Improving cancer treatment relies on understanding the variables that drive key biological processes in cancer. Studies have used molecular ‘omic data to predict diagnostic features, such as cancer type and immune phenotype, but the cost and time of measuring gene expression impedes discovery of clinical transcriptional biomarkers. Recent studies show that molecular information is encoded in histology images, which are collected routinely. For example, deep learning models can be trained to predict prognostic molecular features, such as genomic alterations, microsatellite instability, and gene expression, directly from histology. We are motivated by the possibility of developing practical clinical biomarkers by jointly leveraging genomics and histology using squamous cell carcinoma (SCC) samples, encompassing lung, cervical, and head and neck cancer types. There is a critical unmet need for comprehensive biomarkers that capture the complex biological processes underlying the cancer landscape. Deep learning-based histology biomarkers offer a portable, cost effective, and accessible opportunity to infer genomic information from histology anywhere in the world. Exploring the histology feature space and connecting histologic features with biomarker prediction allows for the requisite explainability for clinical integration.

Who is involved: Hanna Hieromnimon