Abstract: ProFAST is a probabilistic factor analysis framework designed for spatially-aware dimension reduction in large-scale spatial transcriptomics (ST) data. By incorporating spatial dependencies via a Conditional Autoregressive (CAR) model, ProFAST enables denoising, batch effect removal, and cell-gene co-embedding across millions of spatial locations. More details can be referred to Wei Liu, et al. (2023) doi:10.1101/2023.07.11.548486.
Introduction
Spatial transcriptomics (ST) provides a powerful lens into tissue architecture by measuring gene expression while preserving spatial coordinates. However, ST data is characterized by high dimensionality and significant technical noise. Traditional dimension reduction methods like PCA or UMAP often treat each spatial "spot" as an independent observation, failing to account for the spatial autocorrelation inherent in biological tissues—the principle that neighboring cells are more likely to be in similar states. ProFAST addresses these challenges by providing a scalable, probabilistic framework for spatially-aware dimension reduction.
Methods: Spatially-Aware Factor Analysis
The core of ProFAST is a probabilistic factor analysis model that leverages an intrinsic Conditional Autoregressive (CAR) model to incorporate spatial context. The workflow begins with the construction of an adjacency matrix, which defines the spatial relationships between spots.
The AddAdj function allows users to specify neighborhoods based on Euclidean distance or K-nearest neighbors, accommodating various ST platforms such as Visium or ST. Following this, the FAST function implements a Variational ICM-EM algorithm to extract low-dimensional embeddings. By treating spatial proximity as a prior, ProFAST effectively denoises the data while preserving biological signal across millions of spatial locations.
Cell-Feature Co-embedding
A distinctive feature of ProFAST is the Non-negative Cell-Feature Co-embedding (NCFM). This approach projects both cells and genes into a shared low-dimensional space. This geometric alignment simplifies the identification of signature genes; instead of relying solely on statistical differential expression, researchers can identify marker genes by their proximity to specific cell clusters in the embedding space.
Key Results and Visualization
ProFAST is designed for high-performance integration and visualization. It can align multiple tissue sections, effectively removing batch effects while maintaining spatial structures across the entire dataset. The package provides specialized tools for interpreting these complex embeddings.
The coembed_plot function generates 2D scatter plots that visualize the relationship between cell groups and their signature features. This allows for an intuitive exploration of the genes driving specific spatial domains. Additionally, diagnostic tools like diagnostic.cor.eigs assist researchers in determining the optimal number of latent factors ($q$), a critical step in transcriptomic analysis.
Conclusion and Limitations
ProFAST represents a significant advancement in the scalability of spatially-aware dimension reduction. While it excels at handling massive datasets, its performance is sensitive to the initial definition of the spatial adjacency matrix. Choosing an inappropriate neighborhood size could lead to over-smoothing or the loss of fine-grained spatial transitions. Future work may focus on automated neighborhood selection to further streamline the analysis of complex tissue architectures.
Summary updated