TEEBot is a tool to predict tissue-specific expression (TSGE) from an individual’s blood transcriptome and demographic information (age, gender and race), with clinical implications. If his/her genotype information is avalibale, it could boost the performance of TEEBot. We trained TEEBot on GTEx version 6 and evaluated its performance in a cross-validation manner. For each gene in each target tissue, we first evaluate its predictability based on LLR test and then fit a lasso regression model to estimate its TSGE. Our models require Whole blood gene expression (WBGE), Whole Blood splicing (WBSp) information, and three demographic ‘confounding’ factors (Age, Race, and Sex), with genetype information as one additonal option.
Download data files from GTEx Portal and dbGaP. The data include gene expression, transcript expression, phenotype and genotype information (optional).
The code to predict the gene expression of a target tissue consists of the following steps:
For each gene, the top PCs of Whole blood transcriptome (top 10 PCs for gene expression and top 20 PCs for splicing profile) as features are used to build a gene specific model to predict its expression in the target tissue. Five-fold cross validation are performed for all the genes across tissues. The model based on lasso regression is implemented using cv.glmnet() function from glmnet R-package.
The prediction accuracy for each gene are evaluated using Pearson correlation test between the predicted expression and the ground truth. Likelihood ratio test is also performed for each gene to assess the independent contribution of blood transcriptome beyond the confounders (age, race and sex). For each gene we provide both its prediatability score (in terms of Pearson correlation coefficient) and FDR of the LLR p-value. We only report the prediction accuracies of genes which pass the likelihood ratio test (FDR<=0.05) and also have pearson correlation coefficient values above a predefined threshold.
Code: regression_model_articleoutput.r
Input are the gene expression, phenotypen and genotype file downloaded from GTEx Portal and dbGaP. \ i) Gene expression data: All_Tissue_Site_Details_Analysis.combined.rpkm.gct \ ii) Phenotype data: phs000424.v6.pht002742.v6.p1.c1.GTEx_Subject_Phenotypes.GRU.txt \ iii) Sample attribute: GTEx_Data_V6_Annotations_SampleAttributesDS.txt \ iv) Gene property file from ncbi: Homo_sapiens.gene_info (downloaded from https://ftp.ncbi.nih.gov/gene/DATA/GENE_INFO/Mammalia/ on march2016) \ v) Genotype data: genotype.v6.chr1_22.Rdata \ These files are read in the code within function "format_gtexdatav6()". We also have a input file "tissue_pair_work_v6.RData" which contains a table with first column as whole blood, second column target tissue and third colum are the number of common samples. All these files needs to be kept within the "workdir/input" folder. Along with these files we need to keep the function.r in the "workdir/input" folder.
The input parameters to run the code are index, run, fold, ENS. 1) "index" are the tissue ids, the ids for the 32 target tissues are as follows 1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,32,34,37,39. 2) "run" indicated which model to compute, M1 or M2. To compute model M1 run=100, and for M2 run=110 3) fold : how many fold cross validation we want. 4) ENS : how many independent predictions we want to do.
Each Target tissue prediction file for example "Whole Blood_xx_lasso_regress_G_Spl_10_20.Rdata", where xx will be replaced by the target tissue name. This Rdata contanis two objects, gene.regress.glmnet and gene.select.LLR. 1) gene.regress.glmnet : is a list, gene.regress.glmnet[["Mean-first"]] contains our desired correlation coefficient that is the predictibility scores. 2) gene.select.LLR : is a list with LLR ratio and pvalue
note: Download the code file (regression_model_articleoutput.r) and the function.r in the same folder. Example path to store the codes is "/home/Desktop/Imputation". Within Imputation make a folder named "input" and "output". In the "output" folder the output of the codes automatically gets saved.
For the code the following R-packages needs to be installed "glmnet", "data.table", "foreach", "doMC", "ROCR", "lmtest".