ASSESSING PREDICTIVE MODELING APPROACHES FOR OPTIMIZING THERAPY IN ADVANCED LARYNGEAL AND HYPOPHARYNGEAL CANCERS

Background: Laryngeal (LC) and hypopharyngeal (HPC) cancers are often diagnosed at an advanced stage (III or IV) due to the absence of early symptoms and biomarkers. First-line therapies include both surgical and non-surgical approaches. Since surgical removal, comprising total and partial laryngectomy, is highly invasive, organ-preservation strategies are frequently preferable as they help maintain organ functionality. Induction chemotherapy (ICT) is one such preservation approach; however, it is not suitable for all patients and may contribute to drug resistance in some cases. Identifying a molecular signature of ICT response in both responsive and non-responsive patients could enable prediction of ICT outcomes in patients with advanced disease, providing valuable insights for therapy decisions. Similarly, a predictive model for organ functionality could help identify patients who may benefit from alternative approaches. Moreover, exploring new combined therapies for non-responsive patients could improve survival outcomes. Methods: We conducted a retrospective analysis on a cohort of FFPE samples from patients with advanced LC and HPC treated solely with ICT across various European hospitals. Laser capture microdissection was used collecting samples enriched in tumoral cells, followed by RNA-extraction, library preparation, and RNA-sequencing. Leveraging machine learning algorithms, we generated three predictive models: (i) a combined model integrating ICT response and LED, (ii) a LED-only model, and (iii) an ICT-response model. Models were trained with iterative validation across 100 random train-test splits to ensure robustness and generated using transcriptomic, clinical, and integrated data. In addition, in vitro cisplatin-resistant cell models were established to investigate resistance-associated genes, which were subsequently integrated into predictive models. Finally, immune-related gene expression was assessed to explore the potential integration of immunotherapy. Results: Survival analyses demonstrated that ICT response strongly correlated with clinical outcomes (OS and PFS), with resistant patients showing significantly poorer prognosis. Similarly, loss of laryngo-esophageal functionality (LED) was also associated with reduced survival and quality of life, underscoring the need for functional prediction alongside therapeutic response. The combined model integrating ICT response and LED achieved high predictive accuracy, with particularly low rate of misclassification when transcriptomic data were included. The LED model demonstrated that molecular data, rather than clinical variables, drive the prediction of organ functionality, reaching robust accuracy without improvement from clinical data integration. Conversely, the ICT-response model significantly benefited from the inclusion of both clinical and molecular features, achieving AUC values above 80%, and further improved by incorporating cisplatin-resistance genes, such as OLFM4 and MGMT, identified in vitro. Importantly, immune profiling revealed that most tumor exhibited an immune-“COLD” phenotype, although a subset of partial responders with preserved organ functionality displayed a “HOT” phenotype, suggesting that ICT combined with immune checkpoint inhibitors could represent a promising strategy for selected patients. Conclusions: This study demonstrates the feasibility of integrating RNA-sequencing data and machine learning to predict ICT response and post-treatment functionality in LC and HPC. The combination of molecular and clinical data allows a more accurate patient stratification, supporting the development of personalized therapeutic strategies and the implementation of predictive biomarkers in clinical practice.