Unsupervised pattern identification in spatial gene expression atlas reveals mouse brain regions beyond established ontology

Robert Cahill, Yu Wang, R. Patrick Xian, Alex J. Lee, Hongkui Zeng, Bin Yu, Bosiljka Tasic, Reza Abbasi-Asl

TL;DR

This paper introduces osNMF, a novel unsupervised framework that leverages stability-driven non-negative matrix factorization to identify biologically meaningful patterns (principal patterns or PPs) in 3D spatial gene expression data of the mouse brain. These data-driven PPs reveal a new brain ontology that complements, and in some cases extends, the established expert-annotated atlas, offering a more objective and comprehensive view of brain organization.

What Problem Does This Paper Address?

The explosion of spatial transcriptomics data, which captures gene expression within its native tissue context, presents a significant challenge: how to extract meaningful biological insights from these complex, high-dimensional datasets. Traditional methods often rely on expert-defined atlases, which can be time-consuming, labor-intensive, and potentially biased. Furthermore, existing computational tools for analyzing spatial gene expression data often lack a systematic approach to ensure the stability and interpretability of the discovered patterns, making it difficult to trust their biological relevance. This paper aims to bridge this gap by developing a data-driven, unsupervised, and stable framework for discovering and interpreting spatial gene expression patterns in the mouse brain.

Methods

The core of the osNMF framework consists of three modules:

  1. Imputation Module: Utilizes a k-nearest neighbors (kNN) algorithm to estimate and fill in missing voxel data in the 3D spatial gene expression dataset. This is crucial for handling the inherent sparsity and missing values in such data.
  2. Stability-Driven Non-Negative Matrix Factorization (staNMF) Core: This is the novel component. It employs a stability-driven approach to Non-Negative Matrix Factorization (NMF) to decompose the gene expression data into a set of "principal patterns" (PPs). Unlike standard NMF, staNMF prioritizes the stability and reproducibility of these patterns across multiple runs and slight data perturbations. This ensures that the identified patterns are robust and not artifacts of the specific analysis run. The method uses Coordinate Descent for optimization.
  3. Ontology Identification Module: A search algorithm that systematically compares the discovered PPs to known anatomical regions from the Allen Common Coordinate Framework (CCF). It identifies combinations of expert-annotated regions that best correlate with each PP, allowing for the mapping of data-driven patterns to established brain atlases.

Datasets

The primary dataset used is the Allen Mouse Brain Atlas (ABA). This dataset comprises in situ hybridization (ISH) measurements for 4,345 genes at 200 µm isotropic resolution from adult mouse brains (56 days old). The Allen Mouse Brain Common Coordinate Framework (CCFv3) was used as the 3D reference atlas.

Key Insights and Ideas

  • Data-Driven Brain Ontology: The most significant insight is that purely data-driven PPs, derived solely from gene expression patterns, can reveal a brain organization that aligns with, yet also extends beyond, the established expert-annotated ontology. This suggests that gene expression itself holds a fundamental organizational principle for brain regions.
  • Stability is Key for Interpretability: The paper strongly emphasizes the importance of pattern stability for biological interpretation. By using a stability-driven approach, osNMF generates more reliable and reproducible patterns compared to methods like PCA.
  • PPs as Biologically Meaningful Units: The identified PPs are not just statistical artifacts; they represent spatially coherent and distinct regions of the mouse brain, often correlating with combinations of known anatomical areas. This suggests that these PPs capture fundamental biological units of brain organization.
  • Beyond Established Boundaries: osNMF can identify brain regions or patterns that don't neatly fit into existing hierarchical classifications, potentially uncovering novel functional or developmental units.

Key Results

  • 11 Stable PPs Identified: The osNMF framework identified 11 stable principal patterns (PPs) in the adult mouse brain dataset.
  • Superior Spatial Coherence: osNMF PPs demonstrated significantly higher spatial coherence (average Moran's I of 0.58 ± 0.12) compared to PCA PPs (0.47 ± 0.15), indicating more distinct and organized spatial expression patterns.
  • Improved Reconstruction Accuracy: osNMF achieved a 24% higher gene reconstruction accuracy (0.62 ± 0.22) compared to PCA (0.37 ± 0.37), meaning the PPs better capture the original gene expression data.
  • Strong Correlation with CCF: The identified PPs showed high correlations with combinations of expert-annotated CCF regions, with an average maximum correlation coefficient of 0.74 ± 0.04.
  • Novel Ontology Discovery: The PPs revealed new organizational principles, with some PPs correlating with combinations of CCF regions that differ from the established ontology. For instance, PPs 1-5 showed strong correlations with the cerebral cortex but didn't fit neatly into coarse or medium-level CCF regions.
  • Identification of Marker Genes and Networks: The framework successfully identified marker genes for each PP and constructed putative local gene interaction networks, providing leads for further experimental validation.

Limitations and Future Work

  • Resolution Limitation: The ABA dataset was downsampled to 200 µm isotropic resolution for computational efficiency. This relatively low resolution might miss finer details and could be a source of errors or biases.
  • Single Modality Data: The analysis was based solely on ISH gene expression data. Future work will integrate other modalities like MRI and axonal projections to provide a more precise characterization of brain regions.
  • Linear Interactions: The current gene interaction network construction relies on linear relationships. Future work will explore incorporating nonlinear interactions using supervised methods.
  • Validation Needed: While the identified marker genes and networks show promising alignment with known biology, they require experimental validation.
  • Broader Applicability: The authors express optimism that osNMF can be applied to new ST datasets and modalities, and suggest exploring stability-driven methods with deep neural networks (DNNs) and other advanced algorithms. The code for osNMF is publicly available.

Summary updated

Abstract

The rapid growth of large-scale spatial gene expression data demands efficient and reliable computational tools to extract major trends of gene expression in their native spatial context. Here, we used stability-driven unsupervised learning (i.e., staNMF) to identify principal patterns (PPs) of 3D gene expression profiles and understand spatial gene distribution and anatomical localization at the whole mouse brain level. Our subsequent spatial correlation analysis systematically compared the PPs to known anatomical regions and ontology from the Allen Mouse Brain Atlas using spatial neighborhoods. We demonstrate that our stable and spatially coherent PPs, whose linear combinations accurately approximate the spatial gene data, are highly correlated with combinations of expert-annotated brain regions. These PPs yield a brain ontology based purely on spatial gene expression. Our PP identification approach outperforms principal component analysis and typical clustering algorithms on the same task. Moreover, we show that the stable PPs reveal marked regional imbalance of brainwide genetic architecture, leading to region-specific marker genes and gene coexpression networks. Our findings highlight the advantages of stability-driven machine learning for plausible biological discovery from dense spatial gene expression data, streamlining tasks that are infeasible by conventional manual approaches.
BibTeX
@inproceedings{cahill2024unsupervised,
  author = {Robert Cahill and Yu Wang and R. Patrick Xian and Alex J. Lee and Hongkui Zeng and Bin Yu and Bosiljka Tasic and Reza Abbasi-Asl},
  title = {{Unsupervised pattern identification in spatial gene expression atlas reveals mouse brain regions beyond established ontology}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2024},
  doi = {10.1073/pnas.2319804121},
}

Loading related content...

Figure