Panoramic stitching of heterogeneous single-cell transcriptomic data

Brian Hie, Bryan D. Bryson, Bonnie BergerbioRxiv 2018 · DOI

TL;DR

This paper introduces Scanorama, a novel method for integrating diverse single-cell RNA sequencing (scRNA-seq) datasets. Unlike previous approaches that often merge dissimilar cell types, Scanorama intelligently identifies and combines only shared cell populations, leading to more biologically accurate results and significantly faster processing times.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) is a powerful tool for understanding cellular function and diversity. Researchers are generating vast amounts of scRNA-seq data from various experiments, labs, and technologies. To gain a comprehensive view of biology and disease, it's crucial to integrate these disparate datasets. However, existing integration methods often struggle with datasets that have different cell compositions or experimental biases. They tend to merge data indiscriminately, even when cell types don't overlap, leading to misleading biological interpretations. This paper addresses the need for a method that can accurately integrate heterogeneous scRNA-seq datasets while preserving distinct cell populations and overcoming technical variations.

Methods

Scanorama is inspired by panorama stitching algorithms used in computer vision. Its core strategy involves:

  • Mutual Nearest Neighbors (MNN) Matching: Instead of pairwise comparisons, Scanorama finds mutual nearest neighbors across all pairs of datasets. This means it looks for cells in one dataset that are closest to cells in another, and vice-versa, ensuring a robust match.
  • Dimensionality Reduction: To speed up the computationally intensive nearest neighbor search, Scanorama first compresses the high-dimensional gene expression data into a lower-dimensional embedding using a randomized Singular Value Decomposition (SVD).
  • Approximate Nearest Neighbor Search: Further acceleration is achieved by employing an approximate nearest neighbor search algorithm based on hyperplane locality sensitive hashing and random projection trees.
  • Nonlinear Merging: Once mutual nearest neighbor matches are identified, Scanorama uses a nonlinear batch-correction strategy to merge datasets. This involves computing translation vectors between matched cells and progressively building larger "panoramas" of integrated data.

Datasets

The paper evaluates Scanorama using a variety of datasets:

  • Simulated Data: Used to test the method's ability to handle non-overlapping cell types.
  • Jurkat and 293T Cell Lines: Three datasets were used: one of Jurkat cells, one of 293T cells, and a 50:50 mixture of both. This was used to demonstrate Scanorama's ability to separate distinct cell types where other methods fail.
  • Large-Scale Integration: The primary benchmark involved integrating 26 diverse scRNA-seq datasets from 9 different technologies, totaling 105,476 cells. This dataset includes:
  • 293T cells
    • Jurkat cells
    • Mouse neurons
    • Macrophages (Mtb infected and uninfected)
    • Mouse hematopoietic stem cells (HSCs)
    • Human pancreatic islet cells
    • Peripheral blood mononuclear cells (PBMCs)
    • Various immune cell subtypes (B cells, monocytes, T cells, NK cells)

Key Insights and Ideas

  • "Panoramic Stitching" for Single Cells: The analogy to image panorama stitching is powerful. Just as stitching combines overlapping images into a seamless whole, Scanorama stitches together scRNA-seq datasets by identifying shared cellular "features" (cell types).
  • Robustness to Heterogeneity: The core innovation is Scanorama's ability to handle datasets with no shared cell types, a critical limitation of previous methods. This is achieved by focusing on mutual nearest neighbors, ensuring that only genuinely shared cell populations are merged.
  • Speed is Key: The algorithmic optimizations (randomized SVD and approximate nearest neighbors) are crucial for making large-scale integration feasible. The dramatic speedup over existing methods is a major practical advantage.
  • Interpretable Alignments: The alignment vectors generated by Scanorama can be interpreted to understand the biological basis of discrepancies between datasets, offering insights into processes like infection responses.

Key Results

  • Accurate Integration: Scanorama successfully integrates heterogeneous datasets, clustering cells by cell type rather than experimental batch. This is demonstrated by a median Silhouette Coefficient of 0.34 on a large dataset, compared to 0.01 for scran MNN and -0.20 for Seurat CCA.
  • Preservation of Cell Types: Scanorama correctly separates distinct cell types, as shown in the Jurkat/293T experiment where other methods failed to distinguish them.
  • Significant Speedup: For a large integration task (26 datasets, 105,476 cells), Scanorama completed the nearest neighbor search in ~5 minutes and batch correction in <25 minutes. In contrast, existing methods took over 27 hours.
  • Biological Insights: Scanorama enabled the identification of a rare subpopulation of pancreatic beta cells with upregulated ER stress genes and revealed genes enriched for bacterial infection processes in macrophage datasets.

Limitations and Future Work

  • Computational Efficiency: While significantly faster, the paper notes that further improvements in computational efficiency could be achieved through random or structure-preserving sampling of data.
  • Unknown Cell States: Scanorama can be used to interrogate systems with unknown cell states, suggesting its potential for discovering novel biological patterns.
  • Modularity: The authors highlight that Scanorama is modular and can be combined with other tools for clustering and visualization, indicating potential for broader applications.
  • Data Quality: The paper mentions removing low-quality cells based on gene counts, implying that data quality is a prerequisite for effective integration. Future work might explore how Scanorama handles datasets with varying levels of data quality.

Summary updated

Abstract

Abstract Researchers are generating single-cell RNA sequencing (scRNA-seq) profiles of diverse biological systems 1–4 and every cell type in the human body. 5 Leveraging this data to gain unprecedented insight into biology and disease will require assembling heterogeneous cell populations across multiple experiments, laboratories, and technologies. Although methods for scRNA-seq data integration exist 6,7 , they often naively merge data sets together even when the data sets have no cell types in common, leading to results that do not correspond to real biological patterns. Here we present Scanorama, inspired by algorithms for panorama stitching, that overcomes the limitations of existing methods to enable accurate, heterogeneous scRNA-seq data set integration. Our strategy identifies and merges the shared cell types among all pairs of data sets and is orders of magnitude faster than existing techniques. We use Scanorama to combine 105,476 cells from 26 diverse scRNA-seq experiments across 9 different technologies into a single comprehensive reference, demonstrating how Scanorama can be used to obtain a more complete picture of cellular function across a wide range of scRNA-seq experiments.
BibTeX
@misc{hie2018panoramic,
  author = {Brian Hie and Bryan D. Bryson and Bonnie Berger},
  title = {{Panoramic stitching of heterogeneous single-cell transcriptomic data}},
  year = {2018},
  doi = {10.1101/371179},
  howpublished = {bioRxiv},
}
Figure