MLSpatial: A machine-learning method to reconstruct the spatial distribution of cells from scRNA-seq by extracting...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) provides cellular gene expression data but loses spatial information during dissociation.
• Spatial distribution of cells is critical for understanding biological processes, especially in cancer immunotherapy.

Purpose of the Study:

• To develop a novel computational method, MLSpatial, for predicting cell spatial locations using only scRNA-seq data.
• To learn the relationship between gene expression patterns and cell spatial locations.

Main Methods:

• Developed MLSpatial, a computational method leveraging gene expression patterns to infer spatial cell positions.
• Utilized drosophila embryo (FISH and scRNA-seq) and MERFISH (osteosarcoma) datasets for model training and validation.
• Compared MLSpatial's performance against existing methods using correlation coefficients for cell-to-cell distances.

Main Results:

• MLSpatial achieved a high correlation (0.99) in predicting cell-to-cell distances in drosophila embryo FISH data.
• Demonstrated moderate success (0.514 correlation) in capturing spatial relationships in osteosarcoma MERFISH data.
• Successfully reconstructed spatial gene expression patterns in drosophila embryo using scRNA-seq data alone.

Conclusions:

• MLSpatial accurately restores relative cell positions from scRNA-seq data, with performance varying by cell type.
• The method shows promise for reconstructing spatial distributions when scRNA-seq and spatial data share similar biological backgrounds.