Trainable clustering framework for spatial transcriptomics

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Spatial transcriptomics (ST) provides high-resolution tissue architecture insights by integrating gene expression with spatial data.
• Spatial domain identification is crucial for linking gene expression to tissue morphology and microenvironment analysis.

Purpose of the Study:

• To introduce a trainable clustering framework for unified spatial domain identification.
• To optimize feature learning and cluster assignments for improved spatial transcriptomics analysis.

Main Methods:

• A cohesive architecture unifying four strategies: ACT, FACT, Scatter, and Ensemble.
• Coupling autoencoder-driven feature learning with an Mclust-assisted clustering layer.
• Utilizing a trainable loss function for joint optimization of representation and cluster assignments.

Main Results:

• The framework achieves competitive accuracy across human DLPFC, mouse brain, and breast cancer datasets.
• It reliably identifies spatial domains while preserving complex tissue architecture.
• Cross-platform generalizability and robustness were evaluated using Stereo-seq and Slide-seq data.

Conclusions:

• The proposed framework offers a robust and accurate method for spatial domain identification in spatial transcriptomics.
• It advances the analysis of tissue microenvironments and cellular interactions.
• The open-source implementation facilitates broader application in biological research.