SpaConTDS: A multimodal contrastive learning framework for identifying spatial domains by applying tuple disturbing...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Multimodal spatial transcriptomics (ST) data is crucial for understanding cellular structure and function.
• Accurate identification of spatial domains is essential for ST data analysis.
• Existing methods face challenges in integrating diverse ST data and correcting batch effects.

Purpose of the Study:

• To develop a novel framework, SpaConTDS, for enhanced multimodal spatial transcriptomics data analysis.
• To improve the accuracy of spatial domain identification and downstream task performance.
• To enable robust integration of multiple tissue sections and batch effect correction without prior alignment.

Main Methods:

• SpaConTDS integrates reinforcement learning with self-supervised multimodal contrastive learning.
• It employs data augmentation and pseudo-label tuple perturbation for sample generation.
• Reinforcement learning dynamically optimizes model hyper-parameters.

Main Results:

• SpaConTDS achieves state-of-the-art accuracy in spatial domain identification.
• It outperforms existing methods in denoising, trajectory inference, and UMAP visualization.
• The framework effectively integrates multiple tissue sections and corrects batch effects without alignment.

Conclusions:

• SpaConTDS provides robust fused representations of multimodal ST data.
• It offers a flexible and powerful tool for various spatial transcriptomics analyses.
• The framework enhances the investigation of cellular structure and function through improved data integration and analysis.