FIST

FIST imputes missing mRNA expression in spatial-transcriptomics RNA sequencing (sptRNA-seq) data by applying graph-regularized tensor completion to recover spatially resolved transcriptomes.

Key Features:

• Graph-regularized tensor completion (CPD): Models imputation as a tensor completion problem in Canonical Polyadic Decomposition (CPD) form.
• 3-way sparse tensor representation: Represents sptRNA-seq data as a 3-way sparse tensor with modes for genes (p-mode) and spatial coordinates (x-mode and y-mode).
• Integration of protein-protein interaction and spatial graphs: Incorporates a protein-protein interaction network and a spatial graph of capture spots to inform imputation.
• Cartesian product graph regularization: Regularizes the tensor completion with a Cartesian product graph combining the PPI network and the spatial graph to capture high-order relations.
• Imputation for high dropout sptRNA-seq: Addresses high dropout rates caused by in-situ capture and amplification failures by imputing missing gene expressions.
• Experimental validation: Evaluated by cross-validation on ten 10x Genomics Visium spatial transcriptomic datasets and compared to leading single-cell RNA-seq imputation methods.
• Tissue-level spatial analysis: Enables analysis of spatial characteristics and functions, demonstrated in a mouse kidney case study.

Scientific Applications:

• Spatial gene expression imputation: Recover missing mRNA expression values in sptRNA-seq datasets.
• Spatial pattern discovery: Reveal tissue-specific spatial organization of gene expression, as shown for mouse kidney.
• Functional inference with network priors: Improve interpretation of gene functions by integrating protein-protein interaction networks.
• Benchmarking imputation methods: Provide comparative performance assessment against single-cell RNA-seq imputation approaches using 10x Genomics Visium data.

Methodology:

Models sptRNA-seq data as a 3-way sparse tensor (genes p-mode; x-mode and y-mode spatial coordinates), performs tensor completion in Canonical Polyadic Decomposition (CPD) form, and regularizes the solution with a Cartesian product graph combining a protein-protein interaction network and a spatial graph, with performance assessed by cross-validation on ten 10x Genomics Visium datasets.