From Data-Driven to Expert-Guided: Combining Unsupervised and Semi-supervised Clustering in Spatial Transcriptomics

One of the challenges in spatial transcriptomic experiments is identifying clusters of genes that exhibit similar expression patterns within specific regions of a tissue sample. The SpaRTaCo model, proposed by A. Sottosanti and D. Risso in 2023, offers a fully data-driven approach for the spatial classification of a tissue based on gene expression levels. Additionally, pathologist annotations of tissue samples are often available, albeit with significant variations between annotations and the data-driven analysis. In this work, we present a pivotal study focusing on a prostate cancer tissue sample. We demonstrate the integration of SpaRTaCo with two semi-supervised variants of the model, which incorporate external biological knowledge. This integration aims to uncover meaningful biological insights and specific gene expression patterns that may not be apparent through solely one of the two approaches.