The world’s leading publication for data science, AI, and ML professionals.

Network Analysis and Visualization of Drug-Drug Interactions

Discover how to explore and visualize drug interactions data (in Python) with NetworkX, Pyvis, and Datapane

Visualization of drug-drug interactions on a data subset | Image by author
Visualization of drug-drug interactions on a data subset | Image by author

Drug-drug interactions occur when two or more drugs react with each other, resulting in unexpected (and potentially fatal) adverse reactions.

Being a data scientist with a background in clinical pharmacy, I was keen to leverage my domain knowledge to explore the data around drug interactions.

Drug-drug interactions involve linkages between multiple entities, so it would be apt to analyze these drug interactions with networks as the core concept.

Contents

(1) Data Acquisition and Preparation (2) Network Analysis with NetworkX (3) Visualization with Pyvis and DataPane (4) Python Implementation Codes (5) Moving forward

Photo by Volodymyr Hryshchenko on Unsplash
Photo by Volodymyr Hryshchenko on Unsplash

Disclaimer: This article is NOT intended to provide any form of medical advice or recommendations. Please consult your doctor or pharmacist for professional advice relating to any drug therapy.


(1) Data Acquisition and Preparation

The Stanford Network Analysis Project (SNAP) has a public Biomedical Network Dataset Collection comprising many datasets with information on relationships between entities. Some examples include disease-drug association, relations between biological processes, and even similarity networks between butterflies.

I used the ChCh-Miner dataset for this project, which comprises a network of interactions between Food and Drug Administration (FDA) approved drugs.

The raw data is stored as a CSV file of 48,464 rows and 2 columns, where each row represents an interaction between a particular pair of drugs.

The values are in the form of DrugBank (DB) codes, so they need to be matched with actual drug names. This is done manually by searching iteratively for the corresponding drug names from the public DrugBank database.

After further data pre-processing (e.g. removal of compounds without drug names), we get a Pandas DataFrame that looks like this:

Sample rows of processed drug interactions data | Image by author
Sample rows of processed drug interactions data | Image by author

A downside of this dataset is that the severity of drug interactions is not provided. Therefore, a custom column (named ‘Weight) filled with 1’s was added to indicate ‘equal’ severity for all interactions.


(2) Network Analysis with NetworkX

(i) Overview

A network is simply a collection of connected objects, and they are highly useful for modeling relationships (aka edges) between entities (aka nodes).

Examples of networks | Image used under pyvis BSD License
Examples of networks | Image used under pyvis BSD License

In our dataset, every pairwise drug interaction is a relationship between two nodes (i.e. the two drug compounds).

Given that we have a bunch of undirected linkages between drug entities, we can proceed to perform network analysis with NetworkX, a Python package used for the creation, manipulation, and study of networks.


(ii) Creation of Network Graph

NetworkX can be installed with: pip install networkx[default]

Once done, we generate a network graph object from the drug interactions Pandas DataFrame, and retrieve a basic summary of the network graph:

Summary results from undirected network graph of drug interactions | Image by author
Summary results from undirected network graph of drug interactions | Image by author

We can see that the network has a total of 1,505 drug entities (nodes) with 48,224 documented interactions (edges).

The average degree is 64, meaning that each drug typically interacts with 64 other drugs on average. (More on node degree in the later section)

Network density describes the proportion of potential connections in a network that are actual connections, with 0 representing a graph with zero edges, and 1 representing a complete graph where all nodes are interconnected.

The low network density of 0.04 makes sense because we do not expect too many medications to have interactions with each other.


(iii) Degree Centrality

The degree of a node refers to the number of connections the node has. When working with networks, we deal with a key concept called degree centrality.

To put it simply, the degree centrality of a node is simply the number of edges it has. This means that a drug node with 5 interactions would have a degree centrality of 5.

By calculating the degree centrality of each node, we can get the top 20 drugs most frequently involved in drug interactions:

Top 20 drugs (with corresponding node degree) involved in the most interactions | Image by author
Top 20 drugs (with corresponding node degree) involved in the most interactions | Image by author

The output above reveals that the drug most frequently involved in drug interactions is Phenytoin, with a node degree of 442.

Phenytoin is a commonly used anti-epileptic medication for the treatment of seizure disorders, and its concentration in the blood has to be kept carefully within consistent levels to ensure its effectiveness and safety.

