Integrated multiomics analysis and machine learning refine molecular subtypes and prognosis for thyroid cancer

BACKGROUND: Thyroid cancer (THCA) exhibits high molecular heterogeneity, posing challenges for precise prognosis and personalized therapy. Most existing models rely on single-omics data and limited algorithms, reducing robustness and clinical value. METHODS: We integrated five omics layers from THCA patients using eleven clustering algorithms to identify molecular subtypes. Based on stable prognosis-related genes (SPRGs), we applied 99 combinations of ten machine learning methods to construct a robust prognostic model-Consensus Machine Learning-Driven Signature (CMLS). The model was validated across multiple internal and external cohorts. Immunogenomic characteristics and drug sensitivity were also evaluated. RESULTS: Three molecular subtypes (CS1-CS3) with distinct clinical outcomes and molecular features were identified; CS2 showed the worst prognosis. A nine-gene CMLS was established, demonstrating strong prognostic performance across cohorts. Patients in the low-CMLS group had better outcomes, stronger immune infiltration, higher TMB/TNB, and greater predicted responsiveness to immunotherapy. Conversely, the high-CMLS group exhibited poor prognosis and lower immunotherapy sensitivity. Drug screening identified six candidate agents for high-CMLS patients. CONCLUSION: Our study provides a robust multiomics-based classification of THCA and develops a clinically relevant CMLS model for prognostic prediction and therapy guidance. These findings may facilitate risk stratification and inform personalized treatment strategies in clinical practice.