search

churchlab / low-N-protein-engineering

Code and data to reproduce analyses in Biswas et al. (2020) "Low-N protein engineering with data-efficient deep learning".
56 stars 18 forks source link

readme

low-N-protein-engineering

Code to reproduce analyses in Biswas, Khimulya, Alley et al (2021). Nature Methods.

1. Set up a Docker environment

We use docker to provide a consistent environment and facilitate reprodubility. First build an image:

To use GPU acceleration:

$ docker build -f docker/Dockerfile.gpu -t low-n-gpu .

To use CPU-only: 

$ docker build -f docker/Dockerfile.cpu -t low-n-cpu .

Note you will need a GPU in order to completely reproduce the results, especially where UniRep inference is required. The above has been tested on p3.2xlarge (which have a NVIDIA V100) AWS instances running Deep Learning AMI (Ubuntu 18.04) Version 30.0 - ami-029510cec6d69f121.

2. Dig in!

Here's the repository layout:

  1. In analysis you'll find the code needed to reproduce the analyses in the paper. Every sub-directory within analysis contains a README for what analyses and figures are covered. Sorry in advance for the sometimes cryptic subdirectory names! We have a few instances of (deeply) hardcoded paths in the code. We didn't change these in order to minimize changes to the original code we wrote, figuring this would help with reproducibility.

  2. docker contains files needed to build the docker environment (more instructions below) needed to reproduce analyses and figures.

  3. requirements contains the requirements that are pulled into the docker environment.