Unsupervised Identification of Cancer Attractor States through the Lens of Embryonic Origin and Cancer Hallmarks

Abstract Background Oncogenesis is a highly intricate process characterized by the transition of normal cells into aberrant biological attractor states. These states emerge from the non-random combinatorial interactions between inherent lineage-specific programs, governed by Gene Regulatory Networks (GRNs), and acquired somatic mutations. While the hallmarks of cancer provide a functional framework for transformation, current genomic models often fail to account for the topographical constraints of the cell of origin. We hypothesized that integrating embryological origin (EO) with hallmark-related mutations (HRM) enables the unsupervised identification of stable, pan-cancer attractor states. Methods Utilizing the MSK-MET database (n=25,775), we annotated somatic variants across 15 hallmarks. We developed and compared two unsupervised clustering models using Jaccard-distanced t-SNE and K-means algorithms: one based solely on HRM profiles and another integrating EO with HRM (EO/HRM). Model robustness and attractor stability were evaluated via bootstrap resampling and five-fold cross-validation. Results Hallmark-related mutations were identified in 95.5% of tumors. The EO/HRM model identified 11 distinct clusters and demonstrated superior stability (ARI: 0.74 vs. 0.70) and prognostic performance (C-index: 0.59 vs. 0.56) compared to the HRM-only model. These clusters represent stable functional basins, such as the TP53-driven minimalist attractor in Cluster C6 and the hyper-unstable attractor in Cluster C10. The clusters exhibited unique, organ-specific metastatic predilections independent of histological subtypes. Critically, we observed prognostic inversions, such as Cluster C0 representing a favorable attractor in lung adenocarcinoma but an aggressive state in prostate carcinoma, illustrating that attractor behavior is gated by the embryological landscape. Conclusion Integrating embryological origin as a proxy for inherent cellular programming significantly enhances the identification of stable, prognostically relevant biological attractor states. This framework provides a novel strategy for functional risk stratification and the development of context-aware precision therapies aimed at destabilizing malignant attractors.