CTIMP

Cross Tissue gene expression IMPutation

About

CTIMP uses a multivariate-response penalized regression to predict cross-tissue gene expression. By imposing group-lasso penalty on each row of coefficient matrix, CTIMP reaches higher accuracy and robust prediction over penalized regression trained for each response separately. The approximation procedure proposed for coordinate descent iteration makes training faster and yields small approximation error. Details of CTIMP could be found in our UTMOST manuscript.

The repo also contains codes for replicating analysis in UTMOST manuscript and training cross-tissue gene expression prediction model with GTEx data as a demonstration.

If you use results generated from this software, please cite:

• Hu, et al. A statistical framework for cross-tissue transcriptome-wide association analysis, Nature Genetics, 2019.

Tutorial

Clone repo

git clone https://github.com/yiminghu/CTIMP.git

Compile the optim.c

cd CTIMP
R CMD SHLIB optim.c

Input files and formats

• feature file (e.g. example/X.txt): covariate matrix with each row being an observation and each column being a variable. The first column is the ID of each observation. No header line.

id1 ftr1 ftr2 ... ftrM

• response file (e.g. example/Y*.txt): one file for one response, each file contains two columns: ID and response values. Due to the missingness in reponse matrix, each file may have different number of rows (number of observations).

• info file: files contain column infomation of X, i.e. names of features, see example/ for details

Usage

X=example/X.txt
Y_folder=example/Y_folder/
info=example/info.txt
ntune=5 #number of grids for each tuning parameter
output_path=example/output/
mkdir ${output_path}
output_prefix=test # prefix of output files

Rscript main.R ${X} ${info} ${Y_folder} ${ntune} ${output_prefix} ${output_path}

main.R will perform 5-fold cross-validation to select tuning parameters and generate estimation of the coefficient matrix.

Final coefficient estimates will be saved to ${output_path}/${output_prefix}.est, it will also writes two rdata files:

• ${output_prefix}.cv.evaluation.RData: contains prediction metrics (mse, rsq, adjusted mse) on cross-validated test data
• ${output_prefix}.prediction_on_all_data.RData: contains the final model (trained on entire data with tuning parameters selected from cv) and the final prediction on all data