinder large-scale single cell multi-omic profiling. An efficient alternative is digital dissociation, which involves quantifying cell abundance and purifying bulk samples at high resolution. However, most existing tools for resolving bulk transcriptomes using scRNA-seq as a reference are not easily transferred to other omics (e.g., chromatin accessibility, DNA methylation, protein) due to ambiguous cell markers. Here, we introduce MODE, a novel multimodal autoencoder neural network designed to jointly recover personalized multi-omic profiles and estimate multimodal cellular compositions. MODE is the first algorithm trained on large-scale pseudo-bulk multi-omics derived from an external scRNA-seq reference and an individualized non-RNA reference panel constructed from target tumors. The accuracy of MODE was evaluated through an extensive simulation study, which generated realistic multi-omic data from distinct tissue types. MODE outperformed existing deconvolution pipelines with superior generalizability. Additionally, the high-resolution multi-omic profiles purified by MODE showed strong fidelity and enhanced power to detect differentially expressed genes. We applied MODE to bulk methylome-transcriptome data from two independent brain tumor cohorts, revealing modality-specific cellular landscapes in pediatric medul-loblastoma (MB) and adult glioblastoma (GBM). In MB tumors, MODE accurately predicted the composition of embryonal lineage cells and their marker genes expression. In GBM, the deconvoluted cellular composition revealed an increased myeloid cell abundance associated with poorer event-free survival. Overall, MODE high-resolution multimodal dissociation unravels the origins, evolution, and prognosis of pediatric and adult brain tumors, offering a powerful tool for multi-omic analysis at cell state resolution without large-scale single cell sequencing. MODE is a pipeline freely available at https://github.com/jsuncompubio/MODE .