CPPLS-MLP: a method for constructing cell-cell communication networks and identifying related highly variable genes based on single-cell sequencing and spatial transcriptomics data

Tianjiao Zhang, Zhenao Wu, Liangyu Li, Jixiang Ren, Ziheng Zhang, Guohua WangBriefings in Bioinformatics 2024 · DOI

TL;DR

This paper introduces CPPLS-MLP, a novel computational method that leverages single-cell and spatial transcriptomics data to construct detailed cell-cell communication networks. By integrating constrained partial least squares (CPPLS) and multilayer perceptrons (MLP), CPPLS-MLP not only identifies key genes involved in these interactions but also classifies them, offering a more nuanced understanding of cellular crosstalk than existing methods.

What Problem Does This Paper Address?

Understanding how cells communicate is fundamental to comprehending multicellular organism development, immune responses, and disease progression. While single-cell sequencing (scRNA-seq) and spatial transcriptomics (ST) have revolutionized our ability to study cellular heterogeneity, accurately mapping the intricate web of intercellular communication remains a significant challenge. Existing methods often focus on ligand-receptor pairs or downstream targets, but they struggle to integrate spatial information and simultaneously identify highly variable genes (HVGs) that are crucial for communication. Furthermore, there's a lack of standardized criteria to definitively link genes to cell communication. This paper aims to bridge this gap by developing a method that can construct comprehensive communication networks, identify cell-type-specific HVGs, and classify their roles in these interactions, all while accounting for spatial context.

Methods

CPPLS-MLP combines two powerful techniques:

  • Constrained Partial Least Squares (CPPLS): This regression method is adapted to handle the "small N, large P" problem common in ST data (fewer samples, many genes). It models the relationship between intercellular communication signals (represented by spatial coordinates and cell types) and HVG expression. CPPLS is chosen for its ability to manage continuous, discrete, and combined data, reflecting the complex interactions within tissues.
  • Multilayer Perceptron (MLP): This feedforward neural network is used for gene classification. After identifying potential communication-related genes, MLP leverages its non-linear mapping capabilities to categorize them based on predefined criteria derived from biological databases.

The method involves several key steps:

  1. Data Preprocessing: Gene expression matrices, spatial coordinate matrices, and cell type labels are prepared.
  2. Cell Communication Network Construction: CPPLS is used to model the impact of neighboring cell types on HVG expression, generating a coefficient matrix that reflects the direction and magnitude of regulation.
  3. MIMO Graph Construction: The interactions are visualized as a directed graph, where edge weights are attenuated based on the number of intermediate cells in a communication pathway, reflecting the weakening of signals over distance.
  4. HVG Clustering: K-means clustering is applied to group HVGs with similar communication profiles.
  5. Gene Identification and Classification: CPPLS-MLP filters top HVGs and then uses MLP to classify them into "strong correlation" (e.g., receptor, ligand) and "weak correlation" (e.g., pathway, regulation) categories based on annotations from GeneCards and Gene Ontology.

Datasets

The CPPLS-MLP method was evaluated using two publicly available datasets:

  • Seq-Scope: A spatial transcriptomics dataset from the mouse colon.
  • seqFISH+: A spatial transcriptomics dataset from the mouse somatosensory cortex.

These datasets represent different organs and cell types, allowing for a robust assessment of the method's generalizability.

Key Insights and Ideas

  • Spatial context is crucial: The method explicitly incorporates spatial coordinates, recognizing that cell proximity significantly influences communication. This moves beyond purely expression-based inference.
  • Beyond Ligand-Receptor pairs: CPPLS-MLP doesn't solely rely on known ligand-receptor interactions. It identifies genes whose expression is modulated by intercellular communication, offering a broader view of regulatory mechanisms.
  • Novel gene classification: The introduction of "strong" and "weak" correlation labels, derived from biological databases and refined by MLP, provides a more nuanced way to understand the functional roles of genes in cell communication.
  • MIMO system modeling: The method accounts for multiple inputs and multiple outputs in cellular communication, reflecting the complex, interconnected nature of biological systems.

Key Results

  • CPPLS-MLP demonstrated superior performance in identifying cell-type-specific HVGs compared to existing methods like CCPLS, showing higher uniqueness in HVG identification across various cell types in both Seq-Scope and seqFISH+ datasets.
  • The method achieved higher accuracy in classifying genes involved in cell communication compared to other classification algorithms (LightGBM, Single Tree, Random Forest) on both benchmark datasets.
  • CPPLS-MLP successfully identified known intercellular communication pathways, such as between fibroblasts and B.cell_IgA cells in the colon, and between L5 and L6 neurons in the cortex, validating its biological relevance.
  • The method identified a significant number of genes related to cell communication, outperforming established tools like CellphoneDB and Cellchat in terms of the number of identified genes.

Limitations and Future Work

  • Limited datasets: The authors acknowledge that their performance assessment was based on only two datasets, albeit representative ones. They suggest that integrating more multimodal datasets (e.g., 10x Multiome, Digital Spatial Profiling) would further enhance the method's capabilities.
  • Image processing for ST data: The authors propose utilizing image processing techniques to extract precise cell location information from tissue slice images, which could enrich the data foundation and overcome limitations when only gene expression data is available.
  • Future applications: The authors envision CPPLS-MLP being valuable for exploring biologically meaningful cell-cell interactions and regulatory mechanisms in various tissues and disease contexts. They also highlight the need for reliable computational models to accurately integrate multimodal data for inference.

Summary updated

BibTeX
@article{zhang2024cppls,
  author = {Tianjiao Zhang and Zhenao Wu and Liangyu Li and Jixiang Ren and Ziheng Zhang and Guohua Wang},
  title = {{CPPLS-MLP: a method for constructing cell-cell communication networks and identifying related highly variable genes based on single-cell sequencing and spatial transcriptomics data}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  doi = {10.1093/bib/bbae198},
}
Figure