Enterprise AI Analysis

On simulated tri-omics data with known ground truth, SMART demonstrated superior qualitative and quantitative performance compared to existing methods like MOFA+, MEFISTO, and SpatialGlue. It accurately identified five distinct spatial factors (cell types and background) with precise boundary delineation, outperforming methods that struggled with noise or over-integration. SMART preserved original omics relationships in the embedded space, achieving high Pearson correlation coefficients, especially for RNA and ADT. It also showed excellent spatial clustering coherence with Moran's I scores approaching 1, indicating orderly spatial distribution. These results validate SMART's theoretical superiority in spatial domain identification.

Applying SMART to human lymph node spatial transcriptomics and proteomics data, it successfully identified anatomical structures such as cortex, medullary sinuses, and follicles, outperforming other methods in delineation accuracy. On mouse brain MISAR-seq data (RNA and ATAC), SMART effectively reconstructed brain structures across developmental stages (E11.0 to E18.5), distinguishing regions like dorsal pallium and diencephalon with clear boundaries. For the P22 mouse brain CUT&Tag-RNA-seq dataset, SMART showed superior integration, identifying major anatomical layers with fewer noise artifacts. These real-world applications underscore SMART's robust ability to integrate diverse spatial multi-omics data for accurate biological insights.

SMART is an efficient model with fewer parameters and lower computational complexity, making it suitable for large-scale datasets. On the P22 mouse brain CUT&Tag-RNA-seq dataset (9752 spots, 25,881 genes, 70,470 peaks), SMART consistently demonstrated the shortest runtime and lowest memory consumption compared to methods like MEFISTO. For the Stereo-CITE-seq mouse spleen dataset, SMART ran smoothly even on the largest Bin10 dataset with over 750,000 spots in just 56 seconds, significantly outperforming all other methods that failed to complete computations on similar scales. This efficiency highlights SMART's broad potential for large-scale spatial data analysis across various resolutions (5 µm to 100 µm).

SMART-MS, an extension of SMART, enables multi-omics integration across multiple tissue sections, incorporating batch effect correction and cross-section spatial neighbor graphs. On a human tonsil dataset (RNA and ADT across three sections), SMART-MS accurately identified follicular and germinal center regions, as well as T cell-enriched interfollicular zones, with superior biological signal preservation and batch effect removal compared to MultiVI, Present-BC, and MOFA+. It achieved the highest overall ARI scores and Moran's I scores across sections. This capability provides a comprehensive molecular atlas beyond individual tissue sections, enhancing resolution and coverage for complex biological systems.