Bayesian Hyperspherical Graph Mixture-of-Experts Deciphers Cell-Cell Interaction in Spatial Transcriptomics.

Area of Science:

• Computational Biology
• Systems Biology
• Genomics

Background:

• Spatial transcriptomics (ST) advances tissue biology by providing gene expression with spatial context.
• Existing computational methods for cell-cell interactions (CCIs) struggle with fixed proximity, limited databases, and over-smoothing.
• Limitations hinder discovery of complex, directional, and long-range CCIs crucial for understanding tissue organization and disease.

Purpose of the Study:

• Introduce B-HGME (Bayesian Hyperspherical Graph Mixture of Experts), a scalable, unsupervised framework for ST data analysis.
• Jointly delineate spatial domains and infer CCI networks with uncertainty quantification.
• Overcome limitations of existing methods for discovering heterogeneous, directional, and long-range CCIs.

Main Methods:

• Integrate spatial and gene-regulatory graphs into a dual-scale structure.
• Encode cell representations on a unit hypersphere using coupled message passing.
• Employ a Bayesian mixture-of-experts with a Dirichlet-regularized gating network for edge decoding.

Main Results:

• Achieve state-of-the-art spatial clustering accuracy across diverse ST datasets.
• Uncover biologically coherent and diverse CCIs, including novel interactions beyond curated ligand-receptor pairs.
• Demonstrate accurate localization of canonical markers, confirming biochemical fidelity at single-gene resolution.

Conclusions:

• B-HGME provides a powerful tool for spatial systems biology and hypothesis generation.
• The framework enables discovery of novel ligand-receptor circuits, offering mechanistic insights into development and disease.
• B-HGME facilitates interpretable and confident CCI inference with principled uncertainty estimates.