Refining molecular subtypes and risk stratification of ovarian cancer through multi‐omics consensus portfolio and machine learning

Abstract Ovarian cancer (OC), known for its pronounced heterogeneity, has long evaded a unified classification system despite extensive research efforts. This study integrated five distinct multi‐omics datasets from eight multicentric cohorts, applying a combination of ten clustering algorithms and ninety‐nine machine learning models. This methodology has enabled us to refine the molecular subtyping of OC, leading to the development of a novel Consensus Machine Learning‐driven Signature (CMLS). Our analysis delineated two prognostically significant cancer subtypes (CS), each marked by unique genetic and immunological signatures. Notably, CS1 is associated with an adverse prognosis. Leveraging a subtype classifier, we identified five key genes (CTHRC1, SPEF1, SCGB3A1, FOXJ1, and C1orf194) instrumental in constructing the CMLS. Patients classified within the high CMLS group exhibited a poorer prognosis and were characterized by a “cold tumor” phenotype, indicative of an immunosuppressive microenvironment rich in MDSCs, CAFs, and Tregs. Intriguingly, this group also presented higher levels of tumor mutation burden (TMB) and tumor neoantigen burden (TNB), factors that correlated with a more favorable response to immunotherapy compared to their low CMLS counterparts. In contrast, the low CMLS group, despite also displaying a “cold tumor” phenotype, showed a favorable prognosis and a heightened responsiveness to chemotherapy. This study's findings underscore the potential of targeting immune‐suppressive cells, particularly in patients with high CMLS, as a strategic approach to enhance OC prognosis. Furthermore, the redefined molecular subtypes and risk stratification, achieved through sophisticated multi‐omics analysis, provide a framework for the selection of therapeutic agents.