Optimizing drug-target interaction prediction based on random walk on heterogeneous networks

Area of Science:

• Computational biology
• Pharmacology
• Bioinformatics

Background:

• Predicting drug-target associations is crucial for drug development and understanding drug mechanisms.
• Computational approaches are increasingly vital for discovering these associations as data grows.
• This study applies the random walk with restart (RWR) method to a heterogeneous network of drugs and targets.

Purpose of the Study:

• To evaluate the performance of the RWR method for predicting drug-target interactions.
• To investigate the impact of parameter variation and chemical fingerprint selection on RWR performance.
• To identify novel drug-target associations using a computational approach.

Main Methods:

• Constructed a heterogeneous network of drugs and targets using DrugBank data.
• Applied the random walk with restart (RWR) algorithm.
• Evaluated RWR performance using the ChEMBL15 dataset with varying bioactivity cutoffs (1 µM and 10 µM).
• Assessed prediction accuracy based on the rank of true interactions.

Main Results:

• RWR performance is robust to the choice of chemical fingerprint when parameters are optimized.
• At a 1 µM bioactivity cutoff, RWR achieved prediction accuracies of 47-60% in the top 50 ranks.
• At a 10 µM bioactivity cutoff, RWR achieved prediction accuracies of 32-35% in the top 50 ranks.
• Identified top ten predicted targets for 110 popular drugs.

Conclusions:

• The RWR method applied to heterogeneous networks with chemical features is effective for identifying novel drug-target interactions.
• The study demonstrates the promise and performance of this computational approach.
• Parameter optimization is key to maximizing the predictive power of RWR.