Interpretable and context-free deconvolution of multi-scale whole transcriptomic data with UniCell deconvolve.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Accurate deconvolution of cell type composition from transcriptomic data is crucial for understanding tissue heterogeneity and disease mechanisms.
• Existing reference-based methods often require specific, contextualized datasets, limiting their broad applicability.
• The development of unsupervised or reference-free methods is essential for advancing transcriptomic data analysis.

Purpose of the Study:

• To introduce UniCell: Deconvolve Base (UCDBase), a pre-trained, interpretable deep learning model for cell type deconvolution and identity prediction.
• To enable analysis across diverse transcriptomic data types (Spatial, bulk-RNA-Seq, scRNA-Seq) without reliance on contextualized reference data.
• To demonstrate the model's utility in identifying cellular changes in disease states and characterizing tumor microenvironments.

Main Methods:

• UCDBase was trained on 10 million pseudo-mixtures derived from a large-scale single-cell RNA sequencing (scRNA-Seq) database (>28 million cells, 840 cell types).
• The model employs deep learning for unsupervised deconvolution and cell identity prediction.
• Transfer learning capabilities were utilized to adapt the pre-trained model to specific datasets and tasks.

Main Results:

• UCDBase and its transfer-learning variants demonstrated performance comparable or superior to state-of-the-art reference-based methods in in-silico deconvolution.
• Feature attribute analysis identified gene signatures related to inflammatory-fibrotic responses in ischemic kidney injury and cancer subtypes.
• The model accurately deconvoluted tumor microenvironments, identified pathologic cell fraction changes in bulk-RNA-Seq data, and distinguished normal from cancerous cells in lung cancer scRNA-Seq data.

Conclusions:

• UCDBase provides a powerful, reference-free approach to deconvolve cell type fractions and predict cell identity across multiple transcriptomic data modalities.
• The model enhances the analysis of cellular composition and spatial context, offering insights into disease mechanisms and tissue microenvironments.
• UCDBase represents a significant advancement in transcriptomic data analysis, facilitating broader and more accurate biological interpretation.