RFDesign: Protein hallucination and inpainting with RoseTTAFold

Jue Wang ([email protected])
Doug Tischer ([email protected])
Sidney Lisanza ([email protected])
David Juergens ([email protected])
Joe Watson ([email protected])

This repository contains code for protein hallucination or inpainting, as
described in our
. Code
for postprocessing and analysis scripts included in scripts/.


All code is released under the MIT license.

All weights for neural networks are released for non-commercial use only under the Rosetta-DL license.


  1. Clone the repository:

    git clone https://git.ipd.uw.edu/jue/rfdesign.git
    cd rfdesign
  1. Create environment and install dependencies:

    cd envs
    conda env create -f SE3.yml
  1. Download model weights (see license info above).

    wget https://files.ipd.uw.edu/pub/rfdesign/weights.tar.gz
    tar xzf weights.tar.gz
  1. Configure path to weights. Put a file called config.json in hallucination/ and
    inpainting/ with the path to the weights directory. An example file is in each
    folder to copy from.


If you want/need to configure your environment manually, here are the packages in our environment:


  • If you are running this on digs at the IPD, you don’t need to do steps 3-4.
  • If you are getting output pdbs that are a ball of disconnected segments (as viewed in pymol), this may be due to a problem with the spherical harmonics cached by SE3-transformer. A workaround is to copy the hallucination/cache/ folder (a correct, clean copy of the cache) to your working directory before running hallucinate.py or inpaint.py.


See READMEs in hallucination/ and inpainting/ subfolders.


J. Wang, S. Lisanza, D. Juergens, D. Tischer, et al. Deep learning methods for designing proteins scaffolding functional sites. bioRxiv (2021). link

M. Baek, et al., Accurate prediction of protein structures and interactions using a three-track neural network, Science (2021). link

An earlier version of our hallucination method can be found at the trdesign-motif repo and published at:

D. Tischer, S. Lisanza, J. Wang, R. Dong, I. Anishchenko, L. F. Milles, S. Ovchinnikov, D. Baker. Design of proteins presenting discontinuous functional sites using deep learning. (2020) bioRxiv link

Our work is based on previous hallucination methods for unconstrained protein generation and fixed-backbone sequence design (trDesign repo):

I Anishchenko, SJ Pellock, TM Chidyausiku, …, S Ovchinnikov, D Baker. De novo protein design by deep network hallucination. (2021) Nature link

C Norn, B Wicky, D Juergens, S Liu, D Kim, B Koepnick, I Anishchenko, Foldit Players, D Baker, S Ovchinnikov. Protein sequence design by conformational landscape optimization. (2021) PNAS link


View Github