https://github.com/benedekrozemberczki/tigerlily
TigerLily: Finding drug interactions in silico with the Graph.
https://github.com/benedekrozemberczki/tigerlily
biology ddi deep-learning drug-drug-interaction embedding gradient-boosting graph graph-database graph-embedding graph-machine-learning heterogeneous-graph knowledge-graph machine-learning network-science node node-embedding pharmaceuticals tigergraph unsupervised-learning
Last synced: about 1 month ago
JSON representation
TigerLily: Finding drug interactions in silico with the Graph.
- Host: GitHub
- URL: https://github.com/benedekrozemberczki/tigerlily
- Owner: benedekrozemberczki
- License: apache-2.0
- Created: 2022-02-28T21:48:19.000Z (about 3 years ago)
- Default Branch: main
- Last Pushed: 2022-12-17T16:11:50.000Z (over 2 years ago)
- Last Synced: 2025-03-25T11:21:35.020Z (about 2 months ago)
- Topics: biology, ddi, deep-learning, drug-drug-interaction, embedding, gradient-boosting, graph, graph-database, graph-embedding, graph-machine-learning, heterogeneous-graph, knowledge-graph, machine-learning, network-science, node, node-embedding, pharmaceuticals, tigergraph, unsupervised-learning
- Language: Jupyter Notebook
- Homepage:
- Size: 14.3 MB
- Stars: 99
- Watchers: 2
- Forks: 9
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- Contributing: CONTRIBUTING.md
- License: LICENSE
Awesome Lists containing this project
README
[pypi-image]: https://badge.fury.io/py/tigerlily.svg
[pypi-url]: https://pypi.python.org/pypi/tigerlily
[size-image]: https://img.shields.io/github/repo-size/benedekrozemberczki/tigerlily.svg
[size-url]: https://github.com/benedekrozemberczki/tigerlily/archive/main.zip
[build-image]: https://github.com/benedekrozemberczki/tigerlily/workflows/CI/badge.svg
[build-url]: https://github.com/benedekrozemberczki/tigerlily/actions?query=workflow%3ACI
[docs-image]: https://readthedocs.org/projects/tigerlily/badge/?version=latest
[docs-url]: https://tigerlily.readthedocs.io/en/latest/?badge=latest
[coverage-image]: https://codecov.io/gh/benedekrozemberczki/tigerlily/branch/main/graph/badge.svg?token=30XLVBUIEH
[coverage-url]: https://codecov.io/github/benedekrozemberczki/tigerlily?branch=main
[![PyPI Version][pypi-image]][pypi-url]
[![Docs Status][docs-image]][docs-url]
[![Code Coverage][coverage-image]][coverage-url]
[![Build Status][build-image]][build-url]
[](https://arxiv.org/abs/2204.08206)
----------------------------------------------------------------------
### **Drug Interaction Prediction with Tigerlily**
**[Documentation](https://tigerlily.readthedocs.io)** | **[Example Notebook](https://github.com/benedekrozemberczki/tigerlily/blob/main/example_notebook.ipynb)** | **[Youtube Video](https://www.youtube.com/watch?v=fEWcor96tt8)** | **[Project Report](http://arxiv.org/abs/2204.08206)**
**Tigerlily** is a [TigerGraph](https://www.tigergraph.com/) based system designed to solve the [drug interaction prediction task](https://arxiv.org/abs/2111.02916). In this machine learning task, we want to predict whether two drugs have an adverse interaction. Our framework allows us to solve this **[highly relevant real-world problem](https://www.newscientist.com/article/2143486-side-effects-kill-thousands-but-our-data-on-them-is-flawed/)** using graph mining techniques in these steps:
- **(a)** Using [PyTigergraph](https://github.com/pyTigerGraph/pyTigerGraph) we create a heterogeneous biological graph of drugs and proteins.
- **(b)** We calculate the [personalized PageRank](https://github.com/tigergraph/gsql-graph-algorithms/blob/master/algorithms/Centrality/pagerank/personalized/multi_source/tg_pagerank_pers.gsql) scores of drug nodes in the [TigerGraph Cloud](https://tgcloud.io/).
- **(c)** We embed the nodes using [sparse non-negative matrix factorization](https://scikit-learn.org/stable/modules/generated/sklearn.decomposition.NMF.html) of the personalized PageRank matrix.
- **(d)** Using the node embeddings we train a [gradient boosting](https://lightgbm.readthedocs.io/en/latest/) based drug interaction predictor.
