Introduction

This repository contains code to streamline the creation of a time-resolved drug repurposing heterogenous network benchmark. This dataset aims to better model knowledge graph completion algorithm performance on predicting unseen future approved drug-disease indications.

The knowledge graph is derived off SemMedDB and is extended by approved drug-disease indications from DrugCentral as highlighted in Mayers et. al (2019). This repository updates the original pipeline by replacing notebook executions with python scripts, refactoring code to decrease dependencies and increase speed, adds option flagging, and various bug fixes.

Data is segmented by year. Each year is segmented into a train and test set (or train, test, and valid set); the train set is comprised of triples prior to the given year, and the test set is comprised of approved drug-disease triples after the given year. There are a total of 73 datasets spanning from 1950 to 2023, 6 metanodes (types), and 32 unique relations.

To build the dataset in this repository, a UTS license/account is required. You can apply for one here (may take up to 3 days to be approved). Additionally, you must have access to a PostgreSQL database for initial processing.

A convenient script to integrate the dataset into PyKEEN, a knowledge graph embedding framework and library, can be found here. Installation details can be found below.

Usage

The dataset can be generated automatically via build.sh or manually by stepping through the scripts of interest, sequentially. Either way, there are some required softwares needed prior to usage.

Required

postgres

• Please install pv and postgresql prior to running.

# on ubuntu
sudo apt-get update
sudo apt-get upgrade

sudo apt install pv postgresql

• Create a psql user and password

# login to psql, create a username and password as postgres, quit.
user@server:~$ sudo su postgres 
postgres:~$ createuser [myuser] --createdb --pwprompt
postgres:~$ exit

conda/mamba

• Please install mamba following the directions outlined in the link.

Building the dataset for your own use

• Building the dataset can be done automatically by calling the build.sh script with appropriate flags as seen in the example below.

bash build.sh --apikey [UMLS apikey] --host [psql hostname] --pass [psql password] \
 --port [psql port] --user [psql user]

• When building, it is highly recommended to run the command in a persistent session using screen (deprecated) or tmux (preferred); even with all the tweaks, downloading files and pinging APIs still takes a while.

Manual mode

• If you want to run each script by itself (or enjoy life on hard mode), first install the python environment using the mini_env.yml file. To do so, first ensure you have conda/mamba (preferred) installed (above), then use the following command in your favorite linux shell in the cloned directory:

mamba env create -f mini_env.yml -y

• and activate the environment with:

mamba activate mini_semmed2

pykeen integration

To integrate the dataset with pykeen, do the following things:

• clone the pykeen repository
• copy timeresolvedkg.py file in this repository into your local pykeen repo atpykeen/datasets/timeresolvedkg.py; this file is a pykeen dataset class that will run the entire pipeline (if you supply the minimum flagged information)
• Add the file path to the setup.cfg entrypoint for pykeen. Example:

[options.entry_points]
console_scripts = 
    pykeen = pykeen.cli:main
pykeen.datasets = 
    trkg = pykeen.datasets.timeresolvedkg:TimeResolvedKG

• Finally, rebuild the pykeen library with pip install .

Organization

1. prepare

The prepare folder contains the scripts and notebooks needed to download semmedDB and other mapping.

files used for this repo, as well as some pre-processing steps. Run bash download_requirements.sh with flags to download the appropriate files. Run python preprocessing.py to generate files for downstream data cleaning.

2. build

The build folder contains the notebooks that builds the time-resolved knowledge graph. Run python building.py to create the knowledge graph and appropriate splits by using the flagging options.

tools

The tools folder cotains some helpful scripts, useful for processing project data.

data

All data will be downloaded and written to this folder. Contains a couple small starter files important for running and speeding up the pipeline.

References

Time-resolved network

Michael Mayers, Tong Shu Li, Núria Queralt-Rosinach, Andrew I. Su. (2019) Time-resolved evaluation of compound repositioning predictions on a text-mined knowledge network. BMC Bioinformatics 20, 653. https://doi.org/10.1186/s12859-019-3297-0

SemMedDB

Halil Kilicoglu,Dongwook Shin, Marcelo Fiszman, Graciela Rosemblat, Thomas C Rindflesch. (2012) SemMedDB: A PubMed-scale repository of biomedical semantic predications. Bioinformatics, 28(23), 3158-60. https://doi.org/10.1093/bioinformatics/bts591

DrugCentral

Oleg Ursu Jayme Holmes Jeffrey Knockel Cristian G. Bologa Jeremy J. Yang Stephen L. Mathias Stuart J. Nelson Tudor I. Oprea. (2017) DrugCentral: online drug compendium. Nucleic Acids Research, 45(D1) D932–D939. https://doi.org/10.1093/nar/gkw993