PredictDB Pipeline for GTEx v7 data
Comparison of V6p and V7 models can be found here
Scott's write up
GTEx v7 PredictDB Models - README
Scott Dickinson 11/13/2017
This directory contains gene expression prediction models for 48 different human tissues, meant for use with the Im Lab's software PrediXcan and S-PrediXcan. These models were trained using GTEx v7 data, subsampled to use European samples only. Genotype data was also imputed using the University of Michigan Imputation Server, with HRC (Version r1.1 2016) as a reference panel. For each model, there is also a corresponding file with covariance data for the SNPs in each model. These files are necessary to use S-PrediXcan.
Genotype data - GTEx genotype data was downloaded from dbGaP in the form of a vcf file with all original 635 samples. GTEx standard QC procedures had already been applied at time of download. We then filtered the samples to include Europeans only, using race as the filtering criteria. The data was then uploaded to the University of Michigan Imputation server for imputation, with the specific goal of obtaining more complete coverage of HapMap SNPs. We selected HRC (Version r1.1 2016) as the reference panel. SNPs were then filtered to those with minor allele frequency >= 1%, R2 >= 0.8, biallelic only, and those which would be unambiguously stranded, i.e. the polymorphism is not to the base's complement. SNPs were then mapped to rsids in dbSNP version 150, using position and alleles as matching criteria. Once genotype data was subsetted again when doing analysis on a tissue, SNPs were again filtered to have minor allele frequency >= 1% according to the samples that had RNA sequencing performed for that tissue. For analysis, genotypes were encoded on a continuous range from 0 to 2, with the value denoting the estimated count of the second, or effect, allele.
Expression data - Expression data for each tissue was quantified by the GTEx consortium, and we made extensive use of their pipeline for eQTL analysis, which can be found at https://github.com/broadinstitute/gtex-pipeline. We normalized the RPKM values using the script normalize_expression.py. In short, this script quantile normalizes the expression values to the average empirical distribution observed across samples, and then for each gene, expression values are inverse quantile normalized to a standard normal distribution across samples. Genes were selected based on expression thresholds of >0.1 RPKM in >=10 samples and >5 reads in >=10 samples.
Covariates - Before training models to predict gene expression, expression values were adjusted for the following covariates:
PEER Factors - PEER Factors were calculated using the GTEx pipeline docker container. The number of PEER factors was chosen according to GTEx v7 protocol with the number of samples being the determining factor. If the number of samples was greater than or equal to 350, we used 60 PEER factors. If the number of samples was between 250 and 350, we used 45. Between 150 and 250, we used 30, and less than 150 we used 15. Sex - As specified by GTEx Subject Phenotypes file Genetic Principal Components - We used the top 3 genetic principal compoonents as computed by GTEx. Sequencing Platform - Whole Genome Sequencing for samples was carried out on two different types of machines. This information can be found in the Sample Attributes file.
Expression was adjusted by performing a multivariate linear regression with all covariates, pulling the residual values, and then assigning the residuals to be the new expression values.
Annotations - Gene annotation was derived from gencode v19, using GTEx's collapsed gene model. SNP annotation was derived from the vcf from GTEx, a HapMap annotation file for the CEU population, and data from dbSNP version 150.
Nested Cross Validated Elastic-Net - In previous versions of PredictDB, we employed 10-fold cross-validated elastic-net to tune the parameter lambda, and then estimated the significance of the model. It recently became apparent that this was biasing the significance measures because we were using the same data to tune the parameter lambda and assess the performance. To correct for this problem, we used the following "nested" cross validation procedure:
Randomly split the data into 5 folds.
For each fold:
a. Remove the fold from the data.
b. Use the remaining data to train an elastic-net model using 10-fold cross-validation to tune the lambda parameter.
c. With the trained model, predict on the hold out fold, and get various test statistics for how the model performs.
Calculate the average and standard deviation of each of the significance statistics, where applicable. This should provide a reasonable estimate for how well the model will generalize to new data.
Train a new elastic-net model using all of the data. Again, use 10-fold cross validation to tune the lambda parameter. The non-zero weights from this final model are what are saved in the database, provided the model meets significance criteria.
