Deep Learning for Multi-Omics-Based Prediction of Cancer Recurrence and Treatment Non-Response

About This Grant

PROJECT SUMMARY/ABSTRACT Cancer recurrence significantly contributes to increased morbidity, reduced survival, and poor treatment response in patients with solid tumors, including glioma, lung adenocarcinoma (LUAD), and breast invasive carcinomas (BRCA). As such, understanding the factors that drive rapid cancer recurrence and treatment failure, a common cause of early recurrence, is essential to improving patient outcomes. Despite recent advancements in machine learning (ML) and deep learning (DL) approaches to recurrence prediction, current predictive models face limitations, including complex implementation, reliance on manual genomic feature selection, and limited model explainability. Moreover, treatment response models seldom integrate patient-specific protein-protein interaction (PPI) networks despite their critical roles in cancer development and therapeutic resistance mechanisms. To address these significant gaps, this research proposes innovative multi-modal DL architectures that leverage genomic and clinical data from The Cancer Genome Atlas (TCGA) for robust predictions of early cancer recurrence and therapy non-response in glioma, LUAD, and BRCA. The long-term objective of this research is to uncover potential driving factors of recurrence and treatment failure that may aid in the larger scientific community’s effort to improve clinical oncology outcomes. Aim 1 focuses on constructing multimodal attention-based DL architectures to predict early recurrence and incorporates somatic mutation, mRNA expression, and clinical data. The resulting models will undergo rigorous explainability assessments to verify model validity and determine the most informative clinical and genomic features. The final models will be made accessible through a downloadable user-friendly interface, enabling researchers to apply and fine-tune these predictive tools to independent datasets. Aim 2 focuses on constructing DL models integrating patient-specific PPI networks to predict treatment non-response. Using TCGA genomic and clinical data alongside PPI data from Proteinarium, graph neural networks will capture experimentally validated interactions between protein nodes, prioritizing hub genes potentially involved in therapy failure. As before, the models will be subject to rigorous explainability assessment. This research will advance our understanding of the driving factors behind cancer recurrence and treatment failure and yield tools that can identify potential drug targets and mechanisms of action for future research in an unbiased, data-driven manner. I will conduct this research in the Uzun Translational Bioinformatics Lab at Brown University, which specializes in PPI modeling and DL approaches to recurrence prediction, under the guidance of experts in genomics, pathology, oncology, clinical informatics, and ML/DL. Completing the proposed research will significantly contribute to my career goal of utilizing DL methodologies and omics datasets to conduct targeted cancer therapy research and reduce the incidence of cancer relapse.