Neural network extrapolation to distant regions of the protein fitness landscape

Installation

The code in this module is built on the nn4dms machine learning models and module. This module has the same requirements as nn4dms, with a few additional dependencies. Use conda to set up an environment from env.yml

The model code is in the linked repo nn4dms. This code is required to reproduce some of our inference results. Code and models to generate model predictions are included as the nn4dms_nn-extrapolate andnn-extrapolate-models submodules. Download times may be long if cloning submodules. We include download commands for nn-extrapolation with and without submodules.

# clone repo without submodules
gh repo clone RomeroLab/nn-extrapolation

# checkout submodules individually
git submodule update nn4dms_nn-extrapolate
git submodule update nn-extrapolation-models

# clone full repo including submodules
gh repo clone RomeroLab/nn-extrapolation -- --recurse-submodules

We provide two conda environments to reproduce our analysis. gb1_inf is used to run all Makefile commands. gb1_notebook is used to reproduce data anlysis in the Jupyter notebooks.

# download and install environments
conda env create -f env_inference.yml
conda env create -f env_notebook.yml

# activate conda environments
conda activate gb1_inf
conda activate gb1_notebook

Installation takes ~5 minutes.

Software Requirements

This module has been tested on the following systems:

• macOS: Mojave (10.14.6)
• Linux: Green Obsidian (8.8)

Data Analysis

Extrapolating learned protein fitness lanscapes

Generate predictions for the Wu et al. 1-4 mutant GB1 fitness dataset.

python 01_extrapolation_predictions.py

Generate extrapolation trajectories.

python 01_extrapolation_trajectories.py

Generate plots for Fig 1 and Fig S1 in 01_extrapolation_analysis.ipynb

ML-guided protein design for deep exploration of the fitness landscape

Design sequences using 02_run_sa.py. See example below. Each design can take minutes to hours, depending on the model; this can be accelerated by running on a GPU.

python 02_run_sa.py data/config_example.txt

Generate plots for Fig 2 and Fig S2-3 in 02_designs_analysis.ipynb

Large-scale experimental characterization of ML designed GB1 variants

From the raw fastq files, preprocess the results and determine counts for each variant in the library. (fastq files will be downloadable from the SRA; save the fastq files in a directory fastq_files). This step can take hours to days.

mkdir merged_reads
for d in fastq_files/ ; do
python 03_preprocessing.py fastq_files/${d} ${d:0:6} merged_reads/ designs.csv designs_counts.csv

Generate plots for Fig 3 and Fig S4-8 in 03_design_experimental_analysis.ipynb

ML designed GB1s show improved display and IgG binding

Generate plots for Fig 4 and Fig S9 in 03_design_experimental_analysis.ipynb