OpenPredict KP

Homepage Link:

Description: Translator OpenPredict KP is a collection of machine learning models to predict drug-treats-disease and drug-interacts_with-protein relations. drug-treats-disease predictions are based on a similarity-based machine learning approach called PREDICT, which is implemented in the open-source fair-workflows/openpredict project. drug-interacts_with-protein predictions are based on a linear classifier using ESM2 protein embeddings and MolecularTransformer drug embeddings to predict potential drug-targets interactions, where targets are proteins.

The OpenPredict API can be used to find drugs to treat a given disease, retrieve the most similar drugs and diseases (drug-drug and disease-disease pairs), and drug-target interactions.

Use the /query operation to submit a TRAPI query to the service. Each result (and each corresponding knowledge graph edge) contains a score showing the prediction probability output of the machine learning model. The prediction probability indicates the confidence level of a model in its prediction based on the model's learned patterns, expressed as a percentage (range: [0-1]), reflecting the likelihood of the relation expressed by the knowledge graph edge. For example, a score of 0.8 from the OpenPredict model on the knowledge graph edge (drug X)-[treats]->(disease Y) can be interpreted as "The Translator OpenPredict model thinks there is an 80% likelihood that drug X treats disease Y."
Use the /predict operation to retrieve predictions for given drugs or diseases. The score represents the probability of the drug-disease relation to be in the treats class (range: [0-1])
Use the /similarity operation to retrieve the most similar drugs or diseases. The score in the /similarity operation represents the cosine embedding similarity, with higher values indicating greater similarity (range: [0-1])
Use the /models operation to get information about the OpenPredict prediction model.
Use the /embeddings operation to add new model embeddings.

Modes of Access

The OpenPredict API is maintained by the Clinical Data Knowledge Provider
Translator Reasoner API, registered as a SmartAPI

Use Cases

Jupyter Notebook

Technical User Guide

Documentation
OpenPredict Python package to train and serve a new model.
OpenPredict source code
Testing
Contributing

Knowledge sources Accessed

OMIM from PREDICT reference dataset
DrugBank from PREDICT reference dataset
COHD
Embeddings:
- DrugBank-OMIM PREDICT reference dataset
- NeuroDKG, a curated database of neurological diseases and the drugs that can be used to treat them
- Embeddings generated from clinical co-occurrence counts from COHD