--------------------------------------------------------------------------------
### (A) **Creating and populating a Graph**
As a first step, the basic **TigerLily** tools are imported, and we load the example dataset that integrated DrugBankDDI and the BioSNAP datasets. We create a ``PersonalizedPageRankMachine`` and connect to the host with the ``Graph``. The settings of this machine should be the appropriate user credentials and details; a secret is obtained in the **TigerGraph Graph Studio**. We install the default Personalized PageRank query and upload the edges of the example dataset used in our demonstrations. This graph has **drug** and **protein** nodes, **drug-protein** and **protein-protein** interactions. Our goal is to predict the **drug-drug** interactions.
```python
from tigerlily.dataset import ExampleDataset
from tigerlily.embedding import EmbeddingMachine
from tigerlily.operator import hadamard_operator
from tigerlily.pagerank import PersonalizedPageRankMachine
dataset = ExampleDataset()
edges = dataset.read_edges()
target = dataset.read_target()
machine = PersonalizedPageRankMachine(host="host_name",
graphname="graph_name",
username="username_value",
secret="secret_value",
password="password_value")
machine.connect()
machine.install_query()
machine.upload_graph(new_graph=True, edges=edges)
```
### (B) **Computing the Approximate Personalized PageRank vectors**
We are only interested in describing the neighbourhood of drug nodes in the biological graph. Because of this, we only retrieve the neighbourhood of the drugs - for each drug we retrieve those nodes (top-k closest neighbors) which are the closest based on the Personalized PageRank scores. We are going to learn the drug embeddings based on these scores.
```python
drug_node_ids = machine.connection.getVertices("drug")
pagerank_scores = machine.get_personalized_pagerank(drug_node_ids)
```
### (C) Learning the Drug Embeddings and Drug Pair Feature Generation
We create an embedding machine that creates drug node representations. The embedding machine instance has a random seed, a dimensions hyperparameter (this sets the number of factors), and a maximal iteration count for the factorization. An embedding is learned from the Personalized PageRank scores and using the drug features we create drug pair features with the operator function.
```python
embedding_machine = EmbeddingMachine(seed=42,
dimensions=32,
max_iter=100)
embedding = embedding_machine.fit(pagerank_scores)
drug_pair_features = embedding_machine.create_features(target, hadamard_operator)
```
### (D) Predicting Drug Interactions and Inference
We load a gradient boosting-based classifier, an evaluation metric for binary classification, and a function to create train-test splits. We create a train and test portion of the drug pairs using 80% of the pairs for training. A gradient boosted tree model is trained, score the model on the test set. We compute an AUROC score on the test portion of the dataset and print it out.
```python
from lightgbm import LGBMClassifier
from sklearn.metrics import roc_auc_score
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(drug_pair_features,
target,
train_size=0.8,
random_state=42)
model = LGBMClassifier(learning_rate=0.01,
n_estimators=100)
model.fit(X_train,y_train["label"])
predicted_label = model.predict_proba(X_test)
auroc_score_value = roc_auc_score(y_test["label"], predicted_label[:,1])
print(f'AUROC score: {auroc_score_value :.4f}')
```
Head over to the [documentation](https://tigerlily.readthedocs.io) to find out more about installation and a full API reference.
For a quick start, check out the [example notebook](https://github.com/benedekrozemberczki/tigerlily/blob/main/example_notebook.ipynb). If you notice anything unexpected, please open an [issue](github.com/benedekrozemberczki/tigerlily/issues).
--------------------------------------------------------------------------------
**Citing**
If you find *Tigerlily* useful in your research, please consider adding the following citation:
```bibtex
@misc{tigerlily2022,
author = {Benedek Rozemberczki},
title = {TigerLily: Finding drug interactions in silico with the Graph},
year = {2022},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/benedekrozemberczki/tigerlily}},
}
```
--------------------------------------------------------------------------------
**Installation**
To install tigerlily, simply run:
```sh
pip install tigerlily
```
**Running tests**
Running tests requires that you run:
```
$ tox -e py
```
--------------------------------------------------------------------------------
**License**
- [Apache 2.0 License](https://github.com/benedekrozemberczki/tigerlily/blob/main/LICENSE)
--------------------------------------------------------------------------------
**Credit**
The **TigerLily** logo and the high level machine learning workflow image are based on:
- [Kubos Origami Font](https://www.fontspace.com/kubos-origami-font-f29538)
- [Noun Project Icons](https://thenounproject.com/)
Benedek Rozemberczki has a yearly subscription to the Noun Project that allows the customization and commercial use of the icons.