Analysis of Multi-modal Spatial Omics with MISO
Kyle Coleman, MS: University of Pennsylvania, DBEI Jian Hu, PhD: Emory University, Department of Human Genetics David Zhang, PhD: University of Pennsylvania, DBEI Mingyao Li, PhD: University of Pennsylvania, DBEI
Poster # 29
Multi-modal spatial omics provides the opportunity to analyze multiple omics and imaging data modalities within a spatial context. A prominent goal in multi-modal spatial omics is the consolidation of features from different modalities into unified embeddings that can be used for downstream analyses. While some methods have been developed to integrate modalities from a limited subset of multi-modal spatial omics experiments (e.g., spatially resolved transcriptomics), they are not suitable for most technologies, can only take two modalities as input, and require a great deal of hyperparameter tuning. With the growing number of multi-modal spatial omics technologies, a user-friendly method that can seamlessly integrate the data generated from any of these experiments is desirable. Here we present MISO, a feature extraction and spatial clustering algorithm that can be applied to all modalities from any multi-modal spatial omics experiment. MISO first extracts low-dimensional embeddings for each modality using modality-specific multilayer perceptrons trained to minimize spectral clustering and reconstruction loss functions. MISO then constructs features representing the interactions between each pair of modalities by taking the outer product between the modality-specific embeddings. The modality-specific and interaction feature vectors are concatenated to form embeddings coherent with respect to all modalities. When evaluated on a diverse set of multi-modal spatial omics datasets, including spatially resolved transcriptomics (transcriptomics and H&E histology), spatial ATAC-RNA-seq (chromatin accessibility and transcriptomics), and spatial CITE-seq (transcriptomics, proteomics, and H&E histology), MISO is able to accurately integrate different modalities and separate spots into biologically meaningful spatial domains. Despite its wide range of applicability, MISO requires little hyperparameter tuning and is able to overcome artifacts inherent in low quality modalities. We anticipate that MISO's flexible framework will make it a preferred method among those seeking to uncover novel biological findings through the integration of multiple spatially resolved omics and imaging data modalities.
LIGHTNING TALK - 2023 MidAtlantic Bioinformatics Conference