A Breast Cancer Recurrence Risk Predictor Using a Deep Learning Multi-omics Data Integration Framework

Over the last few decades, many breast cancer patients have become disease-free due to the advancements in detection, diagnosis, and treatment. However, cancer can come back for almost 30% of patients who are disease free after initial treatment. Current methods of recurrence risk prediction use clinical or phenotypic characteristics which may not give a complete understanding of the underlying genetic causes of the recurrence. Although there are some gene profiling available for recurrence prediction such as Oncotype DX, they are limited to estrogen receptor-positive tumors. The aim of this study was to utilize multi-omics data of gene expression, copy number variation, mutation, and microRNA, in the recurrence rate prediction of any breast cancer tumor. For this project, a modified deep learning-based autoencoder structure known as AIME was used to extract accurate integrated embeddings. Unlike a typical autoencoder, where the input and output layers are the same, AIME uses two different data types such as gene expression and copy number variation, and, at the same time, adjusts for the confounding factors which in this case, was the estrogen receptor status. The integrated embedding data from AIME was then applied to a supervised machine-learning approach of Random Forest algorithm to classify samples into disease-free and tumor-regressed categories. As part of the AIME output, a list of the 25 most significant genes for breast cancer recurrence was identified which could be effectively utilized both as biomarkers to improve patient diagnostics and develop targeted therapies, and as molecular signature to investigate mechanisms of tumor progression and recursion. This approach can be expanded to include other cancer types to predict and understand tumor recursion.