LANTSA: Landmark-based transferable subspace analysis for single-cell and spatial transcriptomics

Chuanchao Zhang, Lequn Wang, Xinxing Li, Wei-Feng Guo, Qianqian Shi, Luonan ChenbioRxiv 2022 · DOI

TL;DR

This paper introduces LANTSA, a novel framework for analyzing single-cell and spatial transcriptomics data. LANTSA efficiently captures complex sample relationships using a landmark-based subspace analysis and enables accurate transfer of cell type labels across different datasets and technologies, overcoming computational and cross-platform challenges.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics (ST) are powerful tools for understanding biological systems at a fine resolution. However, analyzing large datasets or integrating data from different sources (e.g., scRNA-seq and ST, or different scRNA-seq protocols) is computationally demanding and challenging due to inherent biological variability and technical differences. Existing methods often struggle to accurately represent these complex relationships, especially for massive datasets, leading to difficulties in tasks like cell clustering, visualization, and transferring annotations between datasets.

Methods

LANTSA employs a two-pronged approach:

  • Landmark-based Subspace Analysis (for single datasets): Instead of analyzing all samples, LANTSA selects a representative subset of "landmark" samples. It then constructs a "representation matrix" that captures the relationships between these landmarks and all other samples. This is achieved through a novel subspace model that preserves both global and local data structures. This approach significantly reduces computational cost (linear in data size) and memory usage compared to traditional methods that consider all pairwise sample relationships.
  • Transfer Strategy (for multiple/large datasets): For integrating datasets or handling large-scale data, LANTSA uses a "learning and prediction" strategy. It first learns a "discriminant matrix" from a well-annotated "learning" dataset (e.g., scRNA-seq). This matrix then projects both the learning and "prediction" datasets into a shared space. Cell labels from the learning dataset can then be transferred to the prediction dataset (which could be scRNA-seq or ST data) using a classifier built on the discriminant vectors. This allows for cross-platform and cross-technology label transfer.

Datasets

LANTSA was evaluated on a variety of datasets, including:

  • scRNA-seq datasets: Ten benchmark datasets with varying technical and biological characteristics were used for clustering evaluation. A large 1.3 million mouse brain cell dataset was used to assess scalability.
  • Spatial transcriptomics datasets: 14 10x Visium arrays were used for spatial clustering.
  • Cross-platform integration: snRNA-seq and scRNA-seq cerebellum datasets were used to evaluate label transfer capabilities.
  • Mouse brain data: A 10x Visium mouse brain dataset was used for in-depth spatial analysis.

The paper references existing benchmark datasets and their ground truth labels, making it clear that no new benchmark datasets were introduced.

Key Insights and Ideas

  • Landmarks as efficient representatives: The core idea of using a small set of "landmark" samples to represent the entire dataset is a clever way to tackle the computational burden of large-scale transcriptomics data.
  • Subspace analysis for better representation: Moving beyond simple dimensionality reduction, LANTSA's subspace analysis aims to capture more nuanced relationships between samples, which is crucial for complex biological data.
  • Transfer learning for cross-platform integration: The ability to learn from one dataset (e.g., scRNA-seq) and apply that knowledge to another (e.g., ST) is a significant step towards harmonizing data from different sources.
  • Preservation of global structure and cell separation: The discriminant matrix learned by LANTSA effectively preserves the overall data structure while simultaneously enhancing the ability to distinguish between different cell types.

Key Results

  • Superior Clustering Performance: LANTSA generally outperformed six popular scRNA-seq clustering methods on ten benchmark datasets, achieving higher Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) scores.
  • Efficient for Large Datasets: LANTSA processed datasets with up to 1.3 million cells in under a minute, demonstrating remarkable scalability. Its runtime scales linearly with the number of landmarks.
  • Accurate Spatial Clustering: On 10x Visium spatial transcriptomics data, LANTSA outperformed most other methods, accurately identifying anatomical structures in mouse brain slices.
  • Effective Cross-Platform Label Transfer: LANTSA achieved high prediction accuracy (0.97 for scRNA-seq and 0.95 for snRNA-seq) when transferring cell annotations between different platforms, showing insensitivity to reference platforms.
  • Robustness to Parameters: LANTSA's clustering performance was found to be robust to variations in its tunable parameters.

Limitations and Future Work

  • Focus on Sample Representation: The authors acknowledge that LANTSA primarily focuses on sample representation and does not directly link sample features to functional gene modules. This limits its direct application for in-depth biological interpretation of gene functions.
  • Separate Analysis of Sample and Feature Representations: Similar to other methods, LANTSA analyzes sample structure and feature identification separately. The paper suggests that a joint analysis of sample and feature representations would be beneficial for future work.
  • Potential for Untraceable Biological Variabilities: The splitting solution in the analysis might lead to untraceable biological variabilities and potentially biased outcomes, warranting further investigation into joint analysis.

Summary updated

Abstract

ABSTRACT Single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics (ST) technologies provide new insights to understand tissue organization and biological function. Accurately capturing the relationships of samples (e.g., sequenced cells, spatial locations) will result in reliable and consistent outcomes in downstream analyses. However, this undertaking remains a challenge for large-volume or cross-platform datasets due to transcriptional heterogeneity and high computational demands. Here, we introduce landmark-based transferable subspace analysis (LANTSA) to solve such challenges for scRNA-seq and ST datasets. Specifically, LANTSA constructs a representation graph of samples for clustering and visualization based on a novel subspace model, which can learn a more accurate representation and is theoretically proven to be linearly proportional to data size in terms of the time consumption. Furthermore, LANTSA uses a dimensionality reduction technique as an integrative method to extract the discriminants underlying the representation structure, which enables label transfer from one (learning) dataset (i.e., scRNA-seq profiles) to the other (prediction) datasets (e.g., scRNA-seq or ST profiles), thus solving the massive-volume or cross-platform problem. We demonstrated the superiority of LANTSA to identify accurate data structures via clustering evaluation on benchmark datasets of various scRNA-seq protocols, 10x Visium, and Slide-seq ST platforms. Moreover, we confirmed the integration capability of LANTSA to transfer cell annotation on large-scale and cross-platform scRNA-seq datasets. Finally, we validated the effectiveness of LANTSA for the identification of multiple mouse brain areas as well as the spatial mapping of cell types within cortical layers by integrating scRNA-seq and ST data.
BibTeX
@misc{zhang2022lantsa,
  author = {Chuanchao Zhang and Lequn Wang and Xinxing Li and Wei-Feng Guo and Qianqian Shi and Luonan Chen},
  title = {{LANTSA: Landmark-based transferable subspace analysis for single-cell and spatial transcriptomics}},
  year = {2022},
  doi = {10.1101/2022.03.13.484116},
  howpublished = {bioRxiv},
}

Loading related content...

Figure