Deep Learning-Based Prognostic Model for Management of Men with Metastatic Castration-Resistant Prostate Cancer
The management of metastatic castration-resistant prostate cancer (mCRPC), especially the vulnerable patient population with diseases progressing on next-generation hormone therapies, remains a challenging clinical task. Recently, deep learning (DL) has revolutionized bioinformatics studies of cancer genomics given its unprecedented ability to characterize the complex nature of high-dimensional genomics data. This project will test the hypothesis that a sophisticated DL model captures intricate patterns of mutations that are predictive of prognosis. The goal of this study is to develop a novel DL prognostic model for mCRPC progressing on next-generation hormone therapies. The model will comprehensively incorporate germline and liquid-biopsy mutation profiles, together with clinical variables and biomarkers, to make predictions. Achievement of the proposed study will have a huge impact on the management of this vulnerable patient population of mCRPC.
The study addresses the unmet need for a prognostic model to guide the clinical management of metastatic castration-resistant prostate cancer (mCRPC) by developing innovative computational approaches. Our approaches integrate clinical variables and genetic/genomic data to predict treatment outcomes for this vulnerable patient population. We benchmarked state-of-the-art variant calling tools and implemented a bioinformatics pipeline to accurately detect germline variants. We applied the pipeline to whole-exome sequencing data from both the Cancer Genome Atlas (TCGA) cohort of prostate cancer and the Johns Hopkins cohort of mCRPC, enabling a systematic analysis of the landscape of germline variants.
Based on the mutation profiles and clinical follow-up data, a sophisticated deep learning model was developed to capture intricate patterns of mutations for prognosticating mCRPC patients. The model incorporates several unique features to address this challenging task, including a graph-based learner for gene mutations in genome-wide interaction networks, optimization based on a Cox regression-like cost function, and a transfer learning design to integrate large genomic data from different cohorts. To facilitate the drug discovery process, we extended the study to the Cancer Dependency Map (DepMap) by predicting cancer cells’ response to genetic knockouts. To promote the Findability, Accessibility, Interoperability, and Reusability (FAIR) of the findings, all relevant results are deposited in open-access repositories, such as the Code Ocean. User-friendly web tools are also developed for the use of biomedical researchers without the requirement of programming skills.
Various bioinformatics tools developed from the proposed study have resulted in at least 7 publications in high-impact journals, including Science Advances and Briefings in Bioinformatics, and presentations at prestigious conferences, such as the American Association for Cancer Research (AACR) Annual Meeting in 2022 and 2023 and the Conference on Intelligent Systems for Molecular Biology (ISMB) in 2022 and 2023. Beyond the research outcomes, the study has played a pivotal role in the PI's (Chiu) activation of his NCI R00 award to continue works relevant to this study, as well as his career transition to an independent faculty position at the UPMC Hillman Cancer Center of the University of Pittsburgh.
Learn More About Their Work
Science Advances Magazine | Volume 7 | Issue 39 | 24 September 2021
Predicting and characterizing a cancer dependency map of tumors with deep learning
Tool: DeepDEP: deep learning of a cancer dependency map using cancer genomics
NIH/NCI R00 Award (PI: Chiu)
Deep learning of drug sensitivity and genetic dependency of pediatric cancer cells
Oxford Academic | Bioinformatic Advances | 12 June 2023
DepLink: an R Shiny app to systematically link genetic and pharmacologic dependencies of cancer