dtischer / trdesign-motif

Codebase for our preprint using trRosetta to design proteins with discontinuous functional sites, found here: https://www.biorxiv.org/content/10.1101/2020.11.29.402743v1.abstract

TrRosetta-based protein hallucination

2021-10-18
Doug Tischer (dtischer at uw.edu)
Jue Wang (jue at post.harvard.edu)
Sidney Lisanza (lisanza at uw.edu)

This repository contains scripts for performing protein design by gradient descent through the structure-prediction neural network TrRosetta. The method is similar to that of trDesign except our focus is generating scaffolds for functional motifs.

This code accompanies the paper:

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, doi:10.1101/2020.07.22.211482.

Requirements

• tensorflow (tested on 1.14)
• scwrl4 (optional, for sidechains. download)
• pyrosetta (2020.10+release.46415fa, for obtaining structural models. see pyrosetta.org)

Installation & Usage

Clone git repository:

git clone https://github.com/dtischer/trdesign-motif.git

Run examples:

cd example
./run_example.sh        # generates scaffold for PD-1 interface motif versus PD-L1
./run_example2.sh       # same task as above, but using a different trRosetta version (see below)

Fold design model from predicted pairwise distances & angles (first run the example above, then run this in the example subfolder):

sequence_design/fold.sh output/

Additional downstream analyses can be done using the scripts in scoring/.

Contents

hallucination: contains hallucination script design.py. see code for full list of command-line options.

hallucination_grid: an alternative hallucination script using a more recent version of trRosetta. This version is slightly more accurate at structure prediction, and also predicts the probability that some binned 3D position is occupied by some other residue, in the reference frame of each residue.

sequence_design: scripts for folding design models from hallucinated pairwise distances and angles (fold.sh) and for using Rosetta Fastdesign to design better sequences onto the hallucinated backbones.

scoring: various scripts to compute metrics on the designs generated by the hallucination script.

Links

• trDesign - free hallucination and fully constrained fixed-backbone sequence design using trRosetta
• trRosetta - protein structure prediction with a convolutional neural network