Interpretable Drug-to-Drug Network Features for Predicting Adverse Drug Reactions.

Area of Science:

• Pharmacovigilance
• Computational Chemistry
• Machine Learning in Healthcare

Background:

• Adverse drug reactions (ADRs) contribute significantly to global hospitalizations, necessitating early prediction during drug development.
• Traditional pre-clinical and clinical drug research phases are time-consuming and costly.
• There is a growing need for advanced data mining and machine learning techniques to enhance ADR prediction.

Purpose of the Study:

• To construct a drug-to-drug network utilizing non-clinical data based on shared ADRs.
• To extract and integrate network features with original drug features for improved predictive modeling.
• To evaluate the efficacy of machine learning models enhanced with network-based features for ADR prediction.

Main Methods:

• A drug-to-drug network was built using common ADRs from non-clinical data sources.
• Node-level and graph-level network features (e.g., weighted degree centrality, PageRank) were extracted.
• These network features were concatenated with original drug features and tested across seven machine learning models.

Main Results:

• All tested machine learning models demonstrated improved performance when network features were included.
• Logistic regression achieved the highest mean AUROC score of 82.1% across all tested ADRs.
• Weighted degree centrality and weighted PageRank were identified as the most influential network features for logistic regression.

Conclusions:

• Network-based approaches significantly enhance the prediction of adverse drug reactions.
• The developed network approach shows promise for application in other health informatics datasets.
• This methodology can be vital for early risk reduction in drug development.