Integrating Mechanistic Information to Predict Drug-Drug Interactions and Associated Relevance for Decision Support

Abstract

While computational methods offer great potential in predicting drug-drug interactions (DDIs), such predictions as of yet have limited utility in supporting clinical decision-making; in particular, there exists especial difficulty in deriving interaction mechanisms from the vast abundance of available information on potential DDIs. Here, we present a backward-chaining inference algorithm that operates on a knowledge graph integrating multiple types of mechanistic information, from metabolizing enzymes to genetic variants. Given two drugs of interest, this algorithm applies complex rules to identify evidence supporting their potential interaction, which in turn suggests their mechanism of interaction. An evaluation of the ruleset using two widely-used drugs with a suspected interaction, the antibiotic levofloxacin and the chemotherapeutic irinotecan, successfully identified pharmacological and biomedical features that support and may explain their interaction. This algorithm represents a first step toward effectively assessing the clinical relevance of identified DDIs, and of identifying pairs of interacting drugs that may be validated in the experimental setting to support clinical decision-making and ultimately improve medication safety.