Inverse similarity and reliable negative samples for drug side-effect prediction

Area of Science:

• Computational chemistry
• Pharmacology
• Bioinformatics

Background:

• In silico prediction of drug side-effects is crucial for efficient drug development, but current methods are hindered by a lack of reliable negative training data.
• Existing computational approaches often rely on validated drug-side effect relationships, which are insufficient for robust model training.

Purpose of the Study:

• To develop a novel computational method for selecting highly reliable negative samples for drug side-effect prediction.
• To improve the performance of predictive models by addressing the challenge of insufficient negative training data.

Main Methods:

• Proposed an inverse similarity hypothesis: dissimilar drugs are less likely to share side-effects.
• Developed a drug similarity integration framework incorporating chemical structures, target proteins, substituents, and therapeutic information.
• Selected negative samples by prioritizing candidate drugs with lower similarity scores to validated positive drugs.

Main Results:

• The drug similarity integration framework demonstrated superior capability in capturing drug features compared to single-property methods.
• Machine learning algorithms (SVM, RBF, KNN) showed significant improvements in F1-score, precision, and recall when using the selected reliable negative samples.
• The proposed method achieved enhanced performance in simulative side-effect prediction for 917 DrugBank drugs.

Conclusions:

• The inverse similarity hypothesis and integrated drug properties are valuable for enhancing drug side-effect prediction accuracy.
• The selection of highly reliable negative samples is a critical factor in improving the performance of computational prediction models.