A Bayesian nonparametric sparse factor regression for predicting sensitivity of cancer cell lines to a range of drugs.
Drug screening studies assay the sensitivity of a range of cancer cell lines across an array of anti-cancer therapeutics. Alongside these sensitivity measurements high dimensional molecular characterizations of the cell lines are available, including gene expression, copy number variation and genomic mutations. lacrosse is a sparse multitask regression model which learns discriminative latent characteristics that predict drug sensitivity and are associated with specific molecular features. We use Bayesian nonparametrics (specifically an Indian buffet process prior) to automatically infer the appropriate number of these latent characteristics. The resulting analysis couples high predictive performance with interpretability since each latent characteristic involves a typically small set of drugs, cell lines and genomic features. lacrosse uncovers a number of drug-gene sensitivity associations missed by single gene analyses.
The lacrosse R package itself can be installed using e.g.
if (!require("devtools")) install.packages("devtools", repos='http://cran.us.r-project.org')
devtools::install_github("davidaknowles/lacrosse/lacrosse")
which will automatically install the required dependencies.
To run lacrosse on the CTRPv2 dataset you will need the PharmacoGx R package [https://academic.oup.com/bioinformatics/article-lookup/doi/10.1093/bioinformatics/btv723], installed as
source("https://bioconductor.org/biocLite.R")
biocLite("PharmacoGx")
You'll also need the tidyverse, magrittr and doMC (or foreach) packages installed.
code/run_lacrosse.R gives a working example of running lacrosse on the CTRPv2 data.