SpaIM: Single-cell Spatial Transcriptomics Imputation via Style Transfer

Abstract

Spatial transcriptomics (ST) technologies have revolutionized our understanding of cellular ecosystems. However, these technologies face challenges such as sparse gene signals and limited gene detection capacities, which hinder their ability to fully capture comprehensive spatial gene expression profiles. To address these limitations, we propose leveraging single-cell RNA sequencing (scRNA-seq), which provides comprehensive gene expression data but lacks spatial context, to enrich ST profiles. Herein, we introduce SpaIM, an innovative style transfer learning model that utilizes scRNA-seq information to predict unmeasured gene expressions in ST data, thereby improving gene coverage and expressions. SpaIM segregates scRNA-seq and ST data into data-agnostic contents and data-specific styles, with the contents capture the commonalities between the two data types, while the styles highlight their unique differences. By integrating the strengths of scRNA-seq and ST, SpaIM overcomes data sparsity and limited gene coverage issues, making significant advancements over 12 existing methods. This improvement is demonstrated across 53 diverse ST datasets, spanning sequencing- and imaging-based spatial technologies in various tissue types. Additionally, SpaIM enhances downstream analyses, including the detection of ligand-receptor interactions, spatial domain characterization, and identification of differentially expressed genes. Released as open-source software, SpaIM increases accessibility for spatial transcriptomics analysis. In summary, SpaIM represents a pioneering approach to enrich spatial transcriptomics using scRNA-seq data, enabling precise gene expression imputation and advancing the field of spatial transcriptomics research.

Fig. 1. Overview of the SpaIM model.

SpaIM comprises the comprises an ST autoencoder and an ST generator. Both ST autoencoder and ST generator are built on the multilayer Recursive Style Transfer (ReST) layers.

Fig. 2. Performance evaluation in the breast cancer dataset.

a Comparison results between SpaIM and existing methods using Pearson correlation coefficient (PCC). b Comparison results between SpaIM and existing methods using structural similarity index measure (SSIM). c Spatial visualization of the ground truth and the predicted gene expressions. d Spatial visualization of spatial domains. e Scatter plot of associated ligand receptor pairs in the raw data and the SpaIM-imputed data.

Fig. 3. Benchmarking performance on the NanoString CosMx SMI dataset.

a Benchmarking results on the Lung9-rep1 dataset, using evaluation metrics including structural similarity index measure (SSIM) and Jaccard similarity (JS). b Spatial visualization of cell types in the Lung9-rep1 dataset. c Comparisons of the number of differentially expressed genes in each cell type. d Spatial visualization of ground truth and the predicted expressions of tumor-related genes.

Fig. 4. SpaIM facilities spatial domain detection.

a Benchmarking results on the NanoString CosMx spatial transcriptomics dataset (Lung5-rep3), using evaluation metrics including structural similarity index measure (SSIM) and Jaccard similarity (JS). b Spatial visualization of cell types in the whole slide. d Spatial visualization of cell types specific Field Of Views (FOVs).

Fig. 5. SpaIM reliably recovers unmeasured genes.

a UMAP visualizations of all cell populations and nontumor populations. b UMAP visualization of the imputation of unmeasured marker genes by SpaIM. c UMAP visualization of the imputation of unmeasured marker genes by Tangram.

Fig.. Gene imputation performance across diverse spatial transcriptomics (ST) datasets.

Comparison results across (a) 21Visium ST datasets; (b) 28 sequencing-base datasets; (c) 25 imaging-based ST datasets; (d) total 53 datasets profiled by diverse ST technologies, using ranked Pearson correlation coefficient (PCC), Jaccard similarity (JS), and Accuracy score.