Genetic Predictors of Prostate Cancer Survival
Therefore, major clinical questions in the management of prostate cancer today is the decision of who to screen for prostate cancer and what to do upon initial diagnosis of prostate cancer. While clinical factors such as age, cancer stage and grade, and PSA levels at diagnosis can be used to stratify patients and guide treatment decisions at the time of diagnosis, the uncertainty that remains in these predictions as to the potential course of the disease makes the decision between treatment or active surveillance difficult. To address these issues, members of the study team have developed a four-kallikrein biomarker panel that improves the prediction, prior to any diagnosis of prostate cancer, of which men may die of prostate cancer (Sjoberg et al., 2018). While these predictors are useful, they still have a range of uncertainty justifying the development of new markers. Improved predictors that can help identify men at risk of dying from prostate cancer, especially those that can be used in a screening situation, are sorely needed.
Here, we propose a germline genomic approach to identify men at risk of dying from prostate cancer. This work builds on our preliminary work that has identified common genetic polymorphisms associated with survival time after diagnosis, independent of known prognostic factors (Li et al, 2018). By leveraging recent computational advances in genomic analysis, we will take a gene-centered approach to identify genes for which genetically controlled transcriptional alterations and/or functional coding mutations influence survival time in prostate cancer. Using these genes, we will build and test models designed to predict death from prostate cancer both prior to any diagnosis for use in a screening setting and after diagnosis to aide in treatment selection.
1. Identify genes for which germline genetic variants alter the risk of dying from prostate cancer. We hypothesize that both variants that alter gene expression levels and that alter the sequence of the translated protein will influence risk of dying from prostate cancer. We will impute all common variants in each of the GWAS datasets using either the Haplotype Reference Consortium panel or a custom panel we are building based on whole genome sequences from prostate cancer cases. We will augment this imputation with exome or genome sequencing data from cohorts in which those are available. For each gene, we will then impute gene expression levels in each individuals using the TWAS approach with custom prostate specific models we are creating. We will also look at coding variation in the data. In call cases, we will test for association with survival time adjusting for age, stage, and grade at diagnosis using a competing risks model.
2. Build and validate a predictive model of prostate cancer outcome based on germline genotype in addition to known clinical predictors to improve patient stratification at the time of diagnosis. We propose that combining signals from numerous genes with a moderate effect each on survival through machine learning approaches will allow for more precise prediction of prostate cancer outcome. Specifically, we will use a random forest approach with the imputed expression levels and coding alteration status for each gene to be included in the model.
3. Improve our 4-kallikrein biomarker predictor of lethal prostate cancer through incorporation of genetic data predictive of both risk and outcome. We hypothesize that death from prostate caner can be modeled as the cumulation of developing the disease and having the disease progress. We will incorporate both known genetic prostate cancer risk predictors from the literature and genes we identify in aim 1 with data on the 4-kallikrein panel. Specifically for the PLCO data, we will use and link the 4-kallikrein panel results previously generated in the PLCO by collaborator Lilja (PLCO project 2010-0042).
Hans Lilja, Memorial Sloan Kettering Cancer Center
Xiaoyu Song, Icahn School of Medicine at Mount Sinai