Understanding how cells behave in the context of the tissue around them is key to figuring out how organs work, how diseases develop, and how cells interact with each other. Traditional single-cell RNA sequencing gives us a detailed picture of which genes are active in individual cells, but it does not tell us where those cells are located in a tissue. On the other hand, subcellular spatial transcriptomics captures where gene activity occurs, but it only measures a limited set of genes.

Researchers at Yale University have developed a new computational tool called VISTA to fill this gap. VISTA works by combining the broad gene coverage from single-cell RNA sequencing with the spatial detail from spatial transcriptomics, using advanced machine learning approaches to predict gene activity for genes that weren’t directly measured in spatial data. Essentially, VISTA can guess the missing expression information in spatial maps, making the data much more complete and useful.

Overview of VISTA

(a): The flowchart of VISTA. VISTA learns a joint distribution from scRNA-seq data and spatial data with graph-aware design. Here we use a Multi-Layer Perceptron (MLP) to encode information from the expression domain and a GNN to encode information from the spatial domain. The results after imputation can be utilized in several downstream analyses. (b): Explanation of uncertainty quantification. The uncertainty quantification of VISTA is based on sampling from the output distribution. For cell uncertainty quantification, we sample from celllevel distribution and compute the median of the cosine similarity between samples and mean gene expression. For gene uncertainty quantification, we sample from gene-level distribution and compute the median of the cosine similarity between samples and mean gene expression.

When tested on multiple datasets, VISTA not only predicted missing gene expression with high accuracy, but it also scaled well to large datasets and worked efficiently. These enhanced spatial expression maps allow researchers to do more powerful analyses, such as finding genes that vary across tissue regions, identifying how cells communicate using ligand-receptor pairs, inferring how gene expression changes over time in space, simulating how tissues might respond to changes, and separating spatial effects from cell-intrinsic behavior.

By linking comprehensive gene expression with precise spatial information, VISTA provides a richer picture of how cells are organized and behave within complex tissues. This makes it a promising tool for many biological and medical applications, from understanding development to unraveling disease mechanisms.

Availability – The codes of VISTA can be found in https://github.com/HelloWorldLTY/VISTA (also in Zendo)