Predicting Oral Cancer From Precursor Lesions: The Case for a Standardized Framework of Analysis to Improve Prediction Modeling.

Abstract

Oral Cancer often occurs from the transformation of precursor lesions, and this offers an opportunity for early detection. Current methods to assess risk of precursor lesion progression to oral cancer incompletely predict risk. A multimodal framework that leverages machine learning is needed to improve prediction. In this manuscript, we propose more standardization in the study of OPMD across sites and propose a framework aligned to the principles of the NCI's Human Tissue Atlas Network (HTAN). We explore whole exome sequencing for genomic risk, methylation array for environmental genetic regulation, RNA sequencing for transcriptional control of cellular programs with spatial transcriptomics to fulfill the spatial requirements of HTAN. Moreover, we propose a machine learning (ML) analysis whereby a learned image-molecular association is created by integrating H&E histology and sequencing results. Once these associations are learned, a "Teacher-Student" model can be created whereby the Student model can be deployed to infer the "molecular associations" from H&E and patient covariates alone. Such a framework integrating multiomic data across multiple different biologic processes and time points can help to better understand and potentially even predict the biologic events that underlie oral cancer initiation.