Awesome
Deep learning enables therapeutic antibody optimization in mammalian cells by deciphering high-dimensional protein sequence space
Overview
This repository contains the scripts to perform therapeutic antibody optimization with deep learning, as described in Mason et al., 2019 [1]. The Python scripts are found in folder scripts/. The resulting in silico predicted binders are then subjected to multiple developability filters in the R script developability_filters.R.
Table of contents
Prepare working environment
Before running any of the scripts, the necessary packages need to be installed. This is done with Conda, the open source package management system [2], and the environment can be loaded via the provided config.yaml file, using following commands:
conda env create -f environment.yml
source activate dms_env
Usage
Deep learning: Python scripts
The full deep learning analysis, written in Python, can be summarized into three consecutive steps:
- Compare classification performance of different machine learning models on rationally designed site-directed mutagenesis libraries.
- The best-performing model, a convolutional neural network (CNN), is tuned with a randomized search on hyper parameters.
- The optimized CNN model is used to identify antigen-binding sequences from an in-silico generated library.
Steps 1 and 3 are performed simultaneously, by running the following command:
python scripts/main.py
This will produce visualizations of the model performances, and create a CSV file with the predicted values for the in-silico generated library.
Inside the main script, hyper parameters were already selected according to the optimized model. However, the randomized search can be run again:
python scripts/model_tuning.py
A folder model_tuning/
will be created, containing the best hyper paramter settings and the corresponding mean cross-validated score. Those parameters can then be included inside the main script (params
).
Applying developability filters
With the results from the previous analysis, developability filters can be applied to the in silico generated library:
R --vanilla < developability-filters.R
The developability filter based on CamSol solubility scores ([3]-[4]) needs to be run on their web server. Additionally, netMHCIIpan [5], version 3.2, needs to be downloaded and installed under following link, otherwise the R script will terminate.
License
Citation
If you use the the code in this repository for your research, please cite our paper.
@article {Mason617860,
author = {Mason, Derek M and Friedensohn, Simon and Weber, C{\'e}dric R and Jordi, Christian and Wagner, Bastian and Meng, Simon and Gainza, Pablo and Correia, Bruno E and Reddy, Sai T},
title = {Deep learning enables therapeutic antibody optimization in mammalian cells by deciphering high-dimensional protein sequence space},
year = {2019},
doi = {10.1101/617860},
URL = {https://www.biorxiv.org/content/early/2019/05/30/617860},
journal = {bioRxiv}
}
Authors
Derek Mason, Jan Dahinden, Simon Friedensohn, Bastian Wagner, Cédric Weber, Sai Reddy