Skip to content

Package to predict dependencies between cell lines and genes using Network Representation Learning (NRL) based Link Prediction

License

Notifications You must be signed in to change notification settings

pstrybol/DeepLinkPrediction_Public

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

24 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

DOI

Setting Up the Environment

Prerequisit: You must have installed python 3.x. All scripts were tested on Ubuntu 18.04.3 LTS

  1. Install DeepLinkPrdiction package (for use of DLP model and manipulation of interaction networks):

    cd DeepLinkPrediction
    python setup.py install
    
  2. Clone and Install OpenNE (For methods GraRep, DeepWalk, node2vec and LINE):

    cd OpenNE/src
    python setup.py install
    
  3. Install forked EvalNE repository (for baseline methods and benchmarking framework). Clone the repository, then:

    cd EvalNE
    python setup.py install 
    

Demo of the DLP model

As a demonstration of how the DLP model can be called outside of the EvalNE framework, you will find a script called DLP_demo.py inside the demo folder. All necessary data is in this demo folder and the script can be run in its entirety in ~298 seconds on a server (hardware: 48-core CPU and 189GB RAM) and ~336 seconds locally (hardware: 4-core CPU and 16GB RAM).

Workflow for a single cancer type

  1. Construct heterogeneous network by integrating the LOF screening data of a certain cancer type (--disease) with a functional interaction scaffold (--ppi_scaffold) using scripts/construct_train_test_edges_separate_args.py.

  2. Construct train/test/validation edges using construct_train_test_edges_separately.py.

  3. Run EvalNE API's

    • Run EvalNE_General_Performance/EvalNE_General_Performance_API.py to calculate the general performance (on gene-gene and gene-cell line interactions) of every method.
    • Run EvalNE_CellLine_specific_Performance/EvalNE_CellLine_specific_performance_API.py to calculate the dependency specific (gene-cell line interactions only) performance of every method.
    • Run EvalNE_CellLine_specific_total_predictions/EvalNE_CellLine_Specific_totalprediction_API.py to predict a probability for each possible gene-cell line combination
  4. Run DLP models with the pretrained embeddings of DeepWalk.

    • Run DLP_baseline_initializer_Performance/DLP_baseline_initializer_performance_API.py twice to calculate the general and dependency specific performance, respectively.
    • Run DLP_baseline_initializer_Target_prediction/DLP_baseline_initializer_targetprediction_API.py to predict a probability for each possible gene-cell line combination
  5. Run scripts/construct_combined_prediction_df.py to construct a combined pandas dataframe which averages the probabilities of each interaction over the three runs.

These 5 steps are automated in run_single_cancertype.sh for a single cancer type. To iterate over several cancer types use run_several_cancertypes.sh

Useful notes

  • The parameter k-step of the method GraRep needs to be a multiple of the embedding dimension
  • The parameter order of the method AROPE needs to be equal to the dimension of the weights parameter
  • EvalNA API scripts are put in separate directories for the following reason: The version of EvalNE used in this work does not allow for the evaluation of different test sets on the same trained model. Hence, if we want to predict on a separate test and and also on a separate prediction set (eg all possible combinations of genes - cell line), 2 separate EvalNE runs are required. Additionally, EvalNE constructs tmp files on each run making it impossible to run EvalNE on 2 separate test sets at the same time impossible, unless each API call to EvalNE is sitauted in its own separate directory.

About

Package to predict dependencies between cell lines and genes using Network Representation Learning (NRL) based Link Prediction

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published