UCBbind is a program for predicting binding affinities for protein-ligand pairs. The program implements two modules:
- Module Y (Transfer prediction module): Uses sequence alignment and Tanimoto similarity to select reference protein-ligand pairs and replicate their binding free energies.
- Module X (Deep learning module): Takes features extracted from protein sequences and ligand SMILES strings to predict binding affinities.
Justin Purnomo, Caitlin Kim, Kunyang Sun, Yingze Wang, and Teresa Head-Gordon
This environment can be built via: conda env create -f env.yml
To train Module X, run python X_prep.py
To train Module Y, run python Y_prep.py
Note: The trained Module X has already been provided. Module Y requires large .idx and .pkl files and the BindingDB dataset, which are not included in the repo due to size. You can download the cleaned BindingDB dataset here:
After downloading BindingDB.csv and placing in the datasets folder, you can train Module Y.
Predictions can be run using python FEpred.py.
The script expects a CSV file with the following columns: Sequence and SMILES. These describe the protein sequence of the query and the ligand SMILES of the query. Optionally, the CSV may include a Value column containing the experimental binding free energy in positive kcal/mol. If the Value column is present, the script will compute and display evaluation metrics comparing predictions to ground truth. If the Value column is absent, the script will still generate predictions but will skip metric computation.
The default test set used in FEpred.py is: `test_fp = 'datasets/PDBbind.csv'.
To assess classification accuracy, you can run python classifier_statistics.py
This script calculates the binder v nonbinder classification accuracy based on a pIC50 threshold of 5 for the binding affinity.