BatchBlend: A Batch-Aware Interpretable Probabilistic Graphical Model for Integrating Spatially Resolved Transcriptomes

While multi-sample spatially-resolved transcriptomics (mSRT) technologies have enhanced our ability to compare different tissues across conditions and time points, they often introduce technical artifacts or batch effects that confound biological interpretability. Current batch effect correction methods either sacrifice interpretability for performance or fail to adequately leverage spatial context. BatchBlend addresses this gap by extending the interpretable framework of Popari with an explicit batch effect term, enabling the separation of technical variation from biological signal while maintaining an interpretable framework. The model is formulated as a non-negative matrix factorization with hidden Markov random fields (NMF-HMRF) similar to Popari with an added sample-specific batch effect term. Through comprehensive evaluation on simulated mSRT data, BatchBlend demonstrates competitive batch effect correction while maintaining moderate biological signal preservation compared to established methods: PyLiger, scVI, and STAligner. In addition, BatchBlend's performance in batch effect correction on mouse brain coronal slices was analyzed when integrating data from MERFISH, MERSCOPE, and STARmap. While BatchBlend offers an interpretable approach to mSRT batch correction, further methodological refinements are needed to fully leverage spatial context during integration, presenting promising directions for future research to enhance both integration quality and biological signal preservation in spatially resolved transcriptomics.