adamyhe/pydreg-supporting-data

pydreg vs. dREG benchmark outputs

Raw benchmark artifacts backing the performance and accuracy comparisons in pydreg, a from-scratch Python port of dREG (Danko Lab). This dataset holds the paired outputs of running both tools' full peak-calling pipeline (run_dREG/pydreg) on the same 12 real PRO-seq/GRO-seq/ChRO-seq libraries, plus the /usr/bin/time -v logs used to compare wall-clock time and peak memory. It is data, not code — see the pydreg repo for the package itself and for the scripts (figures/timing_scripts_*.sh) that produced everything here.

Dataset structure

Two top-level directories, one per tool, each with one 8-file group per library:

{tool}/{library}.dREG.infp.bw          # raw SVR score at every informative position, genome-wide
{tool}/{library}.dREG.raw.peak.bed.gz  # broad candidate peaks, before RF-assisted splitting
{tool}/{library}.dREG.peak.full.bed.gz # final called peaks, all columns (coords, score, prob, ...)
{tool}/{library}.dREG.peak.score.bed.gz
{tool}/{library}.dREG.peak.score.bw    # final called peaks, score column only
{tool}/{library}.dREG.peak.prob.bed.gz
{tool}/{library}.dREG.peak.prob.bw     # final called peaks, prob column only (1 - FDR-style p-value)
{tool}/{library}.time.log              # `/usr/bin/time -v` output for that run

{tool} is dreg (original R dREG) or pydreg. Filenames and the .dREG.-infixed suffix convention match pydreg's own CLI output exactly, so every {tool}/{library}.dREG.* pair is a direct, position-for-position comparison of the two tools on identical input.

Libraries

All 12 are published PRO-seq/GRO-seq/ChRO-seq libraries from GEO:

Library	GEO accession	Biosample	Assay	Source
G1	GSM1480327	K562	PRO-seq	Core et al., Nat Genet 2014
G3	GSM3452725	K562	PRO-seq	Wang et al., Genome Res 2019
G5	GSE89230	K562	PRO-seq	Vihervaara et al., Nat Commun 2017
G6	GSM2545324	K562	PRO-seq	Dukler et al., Genome Res 2017
G7	GSM2545325	K562	PRO-seq	Dukler et al., Genome Res 2017
GM12878_groseq	GSM1480326	GM12878	GRO-seq	Core et al., Nat Genet 2014
K562_groseq	GSM1480325	K562	GRO-seq	Core et al., Nat Genet 2014
Jurkat_PROseq	GSM3309955	Jurkat	PRO-seq	Chu et al., Nat Genet 2018
Jurkat_ChROseq_1	GSM3309957	Jurkat	ChRO-seq	Chu et al., Nat Genet 2018
Jurkat_ChROseq_2	GSM3309956	Jurkat	ChRO-seq	Chu et al., Nat Genet 2018
Jurkat_ChROseq	GSM3309956 + GSM3309957 (pooled)	Jurkat	ChRO-seq	Chu et al., Nat Genet 2018
Jurkat_leChROseq	GSM3309958	Jurkat	leChRO-seq	Chu et al., Nat Genet 2018

How this was produced

• dREG: the original R package (run_dREG.R), scored with its 2017 SVR model (asvm.gdm.6.6M.20170828.rdata). The runs archived in this dataset were all generated against a raw, directly-installed R/CUDA/Rgtsvm stack (a bare run_dREG.bsh call), not the container described below.
• pydreg: this repo's Python port, same model weights (converted to safetensors, see adamyhe/pydreg on the Hub), run with --cores 16.
• Both tools ran on the same machine (NVIDIA Titan Xp GPU + 16 CPU cores), same bigWig inputs, one library at a time, wrapped in /usr/bin/time -v for wall-clock and peak-RSS logging.

Exact invocations: figures/timing_scripts_download.sh (fetches/rebuilds the 12 input bigWigs from GEO) and figures/timing_scripts_pydreg_only.sh. For dREG, the repo's own script instructions now favor figures/timing_scripts_dreg_apptainer.sh (runs dREG inside the Danko-Lab/dREG-apptainer image) for reproducibility going forward: dREG's R/CUDA/Rgtsvm dependency stack is fragile and version-sensitive to install by hand, which containerizing avoids. The runs archived here predate that script and were produced with an equivalent direct, non-containerized run_dREG.bsh call instead; a fresh reproduction of this dataset should use the apptainer script.

Consumed by figures/plot_walltime.py, plot_memory.py, plot_score_exactness.py, and plot_peak_agreement.py to produce the timing note figures, via the shared fetch helper in figures/_common.py.

License

GPL-3.0, matching the pydreg source license these benchmark outputs support.

Citation

If you use this data, please cite pydreg, dREG, and the original data sources listed in the table above:

@article{he2026pydreg,
  author  = {He, Adam Youlin and Danko, Charles G.},
  title   = {pydreg: a fast Python package for identifying active cis-regulatory elements from nascent transcription},
  journal = {bioRxiv},
  year    = {2026},
  doi     = {10.64898/2026.09.06.745329}
}

@article{wang2019dreg,
  author  = {Wang, Zhong and Chu, Tinyi and Choate, Lauren A. and Danko, Charles G.},
  title   = {Identification of regulatory elements from nascent transcription using dREG},
  journal = {Genome Research},
  year    = {2019},
  volume  = {29},
  number  = {2},
  pages   = {293--303},
  doi     = {10.1101/gr.238279.118}
}