Mapping the Gene Space at Single-Cell Resolution: A Study on Gene Signal Pattern Analysis (GSPA)

Academic Background

Single-cell RNA sequencing (scRNA-seq) technology has made significant progress in biological research in recent years, particularly in revealing the organizational structure of the cellular state space. However, although many computational methods have been developed for mapping the cellular state space, research on the mapping or embedding of the gene space remains relatively limited. Gene expression is highly organized, with genes coordinating through complex biological processes and pathways. Yet, due to biological and technical noise (such as the “dropout” phenomenon), accurately quantifying the similarity between genes remains challenging. To address this, this paper proposes a novel method based on graph signal processing (GSP)—Gene Signal Pattern Analysis (GSPA)—which aims to learn rich gene representations from single-cell data and provide support for various biological tasks.

Source of the Paper

This paper was co-authored by Aarthi Venkat, Sam Leone, Scott E. Youlten, and others, with participation from multiple research institutions including Yale University and Boise State University. The paper was published in December 2024 in the journal Nature Computational Science, titled “Mapping the gene space at single-cell resolution with gene signal pattern analysis,” with DOI 10.1038/s43588-024-00734-0.

Research Process and Results

1. Introduction of the Gene Embedding Problem

The study first introduces the gene embedding problem, where gene expression patterns are treated as signals defined on a cell-cell graph. The goal is to construct a mapping from high-dimensional gene space to low-dimensional embedded space, preserving the distances between genes (based on the geometric structure of the cell-cell graph) while ensuring noise robustness and flexibility for downstream tasks.

2. Overview of the GSPA Model

The core idea of GSPA is to treat gene expression patterns as signals on the cell-cell graph and use diffusion wavelets for multi-scale decomposition. Specific steps include:
- Constructing the Cell-Cell Graph: Building the graph based on similarities in gene expression profiles between cells, and defining the diffusion operator to describe transition probabilities between cells.
- Constructing the Diffusion Wavelet Dictionary: Generating multi-scale wavelets through powering the diffusion operator to capture both local and global features of gene signals.
- Decomposition and Embedding of Gene Signals: Projecting each gene signal onto the diffusion wavelet dictionary to obtain its multi-scale representation, and using an autoencoder for dimensionality reduction to generate low-dimensional gene embeddings.

3. Experimental Results and Validation

The study validates the effectiveness of GSPA through simulated data and real single-cell datasets, including:
- Capturing Gene Co-expression Modules: GSPA accurately captures gene co-expression modules and preserves similarities between genes.
- Gene Localization Analysis: The “differential localization” method proposed by GSPA identifies genes expressed locally on the cell-cell graph, which are often closely related to changes in cell state.
- Downstream Applications: GSPA demonstrates its broad applicability in various areas, including cell-cell communication analysis (GSPA-LR), spatial transcriptomics (GSPA-Multimodal), and patient response prediction (GSPA-PT).

4. Case Studies

• Gene Co-expression in CD8+ T Cell Differentiation: The study analyzes CD8+ T cells during acute and chronic infections, identifying key gene modules related to T cell differentiation and revealing the unique role of interferon signaling in chronic infections.
  
• Cell-Cell Communication Analysis with GSPA-LR: GSPA-LR identifies ligand-receptor (LR) pair signal patterns without the need for cell type annotations, uncovering the role of the immune inhibitory receptor PD-1 in immune-related adverse events.
  
• Spatial Transcriptomics Analysis with GSPA-Multimodal: GSPA-Multimodal integrates gene expression and spatial affinity to identify spatially variable genes and reveals complex multicellular signaling networks in human lymph nodes.
  
• Patient Response Prediction with GSPA-PT: GSPA-PT constructs patient vectors to more accurately predict melanoma patients’ responses to immunotherapy and identifies key genes related to T cell function.

Conclusion and Significance

By treating gene expression as signals on the cell-cell graph and combining diffusion wavelets with deep learning techniques, GSPA provides a novel gene embedding method. It not only captures the complex relationships between genes but also offers powerful tools for various biological tasks, such as cell-cell communication, spatial transcriptomics, and patient response prediction. This study lays an important foundation for the field of gene space mapping and opens new research directions for future single-cell data analysis.

Highlights of the Study

• Novel Gene Embedding Method: GSPA is the first to apply graph signal processing techniques to single-cell gene expression data, proposing a diffusion wavelet-based gene embedding framework.
  
• Multi-scale Representation: By constructing a multi-scale diffusion wavelet dictionary, GSPA simultaneously captures local and global features of gene signals, improving the robustness and interpretability of gene embeddings.
  
• Broad Downstream Applications: GSPA can be used not only for gene module identification and cell-cell communication analysis but also extends to spatial transcriptomics and patient response prediction, demonstrating its strong versatility.
  
• Analysis Without Cell Type Annotation: GSPA-LR identifies ligand-receptor pair signal patterns without relying on cell type annotations, providing a more flexible tool for cell-cell communication analysis.
  

Other Valuable Information

The study has open-sourced the code for GSPA, making it accessible for other researchers to apply and improve the method. The code is available on GitHub. Additionally, validation results on multiple real-world datasets further demonstrate GSPA’s practicality and reliability in actual biological research.