An Interpretable Multi-instance Learner Decodes Cellular Recruitment from Spatially Resolved Transcriptomics.

The recruitment of the various types of cells into the tissue microenvironment and how these cells engage with other cells in the tissue sites play critical biological roles. However, it is difficult to study these processes on a genome-wide scale using traditional low-throughput experiments. To address this gap, we developed a fully interpretable multi-instance deep learner, coined "spacer". Spacer digests the transcriptomics and spatial modalities of spatially resolved transcriptomics (SRT) data in alignment with the biological principles of the tissue spatial architecture. We deployed spacer to a panel of 17 high definition and 20 low definition SRT datasets, for studying how stromal and immune cells were recruited into tumors and heart during myocarditis. Spacer discovered novel biological insights not affordable by prior spatial data analysis tools, which were validated by orthogonal immuno-peptidomics, spatial T cell receptor (TCR) sequencing, and single cell sequencing experiments. We discovered genes that encode more immunogenic peptides and that are involved in developmental pathways are more potent in recruiting T cells to local tumor sites, while the recruitment of B cells and macrophages is stipulated by a different molecular program. On the other hand, expression of mucins in the tumor cells was found to repel T cell localization. For the engaging cell side, we also uncovered T cell-intrinsic features that determine their localization, validated by spatial-TCR-seq data. Spacer also unexpectedly revealed that CD4