Sensitivity Prediction on Protein Protein Interaction Networks

This repository contains all the necessary data to train the Deep Graph Network model on the sensitivity prediction task from Protein Protein interaction networks task, as presented in the work

The training and prediction code are available in the paper github repository.

The data is structured as follows:

• training_data directory contains all the needed data to reproduce the DGN training.
  • folds contains the k-fold splitting files for each use case presented in the work (UC1, UC2, UC3);
  • pyg_datalists contains all the subgraphs induced by the DyPPIN dataset, converted in pyg format. There is a datalist for each protein features set, so
    • io.pkl has only the input and output features;
    • io+emb.pkl adds the compressed protein sequence embeddings to the I/O features;
    • io+onehot.pkl adds the onehot encoding of the protein identifiers to the I/O features.
• prediction_data directory contains the data necessary to use the trained DGN to predict sensitivity on arbitrary PPIN subgraphs.
  • biogrid_ppi_with_nodes_having_embedding.pkl contains a PyTorch geometric data object containing the BioGRID interactome restricted to proteins having a precomputed protein sequence embedding, made available via UniProt.
  • biogrid_graphs_stats.csv contains statistics about the PPIN subgraphs used during training, which are used to estimate prediction reliability.
  • dyppin_proteins.txt contains a list of all the proteins present in the PPIN subgraphs of the training dataset.
  • uniprot_embeddings_pca_128.pkl contains the protein sequence embeddings compressed via PCA.
  • ckpts contains the needed model checkpoints needed for the sensitivity prediction.
• case_study_data contains the data relative to the BACH2 case study, presented in Section 4 of the paper.
  • graphs directory contains the PPIN subgraphs relative to the considered Reactome pathways; here we also provide graphs visualizations in pdf format, where we highlight in green the nodes of interest for the case study (INS, GCG, BACH2, AFF3, CUX2). Node labels refer to the BioGRID identifiers.
  • case_study_predictions.tsv contains the DGN model predictions with statistics relative to the data samples;