BioLACE: unifying spatial geometry and marker priors for cohesive cell-type clustering in spatial transcriptomics.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Spatial transcriptomics (ST) offers high-resolution tissue architecture insights.
• Current graph-based deep learning methods for ST lack interpretability and biological prior integration.
• Marker gene information is crucial for accurate cell type identification in ST.

Purpose of the Study:

• To introduce BioLACE, a scalable framework for spatial transcriptomics analysis.
• To unify spatial structure, transcriptomic variation, and marker gene profiles.
• To enhance cell type clustering accuracy and biological interpretability in ST.

Main Methods:

• Developed BioLACE, a framework utilizing a shared Variational Autoencoder (VAE) latent space.
• Implemented three joint optimization objectives: VAE reconstruction loss, graph Laplacian regularizer, and marker-informed contrastive loss.
• Applied BioLACE to MERFISH hypothalamus, mouse spinal cord, and Slide-seq mouse cerebellum datasets.

Main Results:

• BioLACE achieved superior cell type clustering accuracy across diverse ST datasets.
• Demonstrated well-defined, biologically consistent boundaries in reconstructed tissue architectures.
• Generated interpretable latent representations, facilitating biological insights.

Conclusions:

• BioLACE offers a scalable and generalizable approach for modern spatial transcriptomics analysis.
• The framework effectively integrates spatial, transcriptomic, and marker gene information.
• BioLACE enhances the interpretability and accuracy of ST data analysis.