Spatial transcriptomics is a powerful technology that allows researchers to measure gene activity while preserving information about where cells are located within a tissue. Unlike traditional RNA sequencing, which can lose this spatial information when tissue is broken apart, spatial transcriptomics helps researchers see how different cells and gene expression patterns are organized across structures such as the brain or a tumor.

However, spatial transcriptomics data can contain substantial technical noise. Some genes may appear to have little or no expression simply because their RNA was not captured during the experiment. These missing signals, often called dropout events, can make it difficult to distinguish real biological patterns from technical variation.

Researchers from the Department of Computer Science at the University of Central Florida developed a computational framework called SPIDER to address this problem. SPIDER uses information from independently generated single-cell RNA sequencing, or scRNA-seq, data to help clean up spatial transcriptomics datasets.

Overview of SPIDER framework

SPIDER integrates ST data with reference scRNA-seq information by jointly leveraging spatial structure, transcriptional similarity, and pseudo-ST supervision. The framework first constructs three graphs: a spatial neighborhood graph derived from spot coordinates, and two transcriptional similarity graphs (TSGs) built from real-ST and pseudo-ST generated from scRNA-seq data from the same tissue type. A tri-branch graph attention encoder then learns low-dimensional latent representations from these graphs. The pseudo-ST branch incorporates supervision by optimizing cell-type ratio prediction, while an embedding regularization term aligns the real-ST and pseudo-ST embeddings to transfer cell-type structure from the reference domain. A unified graph attention decoder reconstructs gene expression using a zero-inflated negative binomial likelihood. The resulting denoised expression matrix can be used for downstream analyses, including spatial domain identification, marker gene pattern recovery, and improved clustering performance.

The key idea behind SPIDER is to use the detailed cell-type information available from single-cell RNA sequencing as a guide. Single-cell RNA sequencing measures gene expression in individual cells, allowing researchers to identify different cell populations and characterize their gene expression patterns.

SPIDER uses annotated scRNA-seq reference data from the same type of tissue as the spatial transcriptomics dataset. It then combines individual cells from the single-cell dataset to create simulated spatial transcriptomics measurements known as pseudo-ST spots. Because the cell types contributing to these artificial spots are already known, they provide the model with biological information that is normally missing from spatial transcriptomics data.

The program analyzes the data using three different networks, or graphs. One represents the physical proximity of locations within the tissue, another represents similarities in gene expression between locations in the real spatial transcriptomics data, and the third represents gene expression relationships in the simulated data created from single-cell RNA sequencing.

SPIDER then uses graph-based artificial intelligence methods to combine this information. The system learns patterns from the annotated single-cell reference data and transfers those patterns to the real spatial transcriptomics dataset. The goal is to recover biological signals that may have been weakened or lost because of technical noise.

An important feature of SPIDER is that it produces a new, denoised gene expression matrix. Researchers can therefore examine individual genes directly rather than relying only on abstract computational representations of the data. This helps preserve biological interpretability and makes it easier to connect computational results with specific genes and cellular processes.

The researchers tested SPIDER using spatial transcriptomics datasets from the human dorsolateral prefrontal cortex, a region of the brain involved in functions such as decision-making and working memory, as well as datasets from breast cancer tissue.

Across these datasets, SPIDER improved the spatial patterns of gene expression and helped recover known tissue structures. It also produced better clustering results than the other computational approaches evaluated. Clustering is commonly used to identify regions or groups of cells that share similar patterns of gene activity.

Marker genes also showed more continuous and clearly organized spatial expression after the data were processed with SPIDER. These improvements can make it easier for researchers to identify anatomical regions, cellular populations, and biological structures within complex tissues.

The approach is particularly interesting because it connects two complementary technologies. Single-cell RNA sequencing provides detailed information about individual cell types, while spatial transcriptomics shows where gene expression occurs within intact tissue. SPIDER uses information from one technology to improve the quality of the other.

There are limitations. SPIDER depends on having a high-quality single-cell RNA sequencing reference from the same or a closely related tissue. Differences between the reference dataset and the spatial transcriptomics experiment could influence the results. The researchers also report that SPIDER requires more computational time than some competing approaches.

Even so, SPIDER demonstrates how combining single-cell and spatial gene expression data can help researchers obtain clearer biological information from noisy datasets. As spatial transcriptomics becomes increasingly common in areas such as neuroscience, cancer research, and tissue biology, approaches that improve data quality while maintaining gene-level interpretability could become valuable components of spatial transcriptomics analysis.

Availability – The source code and dataset is available at: (https://github.com/compbiolabucf/SPIDER)