Machine learning driven index of tumor multinucleation correlates with survival and suppressed anti-tumor immunity in head and neck squamous cell carcinoma patients

Objectives: Matching treatment intensity to tumor biology is critical to precision oncology for head and neck squamous cell carcinoma (HNSCC) patients. We sought to identify biological features of tumor cell multinucleation, previously shown by us to correlate with survival in oropharyngeal (OP) SCC using a machine learning approach.

Materials and methods: Hematoxylin and eosin images from an institutional OPSCC cohort formed the training set (D_Tr). TCGA HNSCC patients (oral cavity, oropharynx and larynx/hypopharynx) formed the validation set (D_V). Deep learning models were trained in D_Tr to calculate a multinucleation index (MuNI) score. Gene set enrichment analysis (GSEA) was then used to explore correlations between MuNI and tumor biology.

Results: MuNI correlated with overall survival. A multivariable nomogram that included MuNI, age, race, sex, T/N stage, and smoking status yielded a C-index of 0.65, and MuNI was prognostic of overall survival (2.25, 1.07-4.71, 0.03), independent of the other variables. High MuNI scores correlated with depletion of effector immunocyte subsets across all HNSCC sites independent of HPV and TP53 mutational status although the correlations were strongest in wild-type TP53 tumors potentially due to aberrant mitotic events and activation of DNA-repair mechanisms.

Conclusion: MuNI is associated with survival in HNSCC across subsites. This may be driven by an association between high levels of multinucleation and a suppressive (potentially exhausted) tumor immune microenvironment. Mechanistic studies examining the link between multinucleation and tumor immunity will be required to characterize biological drivers of multinucleation and their impact on treatment response and outcomes.