Distance-Preserving Representations for Genomic Spatial Reconstruction.

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Spatial context is vital for single-cell gene expression analysis but often unavailable.
• Technical limitations restrict access to spatial information in many datasets.
• This limits the scope and depth of downstream analyses.

Purpose of the Study:

• To develop a generalizable framework, dpVAE, for reconstructing spatial coordinates from gene expression data.
• To enable the recovery and utilization of spatial context in datasets lacking this information.
• To enhance the applicability of single-cell genomics to diverse research areas.

Main Methods:

• Proposed a representation learning and transfer learning framework, dpVAE.
• Integrated a distance-preserving regularizer into the model's loss function.
• Utilized constrained optimization for spatial context reconstruction during inference.

Main Results:

• Demonstrated dpVAE's effectiveness across 27 diverse public datasets.
• Showcased robustness in training, out-of-sample evaluation, and transfer learning.
• Successfully reconstructed spatial coordinates from gene expression data.

Conclusions:

• dpVAE provides a powerful tool for inferring spatial context in single-cell genomics.
• The framework overcomes limitations of inaccessible spatial data.
• Facilitates broader applications of genomics studies previously hindered by missing spatial information.