Coexpression enables multi-study cellular trajectories of development and disease

Brian Hie, Hyunghoon Cho, Bryan D. Bryson, Bonnie BergerbioRxiv 2019 · DOI

TL;DR

This paper introduces Trajectorama, a novel algorithm that leverages gene coexpression within cell clusters to integrate diverse single-cell RNA sequencing (scRNA-seq) datasets. This approach successfully reconstructs complex developmental trajectories and reveals disease-specific gene networks, overcoming limitations of existing methods that often remove valuable biological variation during integration.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) generates vast amounts of data, enabling researchers to study complex biological systems at an unprecedented resolution. However, integrating data from different studies, labs, and technologies presents a significant challenge. Existing integration algorithms often aim to remove "study-specific" information to create a unified dataset. While this can reduce technical bias, it frequently eliminates crucial biological variation, particularly when trying to capture continuous processes like cell development or disease progression across multiple time points or conditions. This makes it difficult to reconstruct comprehensive, multi-study trajectories of cellular change, limiting our understanding of dynamic biological processes.

Methods

The core innovation of this paper is the shift from analyzing individual cells to analyzing clusters of cells as the fundamental unit. Trajectorama uses gene coexpression within these clusters as its primary feature.

Here's a breakdown of the key methods:

  • Panresolution Clustering: Instead of relying on a single clustering resolution, Trajectorama considers multiple resolutions to capture a hierarchical view of cell populations. This is crucial because coexpression patterns can change with clustering granularity.
  • Coexpression Matrix as Feature: For each cluster, a gene-by-gene Spearman correlation matrix is computed. This matrix, representing the coexpression patterns within the cluster, serves as the feature vector for downstream analysis. This approach is robust to technical variations in gene expression.
  • Coexpression Landscape: These coexpression matrices are then used to construct a "coexpression landscape," typically visualized using a k-nearest neighbors (KNN) graph. This landscape represents the relationships between cell clusters based on their coexpression profiles.
  • Diffusion Pseudotime (DPT): Analogous to pseudotime analysis in gene expression space, DPT is applied to the coexpression landscape to infer continuous trajectories and developmental ordering.
  • Dictionary Learning: To interpret the coexpression landscape, dictionary learning is employed to identify basis coexpression matrices ("dictionary entries") that represent fundamental patterns. These entries can be interpreted as gene networks.
  • Gene Ontology (GO) Enrichment: GO analysis is used to understand the biological processes associated with genes in specific dictionary entries or network modules.

Datasets

The paper utilizes several publicly available scRNA-seq datasets for its analyses:

  • Mouse Neuronal Development: Five large-scale studies profiling mouse brain development from embryonic to adult stages.
  • Human Hematopoiesis: Datasets from bone marrow, cord blood, fetal thymus, and peripheral blood.
  • Microglial Response to Myelin Injury: Integrated data from mouse and human microglia, including samples from different ages, injury models, and human patients with multiple sclerosis.

The paper does not introduce new benchmark datasets but rather demonstrates the utility of Trajectorama on existing, diverse datasets.

Key Insights and Ideas

  • Coexpression as a Robust Feature: Focusing on gene coexpression within cell clusters, rather than individual gene expression, provides a more robust feature that is less susceptible to technical noise and batch effects, while still capturing meaningful biological variation.
  • Unifying Development and Disease: Trajectorama successfully integrates data across disparate studies to reveal continuous developmental trajectories (e.g., neuronal development, hematopoiesis) and identify disease-specific alterations (e.g., microglial response to myelin injury).
  • Preserving Inter-Study Biological Variation: Unlike many existing integrative methods that aim to remove all study-specific signals, Trajectorama's coexpression-based approach preserves biologically relevant inter-study differences, which are crucial for understanding dynamic processes.
  • Interpretability of Coexpression Landscape: The coexpression landscape offers an intuitive way to visualize and interpret complex biological relationships, allowing for the identification of distinct cell states and their transitions.
  • Scalability: The algorithm is designed for efficiency, capable of processing millions of cells in a reasonable timeframe, making it suitable for large-scale consortium efforts.

Key Results

  • Neuronal Development Trajectory: Trajectorama successfully reconstructed a continuous trajectory of mouse neuronal development, ordered by age, from five disparate scRNA-seq studies. This trajectory was strongly correlated with biological age.
  • Hematopoiesis Lineages: The algorithm integrated data from four hematopoietic tissues, revealing organization consistent with erythropoiesis, lymphopoiesis, and myelopoiesis lineages.
  • Disease-Specific Microglial Network: Trajectorama identified a distinct coexpression network associated with myelin injury and neurodegenerative disease in microglia, highlighting key functional modules related to clearance, cytotoxicity, and inflammation. This network revealed novel associations between genes implicated in neurodegeneration.
  • Neuronal Subtype Integration: The method successfully integrated neuronal subtypes across studies, grouping them primarily by subtype rather than by study of origin, demonstrating its ability to resolve finer biological distinctions.
  • Computational Efficiency: Trajectorama can process nearly one million cells in approximately one hour on a standard cloud instance, showcasing its scalability.

Limitations and Future Work

  • Non-PSD Coexpression Matrices: The paper notes that coexpression matrices are not strictly positive semidefinite (PSD) for practical reasons. Learning large numbers of nontrivially sparse PSD matrices is identified as an interesting future challenge.
  • Alternative Coexpression Measures and Clustering: Exploring alternative methods for measuring coexpression and different clustering strategies (panresolution or otherwise) are suggested for future work.
  • Other Feature Spaces: The question of whether other feature spaces can enable multi-study trajectories is raised, with metric learning approaches mentioned as a possibility.
  • Relationship with Other Assocsations: Investigating the relationship between coexpression and other functional associations within single cells (e.g., chromatin accessibility, methylation) is highlighted as an important future consideration.
  • Experimental Validation: While the study identifies potential gene links for disease pathogenesis, experimental validation of causal roles is noted as important future work.

Summary updated

Abstract

Abstract Single-cell transcriptomic studies of diverse and complex systems are becoming ubiquitous. Algorithms now attempt to integrate patterns across these studies by removing all study-specific information, without distinguishing unwanted technical bias from relevant biological variation. Integration remains difficult when capturing biological variation that is distributed across studies, as when combining disparate temporal snapshots into a panoramic, multi-study trajectory of cellular development. Here, we show that a fundamental analytic shift to gene coexpression within clusters of cells, rather than gene expression within individual cells, balances robustness to bias with preservation of meaningful inter-study differences. We leverage this insight in Trajectorama, an algorithm which we use to unify trajectories of neuronal development and hematopoiesis across studies that each profile separate developmental stages, a highly challenging task for existing methods. Trajectorama also reveals systems-level processes relevant to disease pathogenesis within the microglial response to myelin injury. Trajectorama benefits from efficiency and scalability, processing nearly one million cells in around an hour.
BibTeX
@misc{hie2019coexpression,
  author = {Brian Hie and Hyunghoon Cho and Bryan D. Bryson and Bonnie Berger},
  title = {{Coexpression enables multi-study cellular trajectories of development and disease}},
  year = {2019},
  doi = {10.1101/719088},
  howpublished = {bioRxiv},
}
Figure