Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SPHENIC: Topology-Aware Multi-View Clustering for Spatial Transcriptomics

Chenkai Guo, Yikai Zhu, Renxiang Guan, Jinli Ma, Siwei Wang, Ke Liang, Guangdun Peng, Dayu Hu

TL;DR

SPHENIC tackles the challenge of robust spatial clustering in spatial transcriptomics by integrating topology-aware representations via Extended Persistent Homology and enforcing spatial coherence through a dual-regularized objective. It builds a topology-enhanced, multi-view graph fusion network that jointly leverages gene-expression, spatial coordinates, and topology-derived features in an end-to-end framework. The approach achieves state-of-the-art clustering performance across 11 tissue slices (DLPFC, HBC, MBA), with substantial ARI gains and validated through ablations showing the critical roles of EPH and DualRO. While delivering strong practical impact for deciphering tissue architecture, the method incurs computational overhead due to EPH, highlighting a trade-off between topology fidelity and scalability.

Abstract

Spatial transcriptomics clustering is pivotal for identifying cell subpopulations by leveraging spatial location information. While recent graph-based methods modeling cell-cell interactions have improved clustering accuracy, they remain limited in two key aspects: (i) reliance on local aggregation in static graphs often fails to capture robust global topological structures (e.g., loops and voids) and is vulnerable to noisy edges; and (ii) dimensionality reduction techniques frequently neglect spatial coherence, causing physically adjacent spots to be erroneously separated in the latent space. To overcome these challenges, we propose SPHENIC, a Spatial Persistent Homology-Enhanced Neighborhood Integrative Clustering method. Specifically, it explicitly incorporates topology-invariant features into the clustering network to ensure robust representation learning against noise. Furthermore, we design a dual-regularized optimization module that imposes spatial constraints alongside distributional optimization, ensuring that the embedding space preserves the physical proximity of cells. Extensive experiments on 11 benchmark datasets demonstrate that SPHENIC outperforms state-of-the-art methods by 4.19%-9.14%, validating its superiority in characterizing complex tissue architectures.

SPHENIC: Topology-Aware Multi-View Clustering for Spatial Transcriptomics

TL;DR

SPHENIC tackles the challenge of robust spatial clustering in spatial transcriptomics by integrating topology-aware representations via Extended Persistent Homology and enforcing spatial coherence through a dual-regularized objective. It builds a topology-enhanced, multi-view graph fusion network that jointly leverages gene-expression, spatial coordinates, and topology-derived features in an end-to-end framework. The approach achieves state-of-the-art clustering performance across 11 tissue slices (DLPFC, HBC, MBA), with substantial ARI gains and validated through ablations showing the critical roles of EPH and DualRO. While delivering strong practical impact for deciphering tissue architecture, the method incurs computational overhead due to EPH, highlighting a trade-off between topology fidelity and scalability.

Abstract

Spatial transcriptomics clustering is pivotal for identifying cell subpopulations by leveraging spatial location information. While recent graph-based methods modeling cell-cell interactions have improved clustering accuracy, they remain limited in two key aspects: (i) reliance on local aggregation in static graphs often fails to capture robust global topological structures (e.g., loops and voids) and is vulnerable to noisy edges; and (ii) dimensionality reduction techniques frequently neglect spatial coherence, causing physically adjacent spots to be erroneously separated in the latent space. To overcome these challenges, we propose SPHENIC, a Spatial Persistent Homology-Enhanced Neighborhood Integrative Clustering method. Specifically, it explicitly incorporates topology-invariant features into the clustering network to ensure robust representation learning against noise. Furthermore, we design a dual-regularized optimization module that imposes spatial constraints alongside distributional optimization, ensuring that the embedding space preserves the physical proximity of cells. Extensive experiments on 11 benchmark datasets demonstrate that SPHENIC outperforms state-of-the-art methods by 4.19%-9.14%, validating its superiority in characterizing complex tissue architectures.

Paper Structure

This paper contains 15 sections, 16 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Clustering in Spatial Transcriptomics
  4. Topological Data Analysis
  5. Methods
  6. Preliminaries
  7. Topological Invariant Extraction via EPH
  8. Multi-view Geometric Fusion Network
  9. Dual-Regularized Optimization
  10. Experiments
  11. Experiment Settings
  12. Clustering Performance
  13. Ablation Study
  14. Model Analysis
  15. Conclusions

Figures (5)

  • Figure 1: Comparison of static graph learning and invariant topological feature learning.
  • Figure 2: The proposed SPHENIC framework. Invariant topological information is incorporated into the multi-view graph fusion network to enhance the representation of different cells by EPH feature extraction. Then, the Dual-Regularized Optimization module is designed to ensure the spatial distribution and consistency of spatial clustering results.
  • Figure 3: Comparison experiments on HBC dataset between SPHENIC and other baselines. (a) Manual ground-truth annotations and histological image of the HBC dataset. (b) Spatial clustering result visualization by SCANPY, stLearn, SCGDL, DeepST, GraphST, spatial-MGCN, STAIG, and SPHENIC. (c) UMAP visualization of detected clusters by the above baselines.
  • Figure 4: Hyperparameter sensitivity analysis on HBC dataset and DLPFC-151507, respectively, in ARI term.
  • Figure 5: Visualization of spatial clustering by SPHENIC model in different training stages