Reconstruction of the cell pseudo-space from single-cell RNA sequencing data with scSpace

Jie Liao, Jingyang Qian, Ziqi Liu, Ying Chi, Yanrong Zheng, Xin Shao, Junyun Cheng, Yongjin Cui, Wenbo Guo, Penghui Yang, Yining Hu, Hudong Bao, Qian Chen, Mingxiao Li, Bing Zhang, Xiaohui FanbioRxiv 2022 · DOI

TL;DR

This paper introduces scSpace, a novel computational method that reconstructs the spatial organization of cells from single-cell RNA sequencing (scRNA-seq) data. By integrating scRNA-seq with spatial transcriptomics references, scSpace creates a "pseudo-space" that reveals cell-cell relationships and identifies biologically meaningful subpopulations that are missed by traditional methods, offering new insights into tissue architecture and disease mechanisms.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of cellular heterogeneity by providing high-resolution gene expression profiles for individual cells. However, a major limitation of scRNA-seq is that the process of dissociating tissues into single cells destroys crucial spatial information. This loss of spatial context makes it difficult to understand how cells are organized within a tissue and how this organization influences their function and interactions. While spatial transcriptomics technologies can capture spatial gene expression, they often lack single-cell resolution. This paper addresses the gap between these two technologies by developing a computational approach, scSpace, to infer and reconstruct the spatial relationships of cells from scRNA-seq data, effectively bridging the divide between cellular identity and tissue organization.

Methods

The core of scSpace is a transfer learning approach that integrates scRNA-seq data ($X_s$) with spatial transcriptomics reference data ($X_t$). The process involves three key steps:

  1. Latent Biological Feature Representation Extraction: scSpace uses Transfer Component Analysis (TCA) to eliminate batch effects between the two data types and extract a shared latent feature representation that captures true biological characteristics. This step aims to find a transformation $\phi$ such that the distributions of the transformed data, $\phi(X_s)$ and $\phi(X_t)$, are similar.
  2. Spatial Reconstruction: A multi-layer fully connected neural network is trained to map the latent biological features to spatial coordinates. This model, trained on the spatial transcriptomics data, is then used to predict spatial coordinates for the cells in the scRNA-seq data, creating a "pseudo-space."
  3. Space-Informed Clustering: scSpace constructs a gene expression graph and a spatial graph of cells. The spatial graph is then used to transform the gene expression graph into a "space-informed expression graph" by assigning spatial weights to edges. Finally, unsupervised clustering (using the Leiden algorithm) is applied to this space-informed graph to identify spatially heterogeneous cell subpopulations.

Datasets

The authors utilized several datasets for validation and demonstration:

  • Simulated Data: 50 paired scRNA-seq and spatial transcriptomics datasets were simulated using the Splatter R package to mimic batch effects and varying numbers of cell populations and subclusters.
  • Human Dorsolateral Prefrontal Cortex (DLPFC) Data: Human DLPFC spatial transcriptomics data (Visium) was used as a spatial reference, with other slices serving as simulated scRNA-seq data.
  • Mouse V1 Neocortex Data: Mouse V1 neocortex spatial transcriptomics data (STARmap) was used as a spatial reference, with corresponding scRNA-seq data downsampled.
  • Mouse Intestine Data: Mouse intestine scRNA-seq data and corresponding spatial transcriptomics data were used.
  • Mouse Liver Data: Mouse liver scRNA-seq data and its spatial transcriptomics reference were used.
  • Human Developing Heart Data: Human developing heart scRNA-seq data and its spatial transcriptomics reference were used.
  • Human Cortex scRNA-seq Data: Data from Allen Brain Atlas was used to discover fine subpopulations.
  • Human Melanoma scRNA-seq Data: scRNA-seq data from melanoma patients was used to study T cell subpopulations and exhaustion.
  • COVID-19 Data: Single-cell lung atlas data of lethal COVID-19 and normal human lung spatial transcriptomics data were used to study myeloid subpopulation invasion.

