SOmics: A Machine Learning-Driven Spatial Transcriptomics Tool for Mapping Immune Suppression in Ovarian Cancer

High-grade serous carcinoma (HGSC) and ovarian carcinosarcoma (OCS) are aggressive malignancies characterized by poor clinical outcomes and limited responses to immunotherapy. Bulk RNA sequencing captures tumor-wide expression averages but does not preserve spatial information, while routine immunohistochemistry is limited in marker multiplexing. Here, we present SOmics, a machine learning-driven spatial transcriptomics platform that integrates bulk-trained classifiers with spatial gene expression, immune module scoring × and cell type deconvolution. Random Forest and L1-regularized logistic regression models were trained on TCGA ovarian cancer data to differentiate cancer-associated fibroblast (CAF)-high and immune-high transcriptional states, achieving held-out test AUC values of 0.968 (95% CI: 0.898–1.000) and 0.937 (95% CI: 0.848–0.997). Without retraining or feature reselection, these models were applied directly to 10x Genomics Visium spatial transcriptomic data (n = 8) to map CAF- and immune-associated transcriptional patterns within 55 µm multicellular spots. SOmics showed that OCS tumors had an immune silent microenvironment characterized by highly localized stromal- and myeloid-associated suppressive niches. Conversely, HGSC tumors exhibited distinct spatial compartmentalization with localized regions of antigen presentation together with regulatory T cell (Treg) and exhaustion signals. Spatial mapping of candidate marker genes revealed localized expression of metabolic checkpoints (NT5E/CD73, ENTPD1/CD39), exhaustion markers (LAG3, HAVCR2) and extracellular matrix genes (COL1A1, ACTA2). Within tissue spot-level resampling and exploratory tissue-level sensitivity analyses supported the internal stability of the observed CAF-associated differences between OCS and HGSC, although the limited number of original tissue samples requires cautious interpretation and further validation in independent cohorts. These findings indicate that immune suppression is organized differently on OCS and HGSC and that bulk expression states can be mapped to specific regions of the tumor microenvironment. SOmics provides a reproducible framework for identifying candidate suppressive regions and spatial marker genes that may inform future studies of immune resistance, biomarker development and combination treatment strategies in ovarian cancer. The open-access SOmics web interface is publicly available.