Integration of Mouse and Human Single-cell RNA Sequencing Infers Spatial Cell-type Composition in Human Brains

Travis S. Johnson, Zachary B. Abrams, Bryan R. Helm, Peter Neidecker, Raghu Machiraju, Yan Zhang, Kun Huang, Jie ZhangbioRxiv 2019 · DOI

TL;DR

This paper tackles the challenge of mapping cell types in the human brain, which is limited by a scarcity of donor data. By integrating mouse and human single-cell RNA sequencing (scRNA-seq) data, the researchers developed a method to accurately predict and spatially map cell-type composition in human brain tissue. This approach not only overcomes data limitations but also reveals how neural connectivity influences gene expression patterns, offering a more nuanced understanding of brain tissue heterogeneity.

What Problem Does This Paper Address?

Understanding the intricate cellular makeup of the human brain is crucial for deciphering its function, development, and role in disease. While single-cell RNA sequencing (scRNA-seq) has revolutionized cell type identification, its application to human brain tissue is hampered by a limited number of available brain donors. This scarcity makes it difficult to create comprehensive maps of cell-type distribution and understand how this composition varies across different brain regions. The authors aim to bridge this gap by developing a computational approach that leverages existing data, including that from model organisms like mice, to infer human brain cell-type composition and its spatial organization.

Methods

The core of this study lies in an integrative deconvolution approach. The researchers combined scRNA-seq data from both mouse and human brains to create "in silico chimeric" datasets. This allowed them to "fill in" missing cell-type information for human samples. Key steps included:

  • Feature Selection: Identifying genes that are concordant between mouse and human datasets to ensure comparability.
  • Deconvolution: Using Ordinary Least Squares (OLS) regression (chosen over Non-Negative Matrix Factorization Regression for better accuracy) to estimate the proportion of different cell types within bulk tissue samples (from the Allen Human Brain Atlas).
  • Integration of Mouse and Human Data: Appending mouse cell types to human datasets where human data was incomplete, enabling a more comprehensive cell-type signature generation.
  • Incorporating Neural Connectivity: Analyzing how neural projection volumes (derived from mouse brain connectivity data) influence the relationship between neuron nuclei location and mRNA expression, suggesting that gene expression is a more robust indicator of cell location for complex neuronal structures.

Datasets

The study utilized the following datasets:

  • Mouse scRNA-Seq: One dataset (MusNG) from the mouse hippocampus and cortex, serving as a comprehensive source for cell-type signatures.
  • Human scRNA-Seq: Two datasets (HumN from cortex, HumNG from temporal lobe) providing human cell-type profiles.
  • Allen Human Brain Atlas (AHBA): Six human microarray datasets comprising 3,702 samples from various brain regions, used for spatial mapping and deconvolution.
  • Mouse Brain Connectivity Dataset: From the Allen Brain Institute, used to investigate the role of neural connectivity.

Key Insights and Ideas

  • Mouse-Human Data Integration is Powerful: The study demonstrates that combining scRNA-seq data from different species can effectively overcome limitations in human data availability, enabling robust cell-type deconvolution.
  • Gene Expression Over Nuclei Location for Neurons: For large, irregularly shaped cells like neurons, gene expression patterns are a more reliable indicator of cellular location than nuclei position alone, especially when considering neural connectivity.
  • Neural Connectivity Shapes Spatial Gene Expression: The spatial distribution of mRNA in neurons is not solely determined by nuclei location but is significantly influenced by axonal projection volumes, suggesting a dynamic interplay between structure and gene expression.
  • Consistent Spatial Cell-Type Mapping: The developed method achieves high consistency in cell-type mapping across different scRNA-seq datasets and human brain donors, validating its reliability.

Key Results

  • The deconvolution method achieved high accuracy in predicting cell-type proportions, with correlations between true and estimated proportions often exceeding 0.9.
  • 91% of cell-type correlations between different scRNA-seq datasets were significantly positively correlated, indicating robustness.
  • The spatial mapping of cell types across AHBA donors showed high consistency, with an average pairwise Pearson's correlation coefficient (PCC) of 0.87.
  • The study successfully mapped known cell types to their expected anatomic locations, regardless of the species of the scRNA-seq data used.
  • The correlation between neuron nuclei location and mRNA location was significantly improved (PCC = 0.74) only after accounting for neural connectivity, highlighting its importance.

Limitations and Future Work

  • Data Sparsity: scRNA-seq data is inherently sparse, which requires careful filtering and robust analytical methods.
  • Species-Specific Differences: While integration is effective, subtle species-specific differences in gene expression might still introduce biases.
  • Future Directions: The authors advocate for expanding these techniques to include more diverse data types (e.g., histological images) and developing more sophisticated transfer-learning approaches for even more accurate models of tissue heterogeneity. They also suggest exploring machine learning approaches and feature reduction beyond simple feature selection for deconvolution.

Summary updated

Abstract

Technical advances have enabled the identification of high-resolution cell types within tissues based on single-cell transcriptomics. However, such analyses are restricted in human brain tissue due to the limited number of brain donors. In this study, we integrate mouse and human data to predict cell-type proportions in human brain tissue, and spatially map the resulting cellular composition. By applying feature selection and linear modeling, combinations of human and mouse brain single-cell transcriptomics profiles can be integrated to "fill in" missing information. These combined "in silico chimeric" datasets are used to model the composition of nine cell types in 3,702 human brain samples in six Allen Human Brain Atlas (AHBA) donors. Cell types were spatially consistent regardless of the scRNA-Seq dataset (91% significantly correlated) or AHBA donor (p-value = 4.43E-20 by t-test) used in the model. Importantly, neuron nuclei location and neuron mRNA location were correlated only after accounting for neural connectivity (p-value = 1.26E-10), which supports the notion that gene expression is a better indicator than nuclei location of cellular localization for cells with large and irregularly shaped cell bodies, such as neurons. These results advocate for the integration of mouse and human data in models of brain tissue heterogeneity.
BibTeX
@misc{johnson2019integration,
  author = {Travis S. Johnson and Zachary B. Abrams and Bryan R. Helm and Peter Neidecker and Raghu Machiraju and Yan Zhang and Kun Huang and Jie Zhang},
  title = {{Integration of Mouse and Human Single-cell RNA Sequencing Infers Spatial Cell-type Composition in Human Brains}},
  year = {2019},
  doi = {10.1101/527499},
  howpublished = {bioRxiv},
}
Figure