TL;DR
This paper introduces "stKeep," a novel heterogeneous graph learning framework that integrates diverse data modalities (histology, gene expression, spatial location, gene interactions) to dissect the complex tumor microenvironment (TME). stKeep effectively identifies finer cell states, their specific gene programs, and intercellular communication, outperforming existing methods and offering new avenues for cancer prognosis and immunotherapy.
What Problem Does This Paper Address?
Understanding the tumor microenvironment (TME) is crucial for comprehending cancer progression and developing effective treatments. The TME is a complex ecosystem where cancer cells interact with non-cancer cells, influenced by both internal molecular networks and external cell-cell communication (CCC). Spatially resolved transcriptomics (SRT) technologies provide unprecedented detail by capturing gene expression within its spatial context. However, analyzing this rich, multimodal data presents significant computational challenges. Existing methods often struggle to fully capture the intricate relationships between cells, genes, and histological regions, limiting our ability to interpret the TME's complexity and identify key drivers of disease. This paper aims to bridge this gap by developing a computational framework that can integrate these diverse data sources to provide a more comprehensive understanding of TME heterogeneity.
Methods
The core of this work is stKeep, a heterogeneous graph (HG) learning method. stKeep constructs a unified graph where nodes represent diverse entities like genes, cells/spots, and histological regions. Edges represent the relationships between these entities, including gene expression, spatial proximity, and known gene-gene interactions (like gene regulatory networks - GRNs, and protein-protein interactions - PPIs).
stKeep employs a three-module approach:
- Cell Module: Learns "cell-modules" by integrating local hierarchical representations (from genes and regions) with global semantic representations (from neighboring cells/spots). This is achieved through a self-supervised learning framework that uses contrastive loss to ensure representations are comparable across different cell states.
- Gene Module: Learns "gene-modules" by aggregating information from cells/spots and cell states, leveraging known gene-gene interactions (GRN, PPI) through contrastive learning. This helps identify gene modules specific to particular cell states.
- Cell-Cell Communication (CCC) Module: Infers ligand-receptor interactions (LRPs) by using attention-based heterogeneous graphs to aggregate ligand signals from neighboring cells. Crucially, it uses contrastive learning to ensure that learned CCC patterns are comparable across different cell states, allowing for the identification of disease-relevant LRPs.
The method leverages attention mechanisms and contrastive learning to effectively integrate multimodal data and capture complex relationships.
Datasets
The authors utilized several datasets for training and evaluation:
- Human Dorsolateral Prefrontal Cortex (DLPFC) slices: 12 slices used for evaluating cell and gene module identification and CCC inference, compared against existing methods.
- Breast Cancer Samples:
- IDC (Luminal B) and BAS1 (Her2+) samples from 10X Genomics for evaluating cell-module and gene-module identification.
- Triple-negative breast cancer (TNBC) sample for dissecting intratumor heterogeneity and identifying neoplastic myoepithelial cells.
- Colorectal Cancer and Liver Metastasis Samples: Primary colorectal cancer (P1) and paired liver metastasis (LM1, LM2) samples for identifying metastasis cells.
- Single-cell RNA sequencing (scRNA-seq) data: Used to deconvolute SRT data and provide cell type annotations for various cancer types.
Key Insights and Ideas
- Unified Framework for TME Dissection: stKeep's strength lies in its ability to integrate diverse data modalities (histology, spatial, gene expression, gene interactions) into a single heterogeneous graph, providing a holistic view of the TME.
- Hierarchical Representation Learning: By learning both local hierarchical and global semantic representations for cells, stKeep can capture finer cell states and their specific gene programs more effectively.
- Contrastive Learning for Comparability: The use of contrastive learning in both cell and CCC modules is key to ensuring that learned patterns are comparable across different cell states, enabling robust cross-state analysis.
- Identifying Novel Cell States: stKeep successfully identified cell states missed by other methods, such as bi-potent basal populations and neoplastic myoepithelial cells, highlighting its improved sensitivity.
- Functional Significance of Gene Modules: The identified gene modules were not just collections of genes but were linked to specific biological functions and, importantly, correlated with patient survival, underscoring their clinical relevance.
- Therapeutic and Prognostic Potential: The identification of key transcription factors, ligands, and receptors associated with disease progression suggests stKeep can uncover potential therapeutic targets and biomarkers for prognosis and immunotherapy.
Key Results
- Improved Clustering: stKeep's cell module showed better consistency with manual annotations on DLPFC slices compared to Squidpy, STAGATE, and stMVC, suggesting it dissects more heterogeneous cell populations.
- Accurate Gene Module Identification: Gene pairs identified by stKeep's gene module were more closely characterized and showed significantly higher expression correlation than random pairs. These modules also exhibited layer-specific patterns across multiple slices.
- Enhanced CCC Inference: stKeep's CCC inference showed higher biological relevance and specificity compared to COMMOT, as indicated by variations in correlations between spot pairs within clusters and randomly selected pairs.
- Detection of Rare Cell Populations: In breast cancer samples, stKeep identified cell states (e.g., cluster 21 in IDC, cluster 28 in BAS1) that were missed by other methods. These states were linked to key TFs and pathways relevant to cancer progression.
- Identification of Neoplastic Myoepithelial Cells: In TNBC samples, stKeep identified neoplastic myoepithelial cells and their associated gene programs, suggesting a role in cancer invasiveness.
- Metastasis Cell Identification: stKeep successfully identified metastasis cells in colorectal cancer and liver metastasis samples, revealing mechanisms of cancer cell invasion and metastasis.
Limitations and Future Work
- Histological Region Identification: Identifying histological regions from images can be time-consuming. The authors plan to develop computational models using SAM to automate this process.
- Gene-Gene Relation Directions: While stKeep identifies co-related gene pairs, it doesn't explicitly infer the direction of these relations. Future work will focus on developing algorithms to infer directed gene-gene relations using ATAC-seq or Chip-seq data.
- Inter-Tumor Heterogeneity: The current focus is on intra-tumoral heterogeneity. Future work aims to combine stKeep with other SRT integrative tools to address inter-tumor heterogeneity.
- Extensibility to Other Omics: The HG framework is flexible and can be extended to incorporate other spatial omics data like ATAC-seq, proteomics, and MSI by replacing the gene expression feature matrix.
- Computational Cost: While stKeep is relatively fast, processing very large datasets can still be computationally intensive. The authors note that running time is approximately linearly proportional to the number of input cells.
Summary updated