This study focuses on the analysis of spatial transcriptomics data generated using the 10X Genomics Visium CytAssist platform from formalin-fixed paraffin-embedded (FFPE) coronary artery tissue sections. A total of six human samples were analyzed, comprising three from autopsies and three from freshly explanted hearts. The raw spatial gene expression data were retrieved from the NCBI GEO database (Accession Number: GSE283269).
The primary objective of this study is to investigate intercellular
communication by exploring ligand-receptor interactions among spatial
clusters of gene expression spots. To achieve this, we employed a
computational pipeline integrating widely used Python-based tools,
including scanpy, scvi-tools, and
squidpy.
The workflow begins by importing raw gene expression counts and spot
metadata into an AnnData object. After initial quality
control (QC) steps—such as removing low-quality spots and filtering for
in-tissue regions—we selected highly variable genes (HVGs) from
log-normalized data to prepare for advanced modeling. Detailed report
To uncover latent structure in the spatial gene expression data, we
applied the SCVI model from scvi-tools, which
provides a probabilistic framework that accounts for technical noise and
batch effects. The model was trained using raw counts, and its
performance was monitored via the evidence lower bound (ELBO) loss. The
learned latent embeddings were used to construct a neighborhood graph
for Leiden clustering, and clusters were visualized using both UMAP and
spatial tissue projections.
Differential gene expression (DGE) analysis was subsequently
performed using the trained SCVI model to identify
cluster-specific marker genes. These markers were visualized spatially
to reveal the molecular characteristics and anatomical contexts of
distinct cell populations. The final outputs—including the processed
AnnData object and trained model—were saved for future use.
Samples were processed individually: Autopsy1, Autopsy2 (bad
quality), Autopsy3, Explanted1, Explanted2 and Explanted3.
To investigate cell-cell communication, we employed ligand-receptor
(LR) interaction analysis using Squidpy. The core function,
squidpy.gr.ligrec, estimates communication strength between
clusters based on a curated database of known human LR pairs (default:
OmniPath). Mean expression values of ligands (in source clusters) and
receptors (in target clusters) were computed using the raw count layer
(adata.layers['counts']), with clusters defined via Leiden
clustering of the SCVI latent space.
Interaction scores were calculated for each cluster pair as the
product of mean ligand and receptor expression. To assess statistical
significance, a permutation-based test was performed
(n_perms), in which cluster labels were randomly shuffled
to generate a null distribution of interaction scores. P-values were
computed by comparing observed scores against this null, and
interactions with p-values below a defined threshold
(alpha) were deemed significant.
The results—stored in adata.uns['ligrec']—included
matrices of interaction scores and p-values. Downstream visualizations,
generated via squidpy.pl.ligrec, included dot plots,
interaction heatmaps, and Circos plots (via pycircos) to
map the overall communication landscape. Further analysis identified key
ligand-receptor pairs and dominant communication pathways among
clusters, providing insights into the molecular interactions within
coronary artery tissue microenvironments. Processed results shown
here.