Predicting and understanding comprehensive drug-drug interactions via semi-nonnegative matrix factorization

Area of Science:

• Computational chemistry
• Pharmacology
• Network science

Background:

• Drug-drug interactions (DDIs) can cause adverse reactions, necessitating effective preclinical identification.
• Current computational methods predict DDIs but struggle with comprehensive prediction of enhancive and degressive interactions.
• A systematic analysis of structural relationships in DDIs is lacking, hindering understanding of interaction mechanisms.

Purpose of the Study:

• To develop a novel computational model for predicting both conventional and comprehensive drug-drug interactions (DDIs).
• To analyze the structural relationships within DDI networks to understand interaction mechanisms.
• To guide co-prescription decisions through improved DDI prediction and analysis.

Main Methods:

• Treated comprehensive DDIs as a signed network.
• Designed a novel model (DDINMF) based on semi-nonnegative matrix factorization for DDI prediction.
• Represented DDIs as binary and signed networks for structural analysis using NMF.

Main Results:

• DDINMF achieved high performance in conventional DDI prediction (AUROC=0.872, AUPR=0.605) and comprehensive DDI prediction (AUROC=0.796, AUPR=0.579).
• The model outperformed two state-of-the-art approaches.
• NMF analysis revealed significant clustering and structural balance within DDI networks, uncovering hidden knowledge.

Conclusions:

• The developed approach predicts both binary and comprehensive DDIs effectively.
• DDI networks exhibit clear clustering and specific drug degree distributions.
• The study demonstrates that even binary DDI networks imply relationships found in comprehensive networks, and DDI sign occurrence is not random due to network structural balance.