kinaseCancerEmbeddingTool

This repository presents the analysis and code used for Ravanmehr et al. (2021) Supervised learning with word embeddings derived from PubMed captures latent knowledge about protein kinases and cancer.

zenodo

The analysis encompasses a series of steps, some of which require substantial work or computation to (re)produce. Therefore, we have placed several files that represent intermediate steps of the analysis in this zenodo repository.

• clinical_trials_by_phase.tsv
• embedding_SG_dim100_upto2010.npy
• embedding_SG_dim100_upto2014.npy
• embedding_SG_dim100_upto2020.npy
• words_SG_upto2010.txt
• words_SG_upto2014.txt
• words_SG_upto2020.txt

The clinical_trials_by_phase.tsv file represents the output of the YATCP tool (See below). The embedding and words files represent word/concept embedded vectors and labels (See the tensorflow document for details about file formats).

Start to finish tutorial

See the tutorial for a streamlined, start to finish explanation of the entire pipeline.

The notebooks in this repository demonstrate the analysis steps that follow generation of word/concept embedddings.

Jupyter notebooks

Set up your jupyter environment with a kernel with all required packages. There are many ways of doing this. The following is one way.

virtualenv mykernel
source mykernel/bin/activate
(mykernel) $ pip install -r requirments
(mykernel) $ pip install jupyter
(mykernel) $ ipython kernel install --name "local-venv" --user
(mykernel) $ jupyter-lab

This will create a kernel called local-venv that will be visible in the jupyter lab environment (any name can be used).

We provide the following notebooks:

• figure1: This notebook explores and visualizes the input data and shows how Figure 1C and 1D were generated.
• randomForestClassification: This notebook shows how to perform random forest classification using our analysis pipeline, and demonstrates how the AUC and PR plots in the manuscript and supplement were generated.
• novelPredictions: This notebook shows how we used a random forest model to generate de novo predictions.

Scripts

The scripts can be run with the same virtual environment as the notebooks. We provide the following scripts:

• runRandomForest: The script generates all of the ROC/PR plots that are presented in the manuscript and supplemental material.
• pkpki: Generate file with protein kinase (PK) to protein kinase inhibitor (PKI) links. This is not needed for the analysis but was useful to check the data being used for classification by hand for quality control purposes.

running the tool

A driver script is provided (kce_tool.py) as well as Jupyter notebooks that demonstrate the usage of the package. kce_tool.py has a number of commands that are used to implement the different functionalities. Run the script with no arguments to see the commands. Run python kce_tool <command> -h to see the arguments for individual commands.

$ python kce_tool.py 
usage: kcet <command> [<args>]

The kcet commands are:
   pkpki        list of protein kinases and their inhibitors
   pkilist      get list of all protein kinase inhibitors

kinase cancer embedding tool

positional arguments:
  command     Subcommand to run

optional arguments:
  -h, --help  show this help message and exit

Generate list of protein kinase inhibitors and corresponding protein kinases from DrugCentral

We create the file input/drug_kinase_links.tsvwhich is obtained by applying the affinity(multiplicity) threshold 0.03 on data from DrugCentral. The file input/drug_kinase_links.tsv is a list of protein kinase inhibitors (PKI) that have been used to treat cancer. Each PKI is shown together with its known major targets, act-value and a relevant PubMed id (PMID). To generate drug_kinase_links.tsv file, use the following command:

python kce_tool.py pkpki [options]

--max_multiplicity: Limit on the number of PKs that are inhibited per PKI (default=5)

--outputfilename: The output file name, (default=input/drug_kinase_links.tsv)

Generate list of protein kinase inhibitors

For the downstream analysis with the yactp tool, we need to have a file with the name of each PKI on one line and to remove duplicates. To generate this file, enter the following command

python kce_tool.py pkilist 

This script generates the file protein_kinase_inhibitors.txt