Establishment of a Machine Learning Model for the Risk Assessment of Perineural Invasion in Head and Neck Squamous Cell Carcinoma

Abstract

Perineural invasion is a prevalent pathological finding in head and neck squamous cell carcinoma and a risk factor for unfavorable survival. An adequate diagnosis of perineural invasion by pathologic examination is limited due to the availability of tumor samples from surgical resection, which can arise in cases of definitive nonsurgical treatment. To address this medical need, we established a random forest prediction model for the risk assessment of perineural invasion, including occult perineural invasion, and characterized distinct cellular and molecular features based on our new and extended classification. RNA sequencing data of head and neck squamous cell carcinoma from The Cancer Genome Atlas were used as a training cohort to identify differentially expressed genes that are associated with perineural invasion. A random forest classification model was established based on these differentially expressed genes and was validated by inspection of H&E-stained whole image slides. Differences in epigenetic regulation and the mutational landscape were detected by an integrative analysis of multiomics data and single-cell RNA-sequencing data were analyzed. We identified a 44-gene expression signature related to perineural invasion and enriched for genes mainly expressed in cancer cells according to single-cell RNA-sequencing data. A machine learning model was trained based on the expression pattern of the 44-gene set with the unique feature to predict occult perineural invasion. This extended classification model enabled a more accurate analysis of alterations in the mutational landscape and epigenetic regulation by DNA methylation as well as quantitative and qualitative differences in the cellular composition in the tumor microenvironment between head and neck squamous cell carcinoma with or without perineural invasion. In conclusion, the newly established model could not only complement histopathologic examination as an additional diagnostic tool but also guide the identification of new drug targets for therapeutic intervention in future clinical trials with head and neck squamous cell carcinoma patients at a higher risk for treatment failure due to perineural invasion.

Context-Dependent Regulation of Peripheral Nerve Abundance by the PI3K Pathway in the Tumor Microenvironment of Head and Neck Squamous Cell Carcinoma

“…A cross-tabulation analysis revealed a highly significant enrichment of HPV16-positive OPSCC for the SC low group, which was almost absent in the SC high group ( p = 2.33 × 10 −6 ). Moreover, SC high tumors were enriched for perineural invasion based on the histopathologic annotation ( p = 2.50 × 10 −4 ) and a recently established machine learning model ( p = 1.00 × 10 −3 ) ( Table S7 ) [ 39 ]. To further validate these findings, the SC scores based on single-cell RNA-seq data were compared with tumor cells from HPV-negative and HPV-positive tumors.…”

“…This specific difference in the mutational landscape, in combination with the lack of somatic TP53 mutations, may explain the reduced abundance of peripheral nerves in the TME of HPV16-positive OPSCC. Accordingly, HPV16-positive OPSCC do not only exhibit lower SC scores, but also have reduced expression of synaptic markers and electrical activity within tumors and a lower frequency of PNI [ 24 , 39 ]. However, it is worth noting that the potential impact of PI3K-MTOR signaling on neuro- and/or axonogenesis appears to be context-dependent and influenced by the TP53 mutation status.…”

Tumor Neurobiology in the Pathogenesis and Therapy of Head and Neck Cancer

“…These findings suggest that the current diagnostic criteria for PNI should be updated based on the nerve-tumor distance [23]. Weusthof et al established a PNIrelated 44-gene signature based on RNA sequencing data and trained a random forest model to predict occult perineural invasion [73]. However, further efforts are needed to unify the diagnostic criteria for PNI, improve the diagnostic accuracy, and consider the clinical practicality.…”

“…However, more emerging neuroimaging techniques should be leveraged to map nerve involvement and interactions longitudinally in vivo. At the molecular level, multi-omics analyses, such as single-cell RNA sequencing, spatial transcriptomics, and proteomic approaches, have been developed to investigate the molecular characteristics of neurons, tumor cells, and other cellular components in the TME [1,23,73]. Other techniques to explore the mechanism underlying the cancer-neuron crosstalk include the generation of enhancer-based lentiviruses, single-molecule imaging tools, and in vivo genetic perturbation [1].…”