MODE: high-resolution digital dissociation with deep multimodal autoencoder.

Area of Science:

• Single cell biology
• Computational biology
• Multi-omics analysis

Background:

• Tissue complexity challenges cell mapping and tumor microenvironment analysis.
• Existing deconvolution methods primarily focus on transcriptomics and struggle with other omics data.
• Need for advanced computational tools for digital dissociation of bulk tissues.

Purpose of the Study:

• To develop a multimodal autoencoder pipeline (MODE) for digital dissociation of bulk tissues.
• To enable joint prediction of multi-omic profiles and cellular compositions.
• To overcome limitations of transcriptomic-focused deconvolution methods.

Main Methods:

• MODE utilizes a multimodal autoencoder to link multi-dimensional features.
• It constructs pseudo-bulk data using internal non-transcriptomic references and external single-cell RNA sequencing (scRNA-seq) data.
• The pipeline jointly predicts personalized multi-omic profiles and cellular compositions.

Main Results:

• MODE demonstrated superior generalizability and fidelity compared to nine other deconvolution pipelines in rigorous simulations.
• The method was validated using real multi-omic data from diverse tissue types.
• Successful prediction of cellular compositions and multi-omic profiles was achieved.

Conclusions:

• MODE offers a robust and generalizable approach for digital dissociation of bulk tissues.
• The pipeline effectively integrates non-transcriptomic and transcriptomic data for comprehensive analysis.
• MODE advances multi-omic single cell biology by enabling accurate cell mapping and microenvironment decomposition.