Trainable Clustering Framework for Spatial Transcriptomics

Abstract Spatial transcriptomics (ST) enables high-resolution exploration of tissue architecture by integrating gene expression profiles with spatial information, thereby advancing insights into cellular composition, organization, and interactions. Among ST applications, spatial domain identification is critical for linking gene expression patterns to tissue morphology and analyzing the tissue microenvironment. We introduce a trainable clustering framework that unifies four complementary strategies—ACT, FACT, Scatter, and Ensemble—into a cohesive architecture. By coupling autoencoder-driven feature learning with an Mclust-assisted clustering layer, this framework enables joint optimization of representation and cluster assignments through a trainable loss function. Applied to human DLPFC, mouse brain anterior, and human breast cancer datasets, the Proposed framework achieves competitive accuracy in most cases while reliably identifying spatial domains and preserving complex tissue architecture. Additionally, Stereo-seq and Slide-seq datasets are utilized to evaluate cross-platform generalizability and robustness. The proposed framework is implemented in Python. The codes and datasets are available in https://github.com/SababAosaf/FACTSpatialTranscriptomics.