RLoopBench

RLoopBench is a benchmark for evaluating whether DNA foundation models generalize to R-loop-forming sequence prediction. The benchmark compares rule-based features, classical sequence encodings, task-specific deep learning models, and DNA foundation model embeddings under a unified linear-probe evaluation framework.

R-loops are three-stranded nucleic acid structures consisting of an RNA--DNA hybrid and a displaced single-stranded DNA strand. Because R-loop formation is associated with transcription, replication stress, genome instability, and disease-related genome dysfunction, R-loop prediction provides a biologically distinct test case beyond conventional gene regulatory benchmarks.

Repository overview

This repository provides:

• Example linear-probe training code using 3-mer encoding
• Pretrained linear-probe classifiers for each representation
• Inference scripts for applying trained linear probes

Methods included

RLoopBench evaluates the following representation paradigms.

Rule-based method

• QmRLFS-finder

Please refer to http://r-loop.org/?pg=qmrlfs for details.

Classical sequence representations

• 3-mer frequency
• 4-mer frequency
• One-hot encoding

Trained models and inference scripts are provided in models/. These scripts can be run using the conda environment specified in environment.yml.

conda env create -f environment.yml
conda activate rloop_bench

We also provide an example workflow in training/ showing how to generate 3-mer embeddings, visualize embeddings, and train a linear-probe classifier.

Task-specific deep learning models

• DeepER: https://github.com/NjuChenlab/DeepER
• deepRloopPre: https://github.com/PEHGP/deepRloopPre

Please refer to the original repositories for installation and usage details.

DNA foundation model embeddings

• Evo2: https://github.com/arcinstitute/evo2
• Nucleotide Transformer v3: https://huggingface.co/InstaDeepAI/NTv3_650M_pre
• DNABERT-2: https://github.com/MAGICS-LAB/DNABERT_2

The trained linear-probe models and reference inference scripts for these representations are provided in models/.

Please use the foundation-model inference scripts together with each DNA foundation model's own implementation environment. The scripts in this repository are intended as reference implementations for reproducing the embedding-to-linear-probe inference step.