BIOS-07. DATA–DRIVEN GREEDY MULTI-OMIC CONSOLIDATION REDEFINES SURVIVAL-AWARE MOLECULAR GRADES ACROSS CENTRAL NERVOUS SYSTEM TUMORS

Abstract Recent evidence showed that co-occurring 1p loss and 1q gain can re-stratify meningioma into outcome groups that outperform the 2021 WHO grades, demonstrating how a few carefully chosen genomic alterations may rival histology for prognosis. Building on this insight, we devised a greedy, semi-supervised pipeline that discovers the smallest yet most prognostic multi-omic units and constructs survival-aware molecular grades separately for each central-nervous-system (CNS) tumor type. Gene-level copy-number alterations (CNAs), hotspot mutations, RNA expression, and promoter-methylation sites were retrieved from multiple public cohorts for every major CNS cancer. Each locus entered elastic-net Cox models for overall survival (OS), adjusted for age, sex, extent of resection, treatments, and other confounding variables when available. Loci that improved cross-validated concordance after false-discovery correction were retained. A greedy hierarchical merge then combined loci when they were (i) genomically adjacent and (ii) highly co-occurring, proceeding only if the joint subunit further lowered Bayesian Information Criterion and maintained high C-index. These OS-enriched features then powered a tumor-specific Cox-Mixture Model that maximizes intra-cluster molecular coherence and inter-cluster OS separation, letting data—not histology—define new grade boundaries. The workflow rediscovered 1p loss with 1q gain as a dominant high-risk signal in meningioma, validating the approach. In other entities it confirmed well-known markers (e.g., 1p/19q codeletion in oligodendroglioma), offered subtle refinements to WHO stratification where appropriate, and proposed entirely new grading tiers where molecular criteria were lacking. Each cancer ultimately required only three to twenty consolidated features to generate Kaplan–Meier curves that were consistently more distinct, and often more prognostically informative, than existing classifications. By distilling thousands of loci into concise, interpretable marker panels and embedding OS information directly within tumor-specific clustering, our greedy Cox-Mixture framework delivers transparent molecular grading systems that either reinforce, refine, or redefine WHO schemes—providing a practical, data-driven roadmap for precision neuro-oncology.