💻 bioinformatics

pydreg: a fast Python package for identifying active cis-regulatory elements from nascent transcription

The paper introduces pydreg, a fast and maintainable Python reimplementation of the dREG algorithm for identifying active cis-regulatory elements from nascent transcription, which significantly reduces runtime and memory usage while preserving the original method's accuracy and compatibility with modern computing infrastructure.

Original authors: He, A. Y., Danko, C. G.

Published 2026-09-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: He, A. Y., Danko, C. G.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Inside every cell of the human body, a complex system of switches determines which genes are turned on and which are turned off. These switches, known as regulatory elements, act like the control panel for the cell's operations, deciding when to build a specific protein or when to stop. To understand how a cell functions, scientists must locate these active switches. One powerful way to find them is by listening to the cell's raw, unprocessed genetic messages. When a gene is being read, the cell produces a short-lived copy of RNA called nascent RNA. By sequencing this fresh RNA, researchers can see exactly where the cell is currently active, revealing the location of promoters and enhancers that drive biological processes.

For over a decade, a computer program called dREG has been the standard tool for finding these active regions. It works by scanning the patterns of RNA production across the genome, looking for the specific signature of a switch being flipped. However, the original version of this software was built on older technology that has become difficult to run and maintain. It relied on a complex mix of programming languages and specialized hardware support that is no longer easy to update. As a result, many scientists who wanted to use this powerful method found themselves blocked by technical hurdles, unable to run the analysis on their own computers or integrate it into modern research workflows.

To solve this problem, researchers Adam Y. He and Charles G. Danko have created a new version of the software called pydreg. They rewrote the entire program in Python, a language that is widely used and easier to maintain, while keeping the exact same mathematical logic that made the original tool so accurate. The goal was not to change how the science works, but to make the tool faster, lighter, and much easier for anyone to use. The new software preserves the original "brain" of the program—the pre-trained models that recognize the patterns of active genes—but replaces the heavy, outdated machinery underneath with modern, efficient tools.

When the team tested the new software against the old version, the results were nearly identical in terms of accuracy. The new program identified the same active regions with a level of agreement so high that the two sets of results are almost indistinguishable. In fact, when comparing the scores given to millions of potential locations, the new software matched the old one with a correlation so strong it is effectively perfect. This confirms that the new code did not lose any of the scientific precision of the original method. The difference lies entirely in how the work gets done. The new version runs about four and a half times faster than the old one. It also uses significantly less computer memory, requiring only a fraction of the resources needed by the previous version.

The speed and efficiency come from how the new software handles the heavy lifting of its calculations. The original program used a custom, older method to process data on graphics cards, which was designed for hardware from more than a decade ago. The new version uses modern libraries that take full advantage of current computer chips, allowing it to perform the same complex calculations with far less effort. It also streamlines the steps taken before and after the main calculation, reducing the time the computer spends waiting for data. For researchers, this means that an analysis that might have taken hours can now be finished in a fraction of the time, and it can be run on standard equipment rather than requiring specialized, hard-to-find setups.

Beyond just speed, the new software removes the barriers that kept many scientists from using this method at all. It is packaged as a simple tool that can be installed with a single command, and it works seamlessly with other modern tools used in genetic research. The developers have also made the underlying code and the pre-trained models freely available, ensuring that the method remains accessible and can be improved by the community in the future. By updating the infrastructure without changing the science, the researchers have extended the life of a critical tool, ensuring that the ability to map the active switches of the genome remains available to everyone who needs it.

Technical Summary of "pydreg: a fast Python package for identifying active cis-regulatory elements from nascent transcription"

Problem Statement
Active promoters and enhancers (cis-regulatory elements, or CREs) generate characteristic bidirectional transcription patterns detectable via nascent RNA sequencing (e.g., GRO-seq, PRO-seq). The dREG algorithm is a leading method for identifying these active CREs by analyzing read density patterns using Support Vector Regression (SVR) and statistical peak calling. However, the reference implementation of dREG presents significant usability and maintenance barriers. It relies on a complex workflow combining R, compiled C code, shell scripts, and legacy external utilities. Crucially, its primary GPU acceleration library, Rgtsvm, depends on a source build against CUDA and Boost, targets early-generation NVIDIA hardware, and utilizes sparse-workload assumptions that limit performance on modern GPUs. These factors make deployment difficult, hinder maintenance, and prevent the method from leveraging contemporary hardware advancements.

