Artificial intelligence-based epigenomic, transcriptomic and histologic signatures of tobacco use in oral squamous cell carcinoma

Abstract

Oral squamous cell carcinoma (OSCC) biomarker studies rarely employ multi-omic biomarker strategies and pertinent clinicopathologic characteristics to predict mortality. In this study we determine for the first time a combined epigenetic, gene expression, and histology signature that differentiates between patients with different tobacco use history (heavy tobacco use with ≥10 pack years vs. no tobacco use). Using The Cancer Genome Atlas (TCGA) cohort (n = 257) and an internal cohort (n = 40), we identify 3 epigenetic markers (GPR15, GNG12, GDNF) and 13 expression markers (IGHA2, SCG5, RPL3L, NTRK1, CD96, BMP6, TFPI2, EFEMP2, RYR3, DMTN, GPD2, BAALC, and FMO3), which are dysregulated in OSCC patients who were never smokers vs. those who have a ≥ 10 pack year history. While mortality risk prediction based on smoking status and clinicopathologic covariates alone is inaccurate (c-statistic = 0.57), the combined epigenetic/expression and histologic signature has a c-statistic = 0.9409 in predicting 5-year mortality in OSCC patients.

Fig. 1. TCGA methylation analysis results.

A Volcano plot of batched corrected data. Only differentially methylated sites with unadjusted p < 0.1 are included. Unadjusted p < 0.05 and log fold change > +/−0.5 are considered significant. B QQ plot of the batch corrected data, which compares the expected to observed -log₁₀P, and demonstrates an inflation factor = 0.99.

Fig. 2. Methylation and RNA Seq array work flow.

The analysis steps for the array data from the TCGA cohort are shown, with (A) representing the methylation array workflow and (B) representing the RNA Seq workflow.

Fig. 3. Functional pathway analysis of RNASeq biomarkers.

A GO BP top 10 dysregulated pathways. B Top 10 dysregulated KEGG pathways. C Top 5 Reactome gene sets.

Fig. 4. Functional pathway analysis of methylation biomarkers.

A GO BP gene concept network. B Top 10 dysregulated KEGG pathways. C Dot plot of ORA Reactome gene sets.

Fig. 5. Digital histopathology analysis with a deep learning model designed to predict patient smoking status.

A Whole Slide Images from 203 TCGA hematoxylin and eosin stained histopathology slides served as training data for a deep learning model constructed with the Slideflow pipeline. B Expert pathologists annotated regions of interest (ROI) on each WSI. Within each ROI, the WSIs are divided into tiles of size 299 pixels × 299 pixels. Tiles underwent stain normalization and augmentation prior to model training. C UMAP of the post-convolution layer activations from all images in the validation set. Plotted tiles are a subset of all image tiles within the validation set.

Fig. 6. Deep learning model explainability analysis.

A Heat map of the model’s logit score assigned to a given location within the image’s ROI. B UMAP of the post-convolutional layer activations from all images in the model’s validation set with a label of the model’s smoking status prediction (1- heavy smoker, 0- non-smoker). C UMAP in B labeled with the ground truth smoking status prediction. D Heat map of the model’s uncertainty quantification of the outcome prediction assigned to a given location within the image’s ROI. E UMAP in B labeled with the uncertainty quantification. F UMAP in B labeled with the TCGA donating site. G UMAP in B labeled with anatomic site. H UMAP in B labeled with perineural invasion status (PNI).