Joint Embedding-classifier Learning for improved Interpretability (JELI) Python Package

This repository is a part of the EU-funded RECeSS project (#101102016), and hosts the code for the open-source Python package JELI for the collaborative filtering approach.

Statement of need

Interpretability is a topical question in recommender systems, especially in healthcare applications. In drug repurposing, the goal is to identify novel therapeutic indications as drug-disease pairs. An interpretable drug repurposing algorithm quantifies the importance of each input feature for the predicted therapeutic drug-disease association in a non-ambiguous fashion, using post hoc methods. Unfortunately, different importance score-based approaches lead to different results, yielding unreliable interpretations.

We introduce the novel Joint Embedding Learning-classifier for improved Interpretability (JELI). It features a new structured recommender system and trains it jointly on a drug-disease-gene knowledge graph completion task. In particular, JELI simultaneously (a) learns the gene, drug, and disease embeddings; (b) predicts new drug-disease associations based on those embeddings; (c) provides importance scores for each gene. The drug and disease embeddings have a structure that depends on the gene embeddings. Therefore, JELI allows the introduction of graph-based priors on the connections between diseases, drugs, and genes in a generic fashion to recommend and argue for novel therapeutic drug-disease associations. 

Contrary to prior works, the recommender system explicitly includes the importance scores, strengthening the link between the recommendations and the extracted scores while allowing the use of a generic embedding model. The recommendation strategy in JELI can also be readily applied beyond the task of drug repurposing for any sets of items, users, and features.

Install the latest release

Using pip

bash pip install jeli

Docker

```bash

Build Docker image

docker build -t jeli .

Run Docker image built in previous step and drop into SSH

docker run -it --expose 3000 -p 3000:3000 jeli ```

Dependencies

OS: developed and tested on Debian Linux.

The complete list of dependencies for JELI can be found at requirements.txt (pip).

Usage

```python from jeli.JELI import JELI

from stanscofi.utils import loaddataset from stanscofi.trainingtesting import randomsimplesplit import pandas as pd

loads the Gottlieb drug repurposing data set

dataargs = loaddataset("Gottlieb", "./") dataset = Dataset(**data_args)

splits in training and testing sets without leakage

(trainfolds, testfolds), _ = randomsimplesplit(dataset, 0.2, randomstate=1234) train = dataset.subset(trainfolds) test = dataset.subset(test_folds)

classifier = JELI({"cudaon": False, "ndimensions": 10, "random_state": 1234, "epochs": 25})

trains JELI on the training set

classifier.fit(train)

predicts on the testing set

scores = classifier.predictproba(test) classifier.printscores(scores) predictions = classifier.predict(scores, threshold=0.5) classifier.print_classification(predictions)

computes an embedding i (item/drug)

item = pd.DataFrame(dataset.items.toarray()[:,0],index=dataset.itemfeatures,columns=["0"]) i = model.transform(item, isitem=True)

computes an embedding u (user/disease)

user = pd.DataFrame(dataset.users.toarray()[:,0],index=dataset.userfeatures,columns=["0"]) u = model.transform(user, isitem=False)

computes the feature-wise importance scores from embeddings

embs = classifier.model["featureembeddings"] featurescores = embs.sum(axis=1) ```

Licence

This repository is under an OSI-approved MIT license.

Citation

If you use JELI in academic research, please cite it as follows

``` @article{reda2025joint, title={Joint embedding--classifier learning for interpretable collaborative filtering}, author={R{\'e}da, Cl{\'e}mence and Vie, Jill-J{^e}nn and Wolkenhauer, Olaf}, journal={BMC bioinformatics}, volume={26}, number={1}, pages={26}, year={2025}, publisher={Springer} }

```

Community guidelines with respect to contributions, issue reporting, and support

Pull requests and issue flagging are welcome, and can be made through the GitHub interface. Support can be provided by reaching out to recess-project[at]proton.me. However, please note that contributors and users must abide by the Code of Conduct.