Enzyme Promiscuity Prediction

This repository contains code used to compare various different enzyme-substrate promiscuity strategies on family-wide enzyme screening data.

Install

Packages

A python environment can be created directly using the environment.yml file included:

conda env create -f environment.yml

Once an enviornment has been activated, the package can be installed with:

python setup.py install

Featurizations

All featurizations can be handeled directly by the build_features file. Due to the cost of repeatedly using language models to featurize proteins, features are automatically cached in data/program_cache for later use if the --cache-dir argument is set in the program.

Precomputed JT-VAE embeddings were also precomputed using a forked repositoryfrom the original JT-VAE paper. These can be found in data/processed/precomputed_features/.

Dataset

The datasets used in this study and the corresponding structure reference files can be downloaded from the following github repository: https://github.com/samgoldman97/enzyme-datasets, which contains instructions for how datasets, alignments, and structure references were created and processed.

These dataset files are also included within this package directly for convenience in data/processed/.

Testing a model

Launching a simple program

A simple, example program can be executed using the following run call:

python run_scripts/run_combinations_slurm.py configs/2021_06_30_example_launch.json

This will launch an evaluation of a KNN based model that uses the levenshtein distance between enzymes sequences to make predictions about held out enzyme activity for each substrate in the esterase_binary dataset.

Running experiments

Experiments can be run using python train_model.py. Experiments can also be run from config files located in configs using the launcher scripts contained in run_scripts. Specifically, python run_scripts/run_combinations_slurm.py [config file] will launch the expriments defined in the config file, with instructions for config files contained at the top of run_combinations_slurm.py. The config files have an optional flag to run the program on a SLURM cluster for parallelization as done in the original study.

The various provided config files are detailed here:

1. configs/2021_05_25_psar_olea_hyperopt.json: Perform hyperoptimization for various model types on the OleA dataset for PSAR models that try to generalize to new enzymes.

2. configs/2021_05_25_qsar_olea_hyperopt.json: Perform hyperoptimization for various model types on the OleA dataset for QSAR models that try to generalize to new substrates.

3. configs/2021_05_27_psar_multi.json: Use the resulting hyperoptimized parameters to run PSAR analyses on all other datasets.

4. configs/2021_05_28_qsar_multi.json: Use the resulting hyperoptimized parameters to run QSAR analyses.

5. configs/2021_05_25_psar_olea_hyperopt.json: Use the resulting PSAR hyperoptimized parameters to run pooling comparison experiments in the PSAR direction.

6. configs/2021_06_30_example_launch.json: Run an example program launch

After completing a set of experiments, all the results entries from the specific experiment can be collected into a single results file using the script run_scripts/combine_csvs.py. For instance, to combine any experiments in the example launch:

python run_scripts/combine_csvs.py --results-dir results/dense/2021_06_30_example_launch -
-out-file results/dense/2021_06_30_example_launch/combined_csv.csv

These combined results files are used

Making figures

Figures can be constructed usign the scripts contained in the folder make_figs/. Assumign the proper folders. All figure scripts can be run using the command:

source make_figs/make_all_figs.sh