R code for Survival prognosis and variable selection: A case study for metastatic castrate resistant prostate cancer patients

A paper written in relation to [The Prostate Cancer DREAM Challenge](https://www.synapse.org/ ProstateCancerChallenge).

Requirements

Run the following R code to install required packages.

install.packages(
  c("survival",
    "xtable",
    "ggplot2",
    "reshape2",
    "mice",
    "mgcv",
    "timeROC",
    "Bolstad2",
    "ROCR",
    "randomForestSRC",
    "lattice",
    "RColorBrewer",
    "glmnet")
)

One additional package is required, which is not on CRAN, but is a fork of the gbm R package available from GitHub.

install.packages("devtools")
devtools::install_github("uci-cbcl/GBMCI")

Running the code

The Rmd file main.Rmd in the directory Main includes the implementation of the trainModels function. This is the main training function, which relies on the implementations of a number of different functions for the imputation and the training of the different models. These implementations are found in the other subdirectories, and sourced in the beginning of main.Rmd.

Running main.Rmd (using knitr) loads and preprocesses the data, computes various summaries and constructs descriptive figures. It then runs the 5-fold cross-validation, stores the results in a file, refits all models to the entire data set and makes predictions on the validation data. It finally constructs plots of the selection probabilities for stability selection.

There are instructions in main.Rmd for setting seeds to obtain the exact same results as in the paper. The results from the cross-validation are included in the files mainResults12, mainResults1213 and mainResults121314. The validation predictions for the different methods are stored in the directories ValidationResults and ValidationResults2 (two different seeds were used). The results from the validation predictions are stored in the files ValidationResults.txt and ValidationResults2.txt. Note that these validation results were obtained by submission of the predictions via the web interface for the competition. Thus the results cannot be programmatically reproduced.

The Rmd file results.Rmd in the directory Paper reads the cross-validation and validation results and constructs various figures including those in the paper.

Misc.

The directory Write-up contains the original write-up submitted for the competition.