Code repository from the publication "Prediction of designer-recombinases for DNA editing with generative deep learning"

RecGen is a conditional variational autoencoder for the generation of tyrosine site-specific recombinases selective for the defined DNA target site. The repository contains the code that was used to train the RecGen models.

You can find the publication here and the recombinase sequences here

• vae_train_loocv.py: perform leave-one-out cross-validation on the training data.
• vae_train_save.py: train models with all libraries and save models.
• vae_load_predict.py: load trained models and predict for target site of interest.

• training_data_masked.csv: an example how the input data should look like and can be used to test the software. The data is not useful, all mutations have been replaced with stars.
• predict_ts.csv: contains the target sites for which predictions were made for in the publication, can be used for testing.

• Python3.9 with pandas, numpy, pytorch (compiled for cuda), argparse
• A cuda capable GPU

The application has been tested on Arch Linux v5.16.5.arch1-1 with Python 3.9.9, pytorch-gpu 1.10.1, pandas 1.4.0, numpy 1.22.1. To train the models a Nvidia Geforce RTX 3060 was used.

I recommend installing pytorch over conda, which shouldn't take more than a couple of minutes:

conda create -n "pytorch" python=3.9
conda activate pytorch
conda install -c conda-forge pytorch-gpu
conda install -c anaconda pandas
conda install -c anaconda numpy

To download the repository for use:

git clone https://github.com/ltschmitt/RecGen

python vae_train_loocv.py -i example_input/training_data_masked.csv

Expected output in output_loocv/:

• loss.csv: the loss values observed over the course of training
• parameters.txt: the parameters used for training
• prediction_freqs.csv: the frequency of the predicted amino acids for each position
• prediction_hamming.csv: the hamming distances of for example the left out libraries and the predictions
• prediction_strings.csv: the predicted target site + amino acid sequences in a comma seperated format

python vae_train_save.py -i example_input/training_data_masked.csv
python vae_load_predict.py -m saved_models -t example_input/predict_ts.csv -d example_input/training_data_masked.csv 

Expected output in saved_models/:

• parameters.txt: the parameters used for training
• CVAE_0.pt: the model file

Expected output in output_prediction/:

• parameters.txt: the parameters used for prediction
• prediction_str.csv: the predicted recombinase sequences with their target sites

All of these processes are not very demanding, so they should be done within a few minutes.

In case you want to test the application with custom data I recommend to use the --help flag on the scripts to learn about how the parameters can be adapted for your needs.