Cobolt: Joint analysis of multimodal single-cell sequencing data

Boying Gong, Yun Zhou, Elizabeth Purdom

TL;DR

Cobolt is a new computational framework for analyzing single-cell multi-omics data. It can integrate data from different single-cell technologies, including those that measure multiple molecular types (like gene expression and chromatin accessibility) from the same cell, as well as data from technologies that only measure one type. This allows for a more comprehensive understanding of cellular states and subtypes.

What Problem Does This Paper Address?

Single-cell sequencing technologies have revolutionized our ability to study biological systems at the cellular level. Increasingly, researchers are using platforms that can simultaneously measure multiple types of molecular information (e.g., gene expression and chromatin accessibility) from the same cell. This "multi-omics" approach promises deeper insights into cellular function and regulation. However, analyzing such complex, multi-modal data presents significant computational challenges. Furthermore, single-modality datasets (measuring only one type of molecule) are far more common and often have higher throughput. The key challenge is to effectively integrate these diverse datasets – both multi-modal and single-modality – to derive a unified, meaningful representation of the cells. Existing methods often struggle to handle this integration seamlessly, leading to incomplete or biased analyses.

Methods

Cobolt employs a novel application of a Multimodal Variational Autoencoder (MVAE) within a hierarchical Bayesian generative model. The core idea is to project data from different modalities into a shared, lower-dimensional latent space. This latent space aims to capture the underlying biological signal common across all cells and modalities.

Key aspects of the method include:

  • Shared Latent Space: Cobolt learns a single representation for each cell, regardless of whether it comes from a joint-modality platform or a single-modality dataset.
  • Transfer Learning for Integration: For single-modality datasets, Cobolt uses a transfer learning approach. It leverages the information learned from joint-modality data to infer the latent representation for cells with only one type of molecular data.
  • Hierarchical Generative Model: The model is inspired by Latent Dirichlet Allocation (LDA) and assumes that each cell can be characterized by a mixture of underlying biological "categories" or processes. Different modalities (e.g., gene expression, chromatin accessibility) reflect these categories differently.
  • Handling Missing Modalities: The framework is designed to handle cases where not all cells have measurements for all modalities. It can still estimate latent variables for cells with incomplete data.
  • Batch Correction: Cobolt incorporates a mechanism to adjust for technical artifacts ("batch effects") that can arise from different experimental runs or platforms, improving the robustness of the integration.

Datasets

The paper uses several datasets to demonstrate Cobolt's capabilities:

  • SNARE-seq data: A multi-modality dataset measuring paired transcription and chromatin accessibility from adult mouse cerebral cortices (10,309 cells). This dataset is used for both single-modality analysis and as a benchmark for comparing methods.
  • Mouse Cortex Integration Data: This involves integrating the SNARE-seq data with:
  • MOp scRNA-seq data: Single-cell RNA sequencing data from mouse primary motor cortex.
  • MOp scATAC-seq data: Single-cell ATAC sequencing data from mouse primary motor cortex.
  • Human PBMCs Integration Data: This dataset integrates:
  • 10X Multiome datasets: Two multi-modality datasets (ATAC + mRNA) from human peripheral blood mononuclear cells (PBMCs) from different donors (10,970 and 12,012 cells).
  • 10X scRNA-seq data: Single-cell RNA sequencing data from human PBMCs (23,837 cells).
  • 10X scATAC-seq data: Single-cell ATAC sequencing data from human PBMCs (9,030 cells).

The paper also mentions using publicly available datasets from 10x Genomics and the NeMO Archive.

Key Insights and Ideas

  • Unified Representation is Key: The core insight is that a single, shared latent representation is crucial for truly integrating multi-modal and single-modality data. This allows for downstream analyses (like clustering) to be performed on all cells together, regardless of their origin or modality.
  • Transfer Learning Bridges the Gap: Cobolt's ability to use transfer learning from joint-modality data to infer representations for single-modality data is a powerful mechanism for leveraging high-throughput single-cell datasets.
  • Beyond Simple Modality-Specific Analysis: Analyzing each modality separately and then trying to link them (a common approach) often misses important biological signals that are only apparent when modalities are jointly modeled. Cobolt demonstrates that its integrated approach can reveal subtypes missed by single-modality analyses.
  • Robustness to Data Sparsity and Quality: The count-based model and the use of LDA-inspired generative distributions make Cobolt robust to sparse data and low-quality cells, which are common issues in single-cell experiments.

Key Results

  • Improved Subtype Detection: Cobolt successfully identified distinct cell subtypes in multi-modality data that were missed by other methods (e.g., MOFA2, scMM, BABEL) when analyzing only gene expression or chromatin accessibility separately.
  • Effective Integration of Single-Modality Data: When integrating multi-modality data with single-modality data, Cobolt produced a UMAP visualization where cells from different datasets and modalities aligned well, indicating successful integration.
  • Quantitative Evaluation of Integration: Using a nearest-neighbor metric, Cobolt showed superior performance in placing single-modality cells close to their paired counterparts in the multi-modality data compared to other methods like LIGER and Signac. This metric is directly related to the ability to correctly cluster cells across modalities.
  • Hierarchical Cell Type Representation: Cobolt's latent space captured not only broad cell classes but also finer subtypes, demonstrating its ability to resolve biological signals at different resolutions. For example, it could distinguish between subtypes of GABAergic inhibitory neurons.

Limitations and Future Work

  • Limitations:
    • While Cobolt performs well, the paper notes that subtle differences can exist between posterior distributions derived from different subsets of modalities, which might slightly affect downstream analyses like clustering.
  • The paper acknowledges that for optimal integration, it would be ideal to re-call peaks across all datasets, but shows that mapping to existing peak sets can still yield good results.
  • Future Work:
    • The authors suggest that Cobolt's framework can be extended to handle more than two modalities and different combinations of modalities.
  • Future work could explore using the graphical model and inferred parameters to construct gene networks and establish connections between features.
  • Pseudo-time estimation for reconstructing developmental trajectories is mentioned as another potential downstream application.

Summary updated

Abstract

Abstract A growing number of single-cell sequencing platforms enable joint profiling of multiple omics from the same cells. We present Cobolt, a novel method that not only allows for analyzing the data from joint-modality platforms, but provides a coherent framework for the integration of multiple datasets measured on different modalities. We demonstrate its performance on multi-modality data of gene expression and chromatin accessibility and illustrate the integration abilities of Cobolt by jointly analyzing this multi-modality data with single-cell RNA-seq and ATAC-seq datasets.
BibTeX
@inproceedings{gong2021cobolt,
  author = {Boying Gong and Yun Zhou and Elizabeth Purdom},
  title = {{Cobolt: Joint analysis of multimodal single-cell sequencing data}},
  booktitle = {Genome},
  year = {2021},
  doi = {10.1186/s13059-021-02556-z},
}
Figure