RoseTTAFoldNA

open_source

RoseTTAFold2NA is an open-source deep learning model for predicting 3D structures of protein-DNA and protein-RNA complexes, developed by UW IPD.

AI Models & Infrastructure

Research & Education

AI Research Tools

About

RoseTTAFold2NA (RF2NA) is an advanced open-source deep learning model for predicting the three-dimensional structures of protein and nucleic acid complexes. Developed by the Institute for Protein Design (IPD) at the University of Washington, it extends the powerful RoseTTAFold2 architecture to support interactions between proteins, RNA, and DNA molecules—an essential capability for understanding biological systems at a molecular level. The tool leverages SE(3)-Transformer-based neural networks trained on known structural data to generate accurate structural predictions, even for complex systems involving homodimer–DNA interactions and DNA-specific sequence recognition. RF2NA also supports paired protein/RNA multiple sequence alignments (MSAs) for improved prediction quality. Designed for computational biologists, structural bioinformaticians, and researchers in drug discovery, RF2NA is distributed as a fully open-source MIT-licensed repository. Users install it via conda, download pretrained model weights, and run predictions through a command-line shell script. The April 2023 v0.2 update delivered improved homodimer:DNA interaction prediction, enhanced DNA-specific sequence recognition, and bug fixes in the MSA generation pipeline. RF2NA is particularly valuable for scientists studying gene regulation, CRISPR mechanisms, RNA-binding proteins, and nucleic-acid-targeted drug design—areas where understanding protein–DNA/RNA complex geometry is critical. As a research-grade tool requiring GPU hardware and bioinformatics expertise, it is best suited for academic labs and computational biology teams.

Key Features

Protein–Nucleic Acid Complex Prediction: Predicts the 3D structure of complexes involving proteins paired with DNA or RNA, covering a wide range of biologically important interactions.
SE(3)-Transformer Architecture: Leverages NVIDIA SE(3)-equivariant Transformer networks to model the geometry of molecular structures with high accuracy.
Paired MSA Support: Supports paired protein/RNA multiple sequence alignments (MSAs), improving prediction quality for RNA-binding protein complexes.
Homodimer–DNA Interaction Modeling: v0.2 introduced significantly improved predictions for homodimer:DNA interactions and DNA-specific sequence recognition.
Open-Source & Freely Available Weights: Fully open-source under the MIT license, with pretrained model weights downloadable for local research use.

Use Cases

Predicting the 3D structure of CRISPR-Cas protein complexes bound to guide RNA for gene-editing research.
Modeling transcription factor–DNA binding geometries to understand gene regulatory mechanisms.
Studying RNA-binding protein structures for RNA therapeutics and vaccine design.
Supporting structure-based drug discovery targeting nucleic acid–protein interaction interfaces.
Academic research in structural biology requiring high-accuracy protein–nucleic acid complex modeling.

Pros

Cutting-Edge Research Tool: Developed by the world-renowned UW Institute for Protein Design, ensuring state-of-the-art accuracy in protein-nucleic acid structure prediction.
Broad Molecular Coverage: Handles protein–DNA and protein–RNA complexes in a single framework, reducing the need for multiple specialized tools.
Completely Free and Open Source: MIT-licensed with freely available pretrained weights, enabling academic labs to use it without cost or licensing barriers.

Cons

Requires NVIDIA GPU Hardware: The model relies on NVIDIA GPU acceleration and is only supported on Linux, limiting accessibility for users without suitable hardware.
Complex Installation Process: Setup involves conda environments, manual weight downloads, and SE3Transformer installation, which may be challenging for non-technical users.
Research-Grade Maturity: As an academic research tool, it lacks a graphical interface, commercial support, or user-friendly documentation aimed at non-expert users.

Frequently Asked Questions

RoseTTAFold2NA can predict the 3D structures of protein–DNA complexes, protein–RNA complexes, and mixed protein/nucleic acid assemblies, including homodimer–DNA interactions.

Yes, it is fully open-source under the MIT license. The code and pretrained model weights are freely available from the GitHub repository and UW IPD servers.

RoseTTAFold2NA requires a Linux system with an NVIDIA GPU. Installation is managed via conda using the provided RF2na-linux.yml environment file.

After installing the conda environment and downloading the model weights, you run predictions using the provided `run_RF2NA.sh` shell script from the command line.

The April 2023 v0.2 update includes updated model weights for better homodimer:DNA interaction prediction, improved DNA-specific sequence recognition, bug fixes in the MSA generation pipeline, and support for paired protein/RNA MSAs.