Calculates average contribution scores across multiple folds output from the ChromBPNet pipeline
A typical ChromBPNet workflow for a single dataset involves: defining five folds with different combinations of training, test and validation chromosomes; training a model and calculating contribution scores (shap values) for each of these folds; averaging the shap values; and using these averaged shap values to identify transcription factor motifs with TFModisco. In addition to averaging the shap values, this script optionally produces quality control matrix scatter plots for quickly comparing shap values between folds -- values should be correlated but not identical.
Produces json files with necessary metadata for uploading ChromBPNet models to the ENCODE portal
ChromBPNet model datasets that are uploaded to the ENCODE portal
include dozens of files; in addition to an h5 and tar version of each model we've trained, we include regions
bed files of train, test and validation data, numerous log files, bed files of background regions, a bigwig of the
signal track, and others. This script automates the production of a json with metadata of file locations that is then
ingested by another script, which produces tar files that are
then uploaded to the ENCODE portal.