Communication-space representation of multicellular tissue organization

Technical Summary: Communication-Space Representation of Multicellular Tissue Organization

Problem Statement
Multicellular tissue organization is a higher-order property emerging from coordinated intercellular interactions, yet current computational representations of spatial transcriptomics data primarily organize cells based on intrinsic transcriptional identity or local neighborhood composition. While these phenotype-centered approaches have advanced the identification of cellular states and spatial domains, they often fail to resolve higher-order multicellular organization that arises from coordinated communication. Specifically, cells with similar transcriptional profiles may diverge in distinct communication environments, while transcriptionally distinct populations may converge toward shared functional organization through coordinated signaling. Existing methods struggle to capture these "communication ecosystems," particularly in complex contexts such as hierarchical cortical structures, specialized pathological boundaries, and dynamic tissue remodeling.

Methodology: The CommuSpace Framework
The authors introduce CommuSpace, a computational framework that represents tissue organization in a "communication space" defined by local ligand–receptor (LR)-mediated interactions among neighboring cellular populations.

1. Local Communication Architecture Construction:

   • For each focal cell, the framework constructs a communication tensor ($T \in \mathbb{R}^{N \times K \times P}$) by quantifying sender-centric signaling interactions.
   • The tensor dimensions represent focal cells ($N$), neighboring cellular populations ($K$), and ligand–receptor signaling channels ($P$).
   • The interaction score between a focal cell $i$ and a neighboring population $k$ via LR pair $p$ is calculated based on ligand expression in the focal cell, receptor expression in neighbors, and Euclidean distance, weighted by a sender-centric formulation.

2. Non-Negative Tensor Decomposition:

   • CommuSpace employs non-negative tensor decomposition to factorize the communication tensor into recurrent communication programs.
   • This process minimizes reconstruction error under non-negativity constraints, decomposing the tensor into three factor matrices: cell-wise program activities ($U$), neighboring population loadings ($V$), and LR channel loadings ($W$).
   • The matrix $U$ serves as the communication embedding, where each cell is represented by a vector of program activities. Distances in this space reflect similarity in local signaling architecture rather than transcriptional similarity.

3. Ecosystem Identification and Interpretation:

   • Communication Ecosystems (CEs) are identified as stable regions within the communication space using consensus clustering (Louvain algorithm) across multiple resolutions.
   • These ecosystems are interpreted through their cellular composition, dominant communication programs, interacting populations, and specific LR signaling channels.
   • The framework includes scalable procedures (e.g., micro-clustering) to handle atlas-scale datasets and supports alternative signaling formulations (receiver-centric, joint) to ensure robustness.

Key Results
The framework was validated across three distinct biological systems:

• Mouse Primary Motor Cortex (MOp):

  • CommuSpace reorganized cells into cross-lineage communication ecosystems that were not simply clusters of transcriptional identity or neighborhood composition.
  • It revealed hierarchical cortical organization, where parent ecosystems (e.g., CE-6) reproducibly subdivided into daughter ecosystems (CE-6a, CE-6b) with distinct laminar positions and communication programs, preserving multiscale structure better than phenotype-centered methods.
  • Specific ecosystems (e.g., CE-0) were found to recur across anatomically separated layers, representing shared neuro-supportive communication states involving glial and vascular coordination.

• Human Breast Cancer (Xenium):

  • The framework resolved specialized pathological boundary states that were indistinguishable by cellular composition alone.
  • It identified distinct ecosystems at the DCIS boundary (CE-7) and invasive tumor margins (CE-1). Despite similar stromal-vascular compositions, these boundaries engaged distinct communication programs (e.g., ECM remodeling vs. migratory/stromal coupling).
  • CommuSpace revealed ecosystem-associated immune remodeling, showing that CD8+ T cells adopt distinct functional states (activation, adaptation, or suppression) depending on the specific communication ecosystem they inhabit, a pattern obscured in phenotype-centered embeddings.

• Mouse Kidney Injury and Repair:

  • Applied to longitudinal data, CommuSpace distinguished stable tissue scaffolds (canonical nephron compartments) from temporally ordered remodeling transitions.
  • It identified transient acute injury ecosystems (CE-4) and persistent maladaptive/failure-to-repair ecosystems (CE-5, CE-10), linking epithelial state changes to evolving multicellular communication contexts.
  • Unlike other methods that fragmented these transitions, CommuSpace preserved the ordered relationships between acute injury, fibro-inflammatory remodeling, and failed repair.

Significance and Claims
The paper claims that communication ecosystems represent a previously unrecognized higher-order organizational layer of multicellular tissues that cannot be fully explained by cell identity or conventional spatial niches.

• Relational Representation: CommuSpace shifts the paradigm from describing "what cells are" (phenotype) to "how cells organize collectively" (communication architecture).
• Decoupling Lineage from Organization: The framework demonstrates that tissue organization can be resolved through coordinated intercellular communication, allowing transcriptionally distinct populations to converge into shared organizational states and transcriptionally similar populations to separate based on communication context.
• Biological Interpretability: Unlike black-box latent embeddings, communication ecosystems are directly interpretable through their constituent signaling programs, LR pairs, and cellular interactions, providing a grounded framework for understanding tissue architecture in physiological, pathological, and regenerative contexts.
• Robustness: The approach is shown to be robust across different signaling formulations and scalable to large spatial transcriptomics datasets, offering a general foundation for comparing multicellular organization across tissues and time.

The authors conclude that communication space provides a complementary representation principle for spatial biology, linking cellular states, spatial architecture, and multicellular behavior through coordinated intercellular communication.