Positive-Unlabeled Learning for inferring drug interactions based on heterogeneous attributes

Area of Science:

• Computational chemistry
• Pharmacology
• Bioinformatics

Background:

• Drug-drug interactions (DDIs) are critical for effective clinical care but experimentally detecting them is costly and time-consuming.
• Computational methods for DDI inference are highly desirable due to the scarcity of confirmed non-interacting drug pairs (negatives) for training standard classifiers.
• A Positive-Unlabeled Learning approach is proposed to address the lack of negative data in DDI prediction.

Purpose of the Study:

• To develop an efficient computational method for inferring potential drug-drug interactions (DDIs).
• To overcome the challenge of limited negative data in training predictive models for DDIs.
• To classify predicted DDIs into Cytochrome P450 (CYP) enzyme-dependent and independent interactions.

Main Methods:

• Utilized Growing Self-Organizing Maps to infer negative drug pairs from unlabeled data.
• Employed a pairwise similarity function to quantify drug feature overlap.
• Applied a support vector machine classifier for DDI prediction, incorporating inferred negatives.

Main Results:

• Inferred 589 drug pairs as likely non-interacting, serving as negative training data.
• Successfully classified predicted DDIs into CYP-dependent and CYP-independent categories based on their location on the Self-Organizing Map.
• Identified 5300 potential CYP-dependent and 592 CYP-independent DDIs with high confidence.

Conclusions:

• The proposed Positive-Unlabeled Learning method significantly improved F1-scores compared to random negative sampling.
• The inferred DDIs, particularly CYP-dependent interactions, hold potential for enhancing clinical care and drug development.
• The method provides a valuable tool for predicting DDIs, aiding in safer and more effective drug prescriptions.