TFBlearner data

Data accompanying the TFBlearner manuscript.

This includes:

• TFBlearner MultiAssayExperiment object (necessary for training and prediction). This is saved as a R serialized object ( mea_object.rds ), along with its underlying HDF5-backed assays ( *mapped.h5 files). Some files needed to reconstruct the object from scratch are also included (the sequence embeddings in sequenceEmbeddings.rds, the coordinates of the Putative Regulatory Elements in PERs.bed.gz ).
• The (shifted) ATAC fragments overlapping PREs across all (training and prediction) cellular contexts, prepared as (serialized) GRanges R objects (one per context) for speed, in fragments.tar. This is necessary for training, as well as for prediction on one of those contexts.
• The motif matches nearby PREs in motifs.tar (one GRanges rds file per motif), needed for training and prediction.
• The top interactors per TF (used for training and predictions).
• The binding predictions for all TFs across the 43 prediction cell types (see below).
• The GAM models to infer replicability from peaks ( reproducibilityGAM.rds ).
• A table of TF metadata (TF_annotation.tsv) collected from various sources (not needed for the workflow).
• Large supplementary tables listing all datasets used ( Supplementary_tables_* ).

Binding predictions

The binding predictions (in all_predictions_h5.tar.* ) are h5 files, one per TF, each containing two objects: 1) 'predictions' is a matrix of binding probabilities (with PRE regions as rows and cellular contexts as columns), and 2) 'contexts' is a vector of context codes, corresponding to columns of the 'predictions' matrix. The contexts/columns are the same across all prediction files, and are described in 'preds_h5_context_order.tsv'.
To untar the predictions, simply do :

cat all_predictions_h5.tar.0* | tar -xzf -

Note that binding predictions are also browsable (and downloadable on TF at a time) using the web platform www.ethz-ins.org/TFBPlatform .