LynD-substrate-modeling

This repository contains the code, model weights, and raw data for our paper on LynD substrate specificity (Steude et al., 2024).

Citation

If this code or model proves useful for you, please include the following citation in your work:

@article {Steude2024.10.14.618330,
	author = {Steude, Emma G. and Dieckhaus, Henry and Pelton, Jarrett M. and Kuhlman, Brian and Bowers, Albert A.},
	title = {Assessing substrate scope of the cyclodehydratase LynD by mRNA display-enabled machine learning models},
	year = {2024},
	doi = {10.1101/2024.10.14.618330},
	publisher = {Cold Spring Harbor Laboratory},
	URL = {https://www.biorxiv.org/content/early/2024/10/14/2024.10.14.618330},
	eprint = {https://www.biorxiv.org/content/early/2024/10/14/2024.10.14.618330.full.pdf},
	journal = {bioRxiv}
}

Note: While the model training and nomination code is original to this repo, we make heavy use of analysis and visualization functions from the LazDEF analysis repository by Vinogradov et al. We have attempted to note where these tools are used/adapted throughout the source code. We thank the authors of this work for making these useful tools open-source and well-organized.

Repo Contents:

datasets/: Preprocessed sequencing data for selection rounds 1-3 in CSV format.
src/model_weights: Trained ML model weights for LynD NNK7 semi-randomized library rounds 1-3.
src/notebooks/: Jupyter notebook walkthroughs for exploratory data analysis (eda.ipynb) and peptide nomination (nom.ipynb).

src/trainer.py: Script for training new models.
src/analysis.py: Script for analyzing patterns in substrate tolerance and epistasis.
src/pct_mod.py: Script for assessing percent-modification levels for partially randomized libraries (used for epi analysis).
src/runscripts: Example runscripts for the training and analysis code with default options used in the paper.

src/environment.yml: List of python dependencies for installation using conda.

Installation

We use Mamba and git to install and manage the necessary python dependencies:

git clone git@github.com:Kuhlman-Lab/LynD-substrate-modeling.git
cd LynD-substrate-modeling

mamba create -n LynD python=3.10
mamba activate LynD
pip install .

Install PyTorch dependencies:

mamba install pytorch=2.1 torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia

Install other dependencies:

mamba install -c conda-forge omegaconf "numpy<2.0" wandb tqdm pandas torchmetrics pytorch-lightning matplotlib seaborn

Training

To see a full list of training options, run python trainer.py -h. Training for 50 epochs should take around 15-20 minutes on a V100 GPU with 8 CPU workers. To run with default settings:

cd src/runscripts
sh train.sh

Inference (nomination)

Inference and peptide nomination/filtering code is provided in src/nomination.ipynb.

Analysis

Several analysis functions are included in src/analysis.py and example inputs are provided in src/runscripts/analyze.sh

To screen a set of peptides based on activity threshold and Hamming distance constraints:

python analysis.py \
	--features onehot \
	--model MLP \
	--routine screen \
	--variable_region 14 15 16 17 18 19 20 21 \
	--ckpt model_weights/LynD_NNK7_rd3.ckpt \
	--include_C \
    --thresh 0.97 \
    --hamming 2 \
	--sele /proj/kuhl_lab/users/dieckhau/LynD-substrate-modeling/LynD-substrate-modeling/datasets/rd3/sele/ \
	--anti /proj/kuhl_lab/users/dieckhau/LynD-substrate-modeling/LynD-substrate-modeling/datasets/rd3/anti/

To calculate PPV curves for S score and model predictions on a given library:

python analysis.py \
--features onehot \
--model MLP \
--routine threshold \
--variable_region 14 15 16 17 18 19 20 21 \
--ckpt model_weights/LynD_NNK7_rd3.ckpt \
--include_C \
--seed 1234 \
--sele /proj/kuhl_lab/users/dieckhau/LynD-substrate-modeling/LynD-substrate-modeling/datasets/rd3/sele/ \
--anti /proj/kuhl_lab/users/dieckhau/LynD-substrate-modeling/LynD-substrate-modeling/datasets/rd3/anti/

To calculate epistasis matrices for a semi-randomized library:

python analysis.py \
	--features onehot \
	--model MLP \
	--routine epistasis \
	--variable_region 14 15 16 17 18 19 20 21 \
	--ckpt model_weights/LynD_NNK7_rd3.ckpt \
	--include_C \
	--sele /proj/kuhl_lab/users/dieckhau/LynD-substrate-modeling/LynD-substrate-modeling/datasets/rd3/sele/ \
	--anti /proj/kuhl_lab/users/dieckhau/LynD-substrate-modeling/LynD-substrate-modeling/datasets/rd3/anti/

Contact

General questions about the LynD manuscript should be directed to the corresponding author (Dr. Albert Bowers). However, questions regarding this code repository should be raised by opening a GitHub issue on this repo.

License

This work is made available under an MIT license. See the LICENSE file for details.