A model was determined to be "significant" if the average pearson correlation between predicted and observed during nested cross validation was greater than 0.1 (equivalent to R2 > 0.01) and the estimated p-value for this statistic was less than 0.05. See below for how the p-value was calculated.
gene - Ensembl ID of the gene
rsid - rsid of the SNP. rsids are from dbSNP version 150.
varID - String label of the format chromosome_position_allele1_allele2_build. All varIDs are from build 37 of the Human Reference Genome.
ref_allele - Allele which appears on Reference Genome
eff_allele - Alternative/Effect allele.
weight - The weight for this SNP in predicting expression for the gene. In predicting the expression for the gene, the weight is multiplied by the count, or estimated count, of the effect allele in the individual. This value is added to all other weighted SNPs in the model.
This table contains summary statistics about each gene model for the tissue
gene - Ensembl ID of the gene
genename - HUGO symbol for the gene
gene_type - Type of the gene according to Gencode v19. We have only included models for protein coding genes, pseudogenes, and lincRNAs.
alpha - The alpha parameter used with the R package glmnet to train the elastic net model. Here, we used 0.5 for every gene.
n_snps_in_window - The number of cis-SNPs examined for model training. This is the number of SNPs found within 1 million base pairs upstream of the transcription start site and 1 million base pairs downstream of the transcription end site. Sites are determined from Gencode v19.
n.snps.in.model - The number of SNPs within the cis window that have non-zero weights, as found by elastic-net.
lambda_min_mse - The lambda parameter for elastic-net found to have the minimum average mean square error on the hold out folds during cross validation.
test_R2_avg - The average coefficient of determination when predicting values of the hold out fold during nested cross validation. R2 is defined as
1 - sum((y_observed-y_predicted)**2) / sum((y_observed-mean(y_observed)**2)
NB: It is possible for this value to be negative and still achieve a positive correlation between predicted and observed.
test_R2_sd - The standard deviation of the coefficient of determination for each of the five hold out folds during nested cross- validation.
cv_R2_avg - The average coefficient of determination for each of the hold out folds when cross-validation was performed on the entire data set.
cv_R2_sd - The standard deviation of the coefficient of determination for each of the 10 hold out folds when doing cross- validation on the entire data set.
in_sample_R2 - After an optimal lambda parameter has been chosen, cv.glmnet trains an elastic net model with all of the data provided, i.e. without any hold out folds. Expression is then predicted using this model, and the coefficient of determination is calculated.
nested_cv_fisher_pval - Our first attempt at calculating a p-value during nested cross-validation. We calculated a p-value using a correlation test using on each of the five hold out folds, and combined the five p-values using Fisher's method. Since some models yielded predictions with a variance of 0, we were unable to perform a correlation test, and the p-value we reported for that fold was a random number from the uniform distribution on the interval (0,1), which would be the correct distribution under the assumption of the null hypothesis. Under simulated null results, we still found this method did not produce a uniform p-value distribution, with values under 0.05 and over 0.95 being under-represented. We decided not to use this statistic because it did not quite produce the desired results and the random element involved when there was prior knowledge that a model did not predict well.
rho_avg - Average correlation between predicted and observed on the hold out folds when doing nested cross-validation
rho_se - Standard deviation of correlation between predicted and observed on the hold out folds when doing nested cross-validation
rho_zscore - Transformation of rho_avg into a z-score using Stouffer's Method.
pred.perf.R2 - rho_avg squared
pred.perf.pval - p-value for rho_zscore
pred.perf.qval - Deprecated. Previously held q-values calculated from the distribution of p-values, but it was later deemed this analysis was inappropriate.
This table contains info solely meant for reproducibility purposes.
chromosome - Chromosome being analyzed
cv_seed - seed that was set at start of script for randomly splitting samples into different folds
Sample Info Table
Basic info on samples used for training. Contains a single row
n_samples - Number of samples
population - The population studied
tissue - The tissue from which RNA was sequenced
Das S, Forer L, Schönherr S, Sidore C, Locke AE, Kwong A, Vrieze S, Chew EY, Levy S, McGue M, Schlessinger D, Stambolian D, Loh PR, Iacono WG, Swaroop A, Scott LJ, Cucca F, Kronenberg F, Boehnke M, Abecasis GR, Fuchsberger C. Next-generation genotype imputation service and methods. Nature Genetics 48, 1284-1287 (2016).
GTEx Pipeline https://github.com/broadinstitute/gtex-pipeline