Dissection of core promoter syntax through single nucleotide resolution modeling of transcription initiation (CLIPNET data)

This contains data necessary to reproduce the figures in the CLIPNET paper (preprint here) as well as processed data used to train and evaluate CLIPNET. To preserve subdirectory structure, we've packaged the data into tar archives. Please refer to the README documents in our manuscript GitHub repo for more details on file contents: https://github.com/Danko-Lab/clipnet_paper/

 

Pretrained CLIPNET models are archived separately at DOI 10.5281/zenodo.10408622

 

V5: Fixed bug in calculation of profile attribution scores causing them to be off by a factor of exactly 500. Genome-wide DeepSHAP tracks & TF-MoDISco tracks have been accordingly updated. I have not updated the individual examples, as these can be quickly fixed by simply multiplying by 500 when plotting. Additionally, I have uploaded profile and quantity motif calls, which contain genome-wide seqlet annotations. The columns in these files are [chrom, start, end, peak_idx, motif_annotation].

V4: Uploaded individual bigWigs. These have been lifted over using CrossMap from the original hg19 (GSE110638) to hg38 and RPM normalized.

V3: Final version prior to journal submission. Don't recall exact details of what's changed.

V2: evaluation_metrics.tar.gz and evaluation_data.tar.gz have been replaced. Previously, we benchmarked the models by treating each peak in each individual as a separate data point. Here, we instead predicted from the reference genome and compared against the averaged bigWigs.