Efficient integration of multiple spatial transcriptomics data for 3D domain detection, matching, and alignment with stMSA

Spatial transcriptomics (ST) is a powerful methodology that enables the study of genes within tissue architecture by providing gene expression information along with spatial location data. With the increasing availability of ST datasets, researchers are now inclined to explore potential biological features across larger datasets simultaneously, aiming for a more comprehensive understanding. However, existing methods primarily focus on cross-batch feature learning, often overlooking the intricate spatial patterns within individual slices. This limitation poses a significant challenge in effectively integrating features across different slices while considering slice-specific patterns. To address this challenge and enhance the integration performance of multi-slice data, we propose stMSA, a deep graph contrastive-learning model that incorporates graph auto-encoder techniques. stMSA is specifically designed to generate batch-corrected representations while preserving the unique spatial patterns within each slice, simultaneously considering both inner-batch and cross-batch patterns during the integration process. Our extensive evaluations demonstrate that stMSA outperforms state-of-the-art methods in discerning tissue structures across different slices, even when confronted with diverse experimental protocols and sequencing technologies. Furthermore, stMSA exhibits remarkable performance in cross-slice matching and alignment for three-dimensional reconstruction. The source code for stMSA is available at https://github.com/hannshu/stMSA .