Implementation of BANANA (BAsic NeurAl Network for binding Affinity), as described in the BigBind paper.
In addition to pip installing the requirements, you'll need to install torch, dgl, and rdkit.
Once you've installed all the dependencies, you're ready to run the pretrained model. First, create a PDB file containing the pocket residues of your target protein. Then create a smi file with the SMILES strings of the compounds you want to screen (one smiles string per line). Now simply run python inference.py compounds.smi pocket.pdb --out_file out_file.txt. This will load the pretrained weights from data/banana_final.pt, score each compound in compounds.smi, and write all the scores to out_file.txt.
If the inference script fails to load any smiles in the dataset, it will output -100 for the BANANA score.
If you want to train the model yourself, first make sure you've downloaded the BigBind dataset.
Now create a file configs/local.yaml. This contains all the configuration that should differ per-computer. Add in this information:
project: "Your wandb project (optional)"
platform:
bigbind_dir: "/path/to/bigbind/folder"
# lit_pcba_dir is optional, only needed if you want to test your model on LIT_PCBA
lit_pcba_dir: "/path/to/lit_pcba/folder"
# To speed up dataloading and other random things, many pickles are saved.
# cache_dir specifies where they should be saved to
cache_dir: "/path/to/cache/dir"
num_workers: num_workers_for_your_dataloader
Now that you have this, train a model py running python train.py config_name overrides. The config name used in the BigBind paper is classification. This config is found in configs/classification.yaml. Feel free to make your own config. Anything in the config file can be overwritten with command line arguments. For instance, train with a batch size of 2 with python train.py classification batch_size=2.
If you have a Weights and Biases project and specify it in the local config file, the training script will log all the weights and metrics.
The validation scripts are all designed to download models from Weights and Biases. Unfortunately, you'll need to modify them if you don't want to use wandb.
To validate the model on any of the BigBind data splits, run python -m validation.validate run_id=wandb_run_id tag=tag_of_weights_artifact data_split=train|val|test.
If you want to benchmark the model on the BigBind screening benchmarks, run python -m validation.screen run_id=wandb_run_id tag=tag_of_weights_artifact benchmark=bigbind data_split=val|test. To benchmark on LIT-PCBA, you'll need to first run preprocess_lit_pcba.py to produce pocket files for all the targets (note: this script requires PyMol). Once you've done this, run python -m validation.screen run_id=wandb_run_id tag=tag_of_weights_artifact benchmark=lit_pcba. Both these scripts will produce csv files in outputs/.
If you have any questions, email me at [email protected].