Machine Learning–Assisted Analysis of the Oral Cancer Immune Microenvironment: From Single-Cell Level to Prognostic Model Construction

Abstract

Oral cancer is among the most prevalent malignant tumours worldwide; prognosis can be affected by several factors, including molecular subtypes, immune microenvironment and clinical characteristics. In this study, we aimed to apply machine learning methods in conjunction with single-cell sequencing data to characterise the immune microenvironment of oral cancer and build an immune infiltration prediction model to provide a theoretical basis for the personalised therapy and prognosis assessment of oral cancer. Clinico-genomic data were obtained from patients with oral cancer and single-cell sequencing was utilised to delineate the immune cell composition in the tumour microenvironment. Model construction and immune-related gene screening were performed using machine learning algorithms such as Lasso regression, random forest and gradient boosting machine. We assessed the predictive performance of the model by cross-validation on its training dataset and by testing the model on an independent dataset. Certain subsets of immune cells correlate with the prognosis of patients with oral cancer. C-index (given in supplementary) yielded a good discrimination ability (C-index > 0.75) in the training set and validation set. Moreover, the model-identified immune-related genes presented remarkable expression differences in the two different risk groups and played important roles in the response to immune therapy. By exploring the complexity of the oral cancer immune microenvironment with machine learning techniques, in this study, we build a reliable prognostic model based on immune infiltration. The model could be applied in clinical practice to personalisation treatment decision-making and prognosis evaluation.