/decagon

Graph convolutional neural network for multirelational link prediction

Primary LanguageJupyter NotebookMIT LicenseMIT

Decagon: Representation Learning on Multimodal Graphs

Author: Marinka Zitnik (marinka@cs.stanford.edu)

Project website

Overview

This repository contains code necessary to run the Decagon algorithm. Decagon is a method for learning node embeddings in multimodal graphs, and is especially useful for link prediction in highly multi-relational settings. See our paper for details on the algorithm.

Usage: Polypharmacy

Decagon is used to address a burning question in pharmacology, which is that of predicting safety of drug combinations.

We construct a multimodal graph of protein-protein interactions, drug-protein target interactions, and polypharmacy side effects, which are represented as drug-drug interactions, where each side effect is an edge of a different type.

Decagon uses graph convolutions to embed the multimodal graph in a compact vector space and then uses the learned embeddings to predict side effects of drug combinations.

Running the code

The setup for the polypharmacy problem on a synthetic dataset is outlined in main.py. It uses a small synthetic network example with five edge types. Run the code as following:

$ python main.py

The full polypharmacy dataset (described in the paper) is available on the project website. To run the code on the full dataset first download all data files from the project website. The polypharmacy dataset is already preprocessed and ready to use. After cloning the project, replace the synthetic example in main.py with the polypharmacy dataset and run the model.

Citing

If you find Decagon useful for your research, please consider citing this paper:

@article{Zitnik2018,
  title     = {Modeling polypharmacy side effects with graph convolutional networks.},
  author    = {Zitnik, Marinka and Agrawal, Monica and Leskovec, Jure},
  journal   = {Bioinformatics},
  volume    = {34},
  number    = {13},
  pages     = {457–466},
  year      = {2018}
}

Miscellaneous

Please send any questions you might have about the code and/or the algorithm to marinka@cs.stanford.edu.

This code implements several different edge decoders (innerproduct, distmult, bilinear, dedicom) and loss functions (hinge loss, cross entropy). Many deep variants are possible and what works best might depend on a concrete use case.

Requirements

Decagon is tested to work under Python 2 and Python 3.

Recent versions of Tensorflow, sklearn, networkx, numpy, and scipy are required. All the required packages can be installed using the following command:

$ pip install -r requirements.txt

License

Decagon is licensed under the MIT License.