Cycle-consistent deep generative modeling unifies cellular states across unpaired spatial and single-cell modalities

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Current spatial and single-cell technologies provide incomplete views of cellular states.
• Integrating transcriptomic, proteomic, and spatial data is challenging due to unpaired measurements and modality biases.

Purpose of the Study:

• To present MultiTME, a novel multimodal framework for integrating heterogeneous spatial and single-cell data.
• To enable cross-modal translation and harmonization of biological data without paired observations.

Main Methods:

• Utilized a spatially-regularized, cycle-consistent deep generative model.
• Enforced consistency of bidirectional mappings to learn a shared latent representation.
• Applied the framework to benchmark datasets and multimodal colorectal cancer data.

Main Results:

• MultiTME outperforms existing methods in data integration and cross-modal cell typing.
• Improved spatial transcriptomic panel completion and generated high-resolution spatially resolved maps.
• Revealed a proliferative-invasive tumor axis in colorectal cancer and corrected platform biases between Xenium and CosMx.

Conclusions:

• MultiTME effectively integrates multimodal spatial and single-cell data, overcoming key technical limitations.
• The framework facilitates cross-dataset harmonization and enables comprehensive pan-cancer spatial studies.
• MultiTME advances the understanding of spatial biology and disease mechanisms.