Predictive Interpretable Neural Network for Druggability (PINNED)

PINNED is a neural network model designed to produce druggability scores for four separate contributions to protein druggability, as well as a final superscore representing the predicted likelihood that a protein is druggable.

About the model

The model consists of four separate deep neural networks: Sequence and structure, Localization, Biological functions, and Network information. Each subnetwork contains a single output neuron, whose scores are summed to generate an overall druggability score logit.

Prior to model training, fpocket_pipeline.ipynb was used to generate the protein drug pocket scores found in fpocket_output.csv, which were combined with other protein data stored in raw_data. feature_processing.ipynb was used to produce a single processed feature matrix, as well as a list of names of features to be inputted into each of the subnetworks, which can be found in processed_data. Note that the actual feature matrix is not included here due to its size; however, it can be generated by running the feature_processing.ipynb notebook.

Subsequently, PINNED_model.ipynb was used to train and cross-validate the model and produce scores for each of the proteins in the dataset.

Directory

• 🗎 fpocket_pipeline.ipynb

• 🗎 feature_processing.ipynb

• 🗎 PINNED_model.ipynb

• 📁 raw_data

  • 🗎 all_proteins.csv
  • 🗎 dezso_features.csv
  • 🗎 fpocket_output.csv
  • 🗎 gdpc_10-14-22.csv
  • 🗎 go_components_10-14-22.csv
  • 🗎 go_functions_10-14-22.csv
  • 🗎 go_processes_10-14-22.csv
  • 🗎 paac_10-14-22.csv

• 📁 processed_data

  • 🗎 bio_func_names.csv
  • 🗎 localization_names.csv
  • 🗎 network_info_names.csv
  • 🗎 seq_and_struc_names.csv

• 🗎 README.md

Authors

• Michael Cunningham — developed the neural network model
• Danielle Pins — generated the fpocket data

External software

Rights to AlphaFold and fpocket are governed by their respective licenses

• AlphaFold pdb models
• fpocket (2020) source code