Skip to content

Latest commit

 

History

History
53 lines (43 loc) · 1.86 KB

README.md

File metadata and controls

53 lines (43 loc) · 1.86 KB
Raw

Train/test prediction models: GEUVADIS constituent pops

Use elastic net to train and test predictive models with GEUVADIS data. Unlike the previous analysis step, this analysis partitions samples into five constituent 1000 Genomes populations:

  • CEU, the Central Europeans from Utah
  • GBR, British from Great Britain
  • TSI, Tuscans from Italy
  • FIN, Finnish from Finland
  • YRI, Yoruba from Nigeria

All samples are unrelated; the children in the GEUVADIS dataset are excluded. Each population is subsampled to the size of the smallest population (YRI, n = 89). Prediction models are trained in one population and tested in all five populations.

Prerequisites

  • Rscript
  • qsub

The following R packages are used:

  • glmnet
  • dplyr
  • data.table
  • broom
  • optparse
  • dunn.test
  • ggplot2

Install there from CRAN using any standard installation procedure. One way is to type

install.packages(c("glmnet", "data.table", "dplyr", "optparse", "broom", "dunn.test", "ggplot2"))

from within the R console.

Running

From the command line in an SGE environment, execute

./test_prediction_models_crosspop.sh

This will schedule jobs with the correct ordering and hold patterns. Jobs for each train-test case are broken into prediction, collection, and postprocessing. Since each training population is independent, all five training populations can be run simultaneously.

Once all results are available, they can be analyzed by running

./run_geuvadis_compile_prediction_results_onepop.sh

which will compile results for each training population individually and then concatenate them together.

Notes

Similar to the previous analysis step, this analysis requires substantial disc space (> 100Gb). Delete log files as needed once correct execution is confirmed.