MediView with Keywords: phenytoin

Here is an example of what a drug interaction involving Phenytoin looks like:

Image used based on Creative Common's Attribution-NonCommercial 4.0 International License as indicated by DrugBank
Image used based on Creative Common’s Attribution-NonCommercial 4.0 International License as indicated by DrugBank

(3) Visualization with Pyvis and DataPane

(i) Overview

Probably the coolest part about working with networks is the visualization aspect. A tool to help us easily construct such visualizations is Pyvis, a Python library well-known for creating interactive network graphs.

Pyvis can be installed with: pip install pyvis

The sequence to creating a Pyvis visualization from a DataFrame is as such:

  1. Generate a NetworkX graph from the Pandas DataFrame containing the edge list with the [nx.from_pandas_edgelist](https://networkx.org/documentation/stable/reference/generated/networkx.convert_matrix.from_pandas_edgelist.html) method.
  2. Initiate a Pyvis Network class instance with [Network()](https://pyvis.readthedocs.io/en/latest/_modules/pyvis/network.html)
  3. Take the NetworkX graph and translate it to a Pyvis graph format with the .from_nx method of the Network() class

There is plenty of customization available to structure or modify the appearance of the network visualization. In my case, I experimented with the popular Repulsion solver and Barnes Hut physics model.

I started by dumping the entire drug interactions data into a Pyvis Barnes Hut graph and in return, I obtained a terrifyingsea urchin lookalike’ network.

Network visualization (Barnes Hut physics model) of the full interactions dataset | Image by author
Network visualization (Barnes Hut physics model) of the full interactions dataset | Image by author

(ii) Exploring a single drug

To make the graphs more meaningful and interpretable, it would be better to work on a smaller subset of data instead of the full dataset.

Since we discovered earlier that Phenytoin has the most drug interactions, let’s create a Pyvis network (using Repulsion solver) for Phenytoin alone:

Note: Click the ‘fullscreen’ button if viewing this on mobile devices

The above visualization presents Phenytoin as the central node, surrounded by 442 other drug entities that it interacts with.

The beauty of Pyvis is that the visualizations are interactive, meaning that you can actually mouse over the graph to zoom, point or drag the nodes.

Since Pyvis saves the graphs as HTML files locally, so we will need a separate tool to share the results publicly. This is where Datapane comes into the fray.

Datapane is an open-source Python library that allows users to publish objects like HTML files and Python plots on the web for sharing and visualization. You can find the script I wrote for uploading HTML files to Datapane here.


(iii) Exploring multiple drugs

Looking at one drug is probably not insightful, so let’s explore a network comprising a slightly larger data subset.

I picked three drugs (Metformin, Lisinopril, and Simvastatin) for further exploration based on their key roles in the prevalent diseases of diabetes, hypertension, and hyperlipidemia respectively.

Upon filtering the data to keep interactions that involve at least one of these three drugs, the resulting graph looks like this (somewhat resembling shapes of jellyfish and broccoli):

From the network graph, we can readily pinpoint the different drugs which have common interactions with Metformin, Lisinopril, and/or Simvastatin. As before, this network graph is interactive, so go ahead and explore it.

One interesting insight is that there is only one entity that interacts (i.e. has edges) with all three handpicked drugs, and that drug is Quinine.

Quinine interacts with Metformin, Lisinopril, and Simvastatin | Image by author
Quinine interacts with Metformin, Lisinopril, and Simvastatin | Image by author

(4) Python Implementation Codes

I have posted the comprehensive codes and data inside this GitHub repo, so feel free to explore (and star/fork) the contents.

Do experiment with the other functionalities and network customizations within NetworkX and Pyvis, and also try creating your very own networks!

Photo by Fernando Hernandez on Unsplash
Photo by Fernando Hernandez on Unsplash

(5) Moving forward

With the baseline analysis and visualizations complete, here are several potential next steps:

  • Drug interactions have different levels of severity (i.e. minor, moderate, severe), and this was not accounted for due to lack of data. It would be ideal to find and include that information so that the edges can have different thicknesses to reflect varying severity.
  • Add details of the interaction as a pop-up upon hover over the nodes and edges
  • Deploy this project as a Streamlit web app so that users can handpick their own set of drugs to visualize in a network graph.

Update: The Streamlit deployment has recently been completed! Check it out here:

How to Deploy Interactive Pyvis Network Graphs on Streamlit


Before You Go

I welcome you to join me on a Data Science learning journey! Do give this Medium page a follow to stay in the loop of more exciting data science content, or reach out on LinkedIn. Have fun exploring network graphs!

Exploring Illegal Drugs in Singapore – A Data Perspective

COVID-19 Vaccine – What’s the Public Sentiment?


References


Related Articles

Some areas of this page may shift around if you resize the browser window. Be sure to check heading and document order.