A repo for experimenting on creating ligands w/ high binding affinity to a target protein using ML.
Architecture consits of GAN and Transformer model. Smiles data was used for ligands and sequences for proteins.
Databases:
- BindingDB - bindingdb.org - Discription
- ATOM3D - zenodo.org
- UniProt - uniprot.org
You can execute a run via ./run_experiment.py
and see metrics and results via Tensorboard by running:
- Model will be required and path to dataset if you haven't stored them in a sql table from a previous experiment with this model:
--model [model name] --data_path [dataset path]
- If it's your first experiment w/ this model you can save the given dataset in a table so they won't need to be processed again for another experiment by additionally writing
-sql [-ho HOST] [-po PORT] [-db DATABASE]
in command line.- If you trained that model already several times you can specify a checkpoint by giving the absolute path
[--ckpt_path CKPT_PATH]
. As default it will choose the best checkpoint - meaning the one with the least amount of loss.- If you want to use a saved dataset in a table enter
-sql
without specifying the sql connection info.- After a run write in command line:
tensorboard --logdir [path to runs folder]
and open the link provided to see results.
Your SQL credentials will be asked when executing an experiment:
Create a proper structure for my own DL project.USER:
PASSWORD: