Neural Fingerprint-based Molecular Docking

A repository enabling a deep-learning approach to molecular docking. Utilizes differentiable, trainable, neural fingerprints to filter dynamic-sized databases of molecules based on binding favorability to proteins/binding sites.

steps to run:

make sure you have all the necessary python libraries installed with pip install -r requirements.txt

1. use python createModels.py ... to populate the src folder with training scripts

• python createModels.py -f/--fpl <fingerprint length> -p/--protein <protein name>
• can alter hyperparameters in createModels.py if you'd like; to alter architecture, make sure you change the architecture in the base model in createModels.py and train.py
  
  

2. use sbatch trainModels.sh if using slurm

• fill in slurm configuration according to resources available, etc.
• on the last line in trainModels.sh, make sure to fill in the path to the log file as well as path to the scripts. Use $SLURM_ARRAY_TASK_ID to vary names of log and model based on slurm array
  
  

3. results

• succinct results will be populated in {protein}/hpResults.csv
• files with detailed results (which you should check) are named corresponding to the model number and are within {protein}/res/ with various names
• output logs for runs are in {protein}/logs/
• specific scripts for each model with hyperparam sets are in {protein}/trainingJobs/
• model checkpoints based on model number are stored in {protein}/models/

Acknowledgements to Xuhan Liu and the work of Gentile et. al. for some of the code in this repository