This repo contains the code for reproducing the results of the publication "Self-supervised machine learning methods for protein design improve sampling, but not the identification of high-fitness variants" (link to preprint). If you are interested instead in using the implemented features for your own work, an overview of them can be found in the Rosetta documentation here, and a tutorial is available from the Meiler Rosetta workshop 2023 "Tutorial 2: Machine Learning in Rosetta".

Code for running the different design protocols can be found in the folder of each dataset, e.g. emi/avg03.sh. All scripts use the RosettaScripts XML provided in the main folder which are named after the different protocols shown in the paper.

The unique sequences and calculated metrics of each design protocol are available in the dataset folders ("dataset/dataset_designs.csv"), e.g. emi/emi_designs.csv.

The code for analyzing the resulting designs and reproducing figures can be found in the design_analysis.ipynb notebook. In order to run the jupyter notebooks, first create a python environment using the environment.yaml file with either conda or mamba:

# create environment
conda env create -f environment.yaml
# activate environment
conda activate probs_design

The code for training and evaluating the oracle models for each dataset can be found in the model_training.ipynb notebook. The datasets used for training can be found in each dataset folder, e.g. gb1/gb1_mutations_full_data.csv. The already trained models are also available, e.g. gb1/gb1_ridge.joblib.

The Rosetta source code can be found at https://github.com/RosettaCommons/rosetta/. Docker containers for Rosetta (including the Tensorflow/LibTorch extras version) can be found at https://hub.docker.com/r/rosettacommons/rosetta.