dREG pretrained models

Pretrained weights for dREG (Danko Lab), a method for detecting active transcriptional regulatory elements (promoters and enhancers) from PRO-seq/GRO-seq nascent-transcription data.

These are the original dREG model parameters, extracted from the R package's distributed .RData/.RDS files and repacked as framework-agnostic safetensors -- no retraining was performed. See config.json for the full parameter listing.

Used by pydreg, a from-scratch Python port of dREG's inference pipeline. DREGModel.from_pretrained() / DREGPeakSplitForest.from_pretrained() download and load these directly.

Files

file	model	description
svm.model.safetensors.zst	dREG SVR	RBF-kernel epsilon-SVR that scores a genomic position's regulatory potential (~[0, 1]) from a 360-dim multi-scale feature vector. 605,187 support vectors.
rf.model.safetensors.zst	peak-split forest	500-tree random forest regression model used only during peak calling, to decide whether two adjacent local score maxima should be merged into one peak or split into two.
config.json	both	Human-readable model configuration (mirrors the metadata embedded in each safetensors file's header).

Both .safetensors.zst files are zstandard-compressed safetensors -- pydreg's loader decompresses them transparently, or use zstd -d manually.

Provenance

• SVR -- extracted from asvm.gdm.6.6M.20170828.rdata (Zenodo), the pretrained model bundled with dREG. The saved object (class gtsvm, trained via Rgtsvm) is field-compatible with e1071's standard RBF epsilon-SVR dual form -- no re-training, only a format conversion.
• Peak-split forest -- extracted from rf-model-201803.RDS, bundled with the dREG R package (dREG/inst/extdata/), a randomForest regression forest used only in peak_calling_rf.R's find_rf_peaks()/split_peak().

Both files contain the exact weights the original R package ships and uses at inference time -- this repo hosts a format conversion, not a retrained or fine-tuned model.

Usage

from pydreg.models import DREGModel, DREGPeakSplitForest

svr = DREGModel.from_pretrained()            # downloads svm.model.safetensors.zst
rf = DREGPeakSplitForest.from_pretrained()   # downloads rf.model.safetensors.zst

scores = svr.predict(X)   # X: (n_queries, 360) features from pydreg.features

See pydreg's documentation for the full pipeline: informative-position scanning -> multi-scale feature extraction -> SVR scoring -> RF-assisted peak calling -> FDR filtering.

Citation

If you use these models, please cite the original dREG papers:

Danko, C. G., Hyland, S. L., Core, L. J., Martins, A. L., Waters, C. T., Lee, H. W., Baranello, L., Yang, Z., Wong, S. E., Setola, V., Lee, S. K., ... & Siepel, A. (2015). Identification of active transcriptional regulatory elements from GRO-seq data. Nature Methods, 12(5), 433-438. https://doi.org/10.1038/nmeth.3329

Wang, Z., Chu, T., Choate, L. A., & Danko, C. G. (2018). Identification of regulatory elements from nascent transcription using dREG. bioRxiv, 321539. https://doi.org/10.1101/321539

License

GPL-3.0, matching the original dREG R package (License: GPL-3 in its DESCRIPTION). These files are a format conversion of dREG's publicly distributed pretrained model weights, not a redistribution of its source code, but are licensed the same way as the rest of this port.