TL;DR
This paper introduces FAST, a novel reference-free method for deconvoluting spatial transcriptomics data. FAST uniquely integrates gene expression, spatial coordinates, and histology information using regularized non-negative matrix factorization (NMF) to provide more accurate and interpretable cell type composition of tissues, even when single-cell reference data is unavailable.
What Problem Does This Paper Address?
Spatial transcriptomics allows us to map gene expression within the physical context of a tissue. However, current technologies often provide "bulk" gene expression data from relatively large sequencing spots (e.g., 55-100 μm). To understand the cellular makeup at a finer resolution, a process called "deconvolution" is needed. Deconvolution aims to determine the proportion of different cell types and their corresponding gene expression profiles within each spot.
A major challenge in deconvolution is the reliance on "reference data," typically single-cell RNA sequencing (scRNA-seq) data. This reference data might not always be available or representative of the specific tissue being studied, making deconvolution difficult. Existing "reference-free" methods often have limitations: they might make strong assumptions about data distributions, depend heavily on pre-identified marker genes, or fail to leverage crucial spatial and histological information. This creates a need for flexible, robust, and user-friendly reference-free methods that can effectively use all available data.
Methods
The core of FAST is a regularized Non-negative Matrix Factorization (NMF) approach. NMF is a matrix decomposition technique that breaks down a large matrix (gene expression data) into two smaller matrices representing "gene signatures" (what genes are characteristic of a cell type) and "cell proportions" (how much of each cell type is in each spot).
FAST enhances standard NMF by introducing two key regularization terms:
- Spatial and Histology Integration: A graph Laplacian matrix is incorporated. This matrix captures the relationships between spots based on their spatial proximity and, crucially, their histological similarity. This allows the model to learn from the tissue's structure, assuming that spatially and histologically similar spots likely share similar cell compositions.
- Cell Proportion Constraint: A penalty term is added to ensure that the estimated proportions of cell types within each spot sum up to one. This enforces a biologically realistic constraint.
The method is flexible, allowing users to define the adjacency matrix (which defines "neighboring" spots) in various ways, including using spatial coordinates and histology image intensities. The authors propose a specific method for constructing this matrix that balances physical distance with histological information.
Datasets
The paper evaluates FAST using:
- Simulated Data: A dataset was generated from mouse nervous system scRNA-seq data, simulating spatial transcriptomics data with known cell type proportions for evaluation. This dataset involved 18,215 genes and two cell types (Astrocytes and Neurons) across 260 spots.
- Real Data:
- Mouse Olfactory Bulb (MOB) data (10X Visium, 260 spots, 16,218 genes).
- Human Breast Cancer data (likely 10X Visium, details on resolution and gene count not explicitly stated for this specific dataset, but mentioned as being from a different platform).
- Mouse Brain data (10X Visium, higher resolution with 2,702 spots and 32,285 genes).
Key Insights and Ideas
- Histology is More Than Just Pretty Pictures: FAST demonstrates that histological images, often used for visual inspection, can be quantitatively integrated into computational models to improve deconvolution accuracy. This moves beyond simple spatial proximity to capture more nuanced biological relationships between tissue regions.
- Reference-Free Power: The ability to perform accurate deconvolution without relying on scRNA-seq references is a significant step. This democratizes spatial transcriptomics analysis, making it accessible to labs that may not have the resources or expertise to generate comprehensive scRNA-seq datasets.
- Unified Framework: Combining gene expression, spatial coordinates, and histology into a single NMF model is a powerful integration strategy. This holistic approach allows for a more comprehensive understanding of tissue composition than methods that focus on only one or two of these data types.
- Interpretable Factors: The NMF framework inherently produces interpretable "gene signature" and "cell proportion" matrices, making it easier to identify marker genes and understand the biological meaning of the deconvoluted cell types.
Key Results
- Simulation Performance: FAST significantly outperformed two existing reference-free methods (STdeconvolve and CARD-free) in simulation studies. It achieved a higher average Pearson correlation (0.93) and lower RMSE (0.15) for estimating cell proportions, with improved consistency and stability.
- Real Data Applications:
- Mouse Olfactory Bulb: FAST successfully recovered tissue structures and identified distinct cell types corresponding to known MOB layers. It also identified marker genes, such as Kctd12 for the olfactory nerve layer.
- Breast Cancer: FAST helped distinguish between different breast cancer regions (DCIS and IDC) and identified cell types that were consistently abundant in both, providing insights into cancer heterogeneity. It also revealed pathways enriched in these cell types, aligning with existing literature.
- Mouse Brain (High Resolution): FAST was applied to higher-resolution data, successfully mapping cell type proportions to specific brain regions like the hypothalamus and isocortex, demonstrating its scalability.
Limitations and Future Work
- Adjacent Matrix Construction: While FAST offers flexibility in defining the adjacent matrix, the authors acknowledge that improving this component with more advanced techniques, such as deep learning-based feature extraction for histology, could further enhance performance.
- NMF Objective Function: The current implementation uses the Frobenius norm. The authors suggest exploring alternative norms, like Kullback-Leibler divergence, to potentially improve the model's performance.
- Generalizability: FAST is presented as a generalizable method that can be applied to other deconvolution problems beyond spatial transcriptomics, with minor modifications.
The authors envision future work focusing on refining the adjacent matrix construction, exploring alternative mathematical formulations for NMF, and extending FAST to a broader range of deconvolution tasks.
Summary updated