FAST: Scalable Factor Analysis for Spatial Dimension Reduction of Multi-section Spatial Transcriptomics

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Spatially resolved transcriptomics (SRT) techniques are rapidly advancing, generating large-scale datasets.
• Existing spatial dimension reduction methods struggle with the throughput and scale of modern SRT data.
• Efficient and scalable methods are needed to analyze multi-section SRT data while preserving biological and spatial information.

Purpose of the Study:

• To develop a fast and scalable method for spatial dimension reduction of large-scale SRT data.
• To create a generalized probabilistic factor analysis model that accounts for the count-based nature of SRT data.
• To enable the analysis of multiple tissue sections simultaneously while preserving spatial smoothness.

Main Methods:

• Developed FAST (Fast and Efficient Generalized Probabilistic Factor Analysis), a spatially aware dimension reduction model.
• FAST models count data across multiple sections and incorporates local spatial dependencies.
• Employs scalable computational complexity for handling large datasets.

Main Results:

• FAST embeddings showed improved correlation with annotated cell and domain types in simulated and real datasets.
• FAST successfully analyzed a large mouse embryo Stereo-seq dataset (>2.3 million locations) in 2 hours.
• Identified differential immune transcription factor activity and predicted CCNH as a carcinogenesis factor in a breast cancer Xenium dataset.

Conclusions:

• FAST is an efficient and scalable method for spatial dimension reduction of large-scale SRT data.
• The model accurately preserves biological signals and spatial information across multiple tissue sections.
• FAST enables novel biological discoveries by analyzing complex datasets and identifying regulatory relationships.