Predicting Polypharmacy Side-effects Using Knowledge Graph Embeddings.

Area of Science:

• Pharmacology and Computational Biology
• Artificial Intelligence in Medicine
• Drug Interaction Studies

Background:

• Polypharmacy, the use of multiple drugs, is common for complex diseases but carries risks of adverse side effects due to drug interactions.
• Current knowledge of polypharmacy side effects is limited, necessitating advanced prediction methods.
• Machine learning, particularly knowledge graph-based approaches like Decagon, has been explored for predicting these adverse effects.

Purpose of the Study:

• To develop a novel knowledge graph embedding technique for more accurate prediction of polypharmacy side effects.
• To address the limitations of existing models, such as the high false positive rate in the Decagon model.
• To improve the reliability and efficiency of predicting adverse drug reactions in polypharmacy.

Main Methods:

• Modeled polypharmacy side effect data as a knowledge graph, similar to the Decagon model.
• Proposed a new knowledge graph embedding technique utilizing multi-part embedding vectors.
• Employed tensor decomposition for the link prediction task within the knowledge graph framework.

Main Results:

• The proposed tensor decomposition-based approach demonstrated superior performance compared to the Decagon model.
• Achieved a 12% improvement in the area under the ROC curve.
• Showcased a 16% improvement in the precision-recall curve, indicating reduced false positives.

Conclusions:

• The novel knowledge graph embedding technique offers a more accurate and reliable method for predicting polypharmacy side effects.
• Tensor decomposition provides an effective strategy for enhancing link prediction in polypharmacy knowledge graphs.
• This advancement has the potential to improve patient safety by better anticipating adverse drug interactions.