Key Insights and Ideas

  • Bridging the scRNA-seq and Spatial Transcriptomics Gap: The primary innovation is the development of a method that effectively integrates information from both scRNA-seq (cellular identity) and spatial transcriptomics (tissue organization) to reconstruct spatial relationships at the single-cell level.
  • Pseudo-space as a Powerful Representation: The concept of a "pseudo-space" generated by scSpace is crucial. It's not just a visualization but a mathematically derived representation that captures both transcriptional and spatial information, enabling more accurate clustering and analysis.
  • Uncovering Hidden Heterogeneity: scSpace can identify biologically meaningful cell subpopulations that are missed by traditional methods that rely solely on transcriptional information. This is particularly important for understanding complex tissues and diseases.
  • Spatial Context is Key for Cell Identity: The study strongly emphasizes that spatial information is as important as transcriptional information for defining cell identity and function, especially in complex tissues like the brain and in disease contexts.
  • Therapeutic Marker Discovery: The application of scSpace in melanoma and COVID-19 highlights its potential for identifying spatially relevant therapeutic targets by revealing how cell subpopulations interact within their microenvironment.

Key Results

  • Improved Clustering Accuracy: scSpace consistently outperformed traditional clustering methods (Louvain, K-means, Hierarchical clustering) on simulated datasets, achieving significantly higher Adjusted Rand Index (ARI) scores for both all clusters and spatially variable subclusters. For example, on simulated data, scSpace achieved an average ARI of 0.914, significantly higher than other methods.
  • Accurate Spatial Reconstruction: scSpace successfully reconstructed the hierarchical structure of human DLPFC layers and mouse V1 neocortex, with predicted spatial coordinates showing high correlation with original locations.
  • Identification of Novel Subpopulations: In human cortex, scSpace identified 19 refined subpopulations, including two with distinct spatial patterns (C10 and C15) that were difficult to distinguish by Louvain. In melanoma, it identified T cell subpopulations with varying distances to malignant cells, correlating with T cell exhaustion. In COVID-19, it revealed distinct spatial relationships of myeloid cells and identified subpopulations with altered infiltration.
  • Validation of Biomarkers: The identified cell subclusters and their spatial distributions were confirmed by the expression patterns of known marker genes.
  • Clinical Relevance: In melanoma, scSpace identified T cell subpopulations with high T cell exhaustion scores closer to malignant cells, and survival analysis of marker genes in these subpopulations suggested potential therapeutic targets. In COVID-19, it highlighted the role of specific myeloid subpopulations and identified pathways crucial for disease progression.

Limitations and Future Work

  • Dependence on Spatial References: The accuracy of scSpace's reconstruction is dependent on the availability and quality of spatial transcriptomics reference data. The method assumes that the reference data captures similar biological characteristics and spatial organization.
  • Computational Cost: While not explicitly detailed as a limitation, the integration of multiple data types and complex algorithms likely involves significant computational resources.
  • Future Work: The authors suggest that scSpace has broad prospects for discovering spatial therapeutic markers and gaining new insights into diseases like melanoma and COVID-19. Further applications in other complex tissues and diseases are implied. The method's ability to reveal spatial relationships and identify spatially variable subpopulations is a key direction for future research.

Summary updated

Abstract

Abstract Tissues are highly complicated with spatial heterogeneity in gene expression. However, the cutting-edge single-cell RNA-seq technology eliminates the spatial information of individual cells, which contributes to the characterization of cell identities. Herein, we propose s ingle- c ell s patial p osition a ssociated c o- e mbeddings (scSpace), an integrative algorithm to distinguish spatially variable cell subclusters by reconstructing cells onto a pseudo-space with spatial transcriptome references (Visium, STARmap, Slide-seq, etc.). We demonstrated that scSpace can define biologically meaningful cell subpopulations neglected by single-cell RNA-seq or spatially resolved transcriptomics. The use of scSpace to uncover the spatial association within single-cell data, reproduced, the hierarchical distribution of cells in the brain cortex and liver lobules, and the regional variation of cells in heart ventricles and the intestinal villus. scSpace identified cell subclusters in intratelencephalic neurons, which were confirmed by their biomarkers. The application of scSpace in melanoma and Covid-19 exhibited a broad prospect in the discovery of spatial therapeutic markers.
BibTeX
@misc{liao2022reconstruction,
  author = {Jie Liao and Jingyang Qian and Ziqi Liu and Ying Chi and Yanrong Zheng and Xin Shao and Junyun Cheng and Yongjin Cui and Wenbo Guo and Penghui Yang and Yining Hu and Hudong Bao and Qian Chen and Mingxiao Li and Bing Zhang and Xiaohui Fan},
  title = {{Reconstruction of the cell pseudo-space from single-cell RNA sequencing data with scSpace}},
  year = {2022},
  doi = {10.1101/2022.05.07.491043},
  howpublished = {bioRxiv},
}
Figure