Geodesic and Neural Features for Link Prediction in the COVID-19 Biomedical Knowledge Graph

Public repository for SMCDC21 Challenge 2 submission (accepted for presentation at SMC2021).

Motivated by Challenge 2 of the 5th Annual Smoky Mountains Computational Sciences Data Challenge, we analyze the COVID-19 biomedical knowledge graph. After computing geodesic statistics for all nodes in the network, we present several machine learning pipelines for automated hypothesis generation.

Table of Contents

• Table of Contents
• Repo Organization
• Setup
• kg_browser
• Novel Relations
• Requirements
• Contact
• License

Repo Organization

This repo is organized as follows:

covid19-link-prediction
├── classifiers  # saved classification models
├── data
│   ├── betweenness     # pre-computed betweenness data
│   ├── embeddings     # trained DeepWalk embeddings
│   ├── graph     # pickled NetworkX graphs
│   ├── og     # dir for provided challenge data
│   ├── other     # dir for intermediary data/computations
│   ├── shortest_paths     # pre-computed APSP data
│   ├── training     # matrices used for training
│   └── validation     # matrices used for validation
├── dev code     # assorted dev notebooks
├── misc     # for assorted supplemental files (currently only the kg_browser image)
├── paper     # submission pre-print
└── kg_browser.py     # utils for browsing processed data/using models/etc.

Setup

1. Clone this repo
2. Download shortest paths data from here (~6 GB) and relocate it to \smcdc-2021-2\data\shortest_paths (see Repo Organization)
3. Download validation data from here (~2/3 GB) and relocate it to \smcdc-2021-2\data\validation (see Repo Organization)
4. Download challenge dataset from here and relocate it to \smcdc-2021-2\data\og (see Repo Organization)
5. cd into the covid19-link-prediction directory and run the following in your shell: pip install -r requirements.txt

kg_browser

We provide kg_browser, a convenient utility interface for accessing our processed data and models.

For further examples, see the kg_browser demo notebook here.

Novel Relations

Our top 1000 proposed novel relations may be found here. Here are the top 10:

+-----------------------+------------------------------+
| Edge                  |  Estimated Link Probability  |
+=======================+==============================+
| C0035236 <-> C1441604 |           0.999488           |
+-----------------------+------------------------------+
| C0027362 <-> C0020967 |           0.999487           |
+-----------------------+------------------------------+
| C0003062 <-> C0012754 |           0.999484           |
+-----------------------+------------------------------+
| C0086418 <-> C0027934 |           0.999484           |
+-----------------------+------------------------------+
| C0006104 <-> C0333230 |           0.999484           |
+-----------------------+------------------------------+
| C1314650 <-> C2700280 |           0.999464           |
+-----------------------+------------------------------+
| C0543467 <-> C0265883 |           0.99945            |
+-----------------------+------------------------------+
| C0582175 <-> C2697883 |           0.999444           |
+-----------------------+------------------------------+
| C1320226 <-> C0401805 |           0.999403           |
+-----------------------+------------------------------+
| C0206031 <-> C0038454 |           0.99938            |
+-----------------------+------------------------------+

Requirements

This project was created with:

• powerlaw==1.4.6
• scikit_network==0.23.1
• numpy==1.21.1
• networkx==2.5
• pandas==1.2.4
• seaborn==0.11.0
• matplotlib==3.3.2
• scipy==1.7.0
• joblib==0.17.0

Contact

• Lucas Hurley McCabe (email)

License

MIT