SEffNet

SEffNet (Side Effect Network embeddings) is a tool that optimizes, trains, and evaluates predictive models for biomedical networks that contain drug-, target- and side effect-information using different network representation learning methods in an attempt to understand the causes of side effects.

This package was developed during the master's thesis of Rana Aldisi.

Structure

• notebooks: Notebooks that were used for training and evaluation of models, and interpertation of prediction model
• resources: The graphs and materials that are used for training and testing

Installation

seffnet can be installed on python37+ from the latest code on GitHub with:

$ pip install git+https://github.com/seffnet/seffnet.git

Usage

Using the predictive model

If you've installed seffnet locally, you can use the default model from the GitHub repository with:

from seffnet.default_predictor import predictor

# Find new relations for a given entity based on its CURIE
results = predictor.find_new_relations(curie='pubchem.compound:5095')
...

You can get the embeddings for phenotype entities with

import itertools as itt
from seffnet.default_predictor import predictor

phenotype_to_embedding = {
    node_data['identifier']: predictor.embeddings[node_id]
    for node_id, node_data in predictor.node_id_to_info.items()
    if node_data['namespace'] == 'umls'
}
# could use sklearn.metrics.pairwise.cosine_similarity on the values in this dict

You can use the default model in the CLI:

$ seffnet predict pubchem.compound:5095

You can predict on new chemicals via their SMILES strings based on their similarity to chemicals included in the network. Warning: we haven't benchmarked how well this actually works yet.

$ seffnet predictc "C1=CC=C(C=C1)C2=CC=C(C=C2)CCO"

Rebuilding the resources

You can rebuild all the graphs and maps created for this project by running the following:

$ seffnet rebuild

Note that you need to have RDKit package and environment to be able to run this command

Model training and evaluation

You can train an NRL model using the following:

$ seffnet train --input-path ./resources/basic_graphs/fullgraph_with_chemsim.edgelist --evaluation --method node2vec

• For further CLI options and parameters use --help, -h

Optimizing hyperparameters

Network representation learning models can be optimized with:

$ seffnet optimize --input-path ./resources/basic_graphs/fullgraph_with_chemsim.edgelist --method node2vec

• For further CLI options and parameters use --help, -h

Web Application

The web application allows users to get results from the model programmatically. Make sure the extra dependencies have been installed as well using the [web] extra. Unfortunately, this doesn't work when installing directly from GitHub, so see the setup.cfg for the Flask dependencies.

$ pip install -e .[web]

Run development server with:

$ seffnet web --host localhost --port 5000

Run through docker with:

$ docker-compose up

• A user interface can be found at http://localhost:5000
• An auto-generated swagger UI can be found at http://localhost:5000/apidocs

As an example, you can check the chemicals predicted to interact with HDAC6 at http://localhost:5000/predict/uniprot:Q9UBN7?results_type=chemical.