Signal Strength Aware Latent Spaces Reveal Molecularly Distinct Substructures within Human Kidney Tissue.

Area of Science:

• Computational Biology
• Spatial Omics
• Biomedical Data Analysis

Background:

• High-dimensional datasets in spatial omics offer vast biological insights but require meaningful latent representations for interpretation.
• Current methods often struggle to disentangle complex molecular features into interpretable factors.

Purpose of the Study:

• To develop a novel computational approach for dissecting high-dimensional spatial omics data into independent, interpretable latent axes.
• To enable the translation of latent space findings back into the original measurement context for biological interpretation.

Main Methods:

• Utilized a beta-variational autoencoder combined with kernel density estimation.
• Developed a comparative-latent-traversal algorithm for mapping latent features to original measurements.
• Applied the method to imaging mass spectrometry data from human kidney tissue.

Main Results:

• The approach successfully disentangled latent space structure, separating signal strength from relative signal content.
• Identified novel subdivisions within kidney proximal tubules, confirmed to be biologically significant.
• Discovered previously unknown lipid species associated with these subdivisions.

Conclusions:

• The proposed method provides exceptional chemical insight and interpretable latent representations for spatial omics data.
• Demonstrated potential as a powerful tool for hypothesis generation and biological discovery in complex tissues.
• Successfully uncovered new biological heterogeneity within the human kidney.