Methodology
To address these limitations, the authors developed pydreg, a Python port of the dREG pipeline that preserves the original algorithmic logic and pretrained models while modernizing the computational infrastructure.

Pipeline Architecture: pydreg processes strand-specific nascent transcription coverage (bigWig files) through a five-step pipeline:
- Identification of genomic positions with sufficient read coverage.
- Summarization of local read profiles into multi-scale feature vectors (length 360).
- Scoring of feature vectors using the original pretrained SVR model.
- Assembly of high-scoring positions into candidate peaks, refined by a pretrained Random Forest splitting model.
- Statistical significance assignment and False Discovery Rate (FDR) filtering.
- The pretrained model parameters are converted to portable safetensors files, automatically downloaded and cached upon first use.
SVR Inference Optimization: The core computational bottleneck is the SVR inference, which involves calculating pairwise distances between query vectors and 605,187 support vectors using a Radial Basis Function (RBF) kernel.
- Legacy Approach: The original Rgtsvm library used custom sparse-oriented CUDA kernels, which resulted in poor data reuse and high memory traffic for dREG's dense 360-dimensional features.
- pydreg Approach: The authors reimplemented the scoring using CuPy (for NVIDIA GPUs) and MLX (for Apple Silicon), with a CPU fallback using NumPy and Numba. They reformulated the RBF distance calculation using the identity $\|x - s\|^2 = \|x\|^2 + \|s\|^2 - 2x^Ts$ . This allows the dominant computation to be performed via vendor-optimized General Matrix Multiplication (GEMM) kernels. Precomputed row norms and fused elementwise kernels further reduce redundant arithmetic and device memory traffic.
Statistical Peak Calling Optimization: The calculation of p-values for peak calling involves integrating over a correlated five-dimensional Laplace null distribution using randomized quasi-Monte Carlo (QMC) methods.
- Optimization: pydreg skips zero-weight grid points and terminates integration early when the remaining tail's maximum possible contribution falls below $10^{-6}$ . These changes eliminate unnecessary QMC evaluations without degrading statistical accuracy.

Key Results
The authors evaluated pydreg against the original dREG on an NVIDIA Titan Xp GPU with 16 CPU threads using 12 previously published nascent RNA sequencing datasets.

Accuracy: pydreg produces nearly identical results to dREG.
- SVR scores show a Pearson correlation $r > 0.999$ and a Mean Absolute Error (MAE) of 0.00251.
- Final peak calls agree with a Jaccard index $> 0.997$ .
Performance:
- Runtime: pydreg is approximately 4.5 times faster than dREG.
- Memory: pydreg reduces peak host memory (RSS) by a median of 5.4-fold (ranging from 3.8 to 6.8-fold).
- GPU Efficiency: Profiling reveals that pydreg performs 12 times more work per kernel launch and 942 times more work per host-to-device memory copy compared to dREG. Total scoring time is 5.5 times faster, with GPU-active compute time reduced by 4.2 times.
Usability: pydreg is distributed as a pip-installable package (pydreg[gpu] or pydreg[mlx]) with both command-line and programmatic Python interfaces, integrating seamlessly into modern Python-based genomics workflows.

Significance and Claims
The paper claims that pydreg successfully extends the useful life of the dREG algorithm by removing the dependency on legacy, difficult-to-maintain infrastructure. By adopting modern numerical libraries and packaging, the authors assert that pydreg:

Improves Reproducibility and Interoperability: It allows dREG to run in standard Python environments, facilitating integration with contemporary analysis pipelines.
Reduces Computational Barriers: The significant reductions in runtime and memory usage make running dREG locally more practical for a broader range of users.
Leverages Modern Hardware: Unlike the legacy Rgtsvm, pydreg is optimized for current GPU architectures (NVIDIA and Apple Silicon), ensuring the method remains efficient as hardware evolves.

The authors conclude that pydreg provides a maintainable foundation for future development while preserving the high accuracy of the original dREG method.

Technical Summary of "pydreg: a fast Python package for identifying active cis-regulatory elements from nascent transcription"

More like this