Prediction of Drug Combinations with a Network Embedding Method.

Area of Science:

• Computational biology
• Pharmacogenomics
• Bioinformatics

Background:

• Complex diseases involve multiple targets, necessitating drug combinations for effective treatment.
• Traditional methods for identifying drug combinations are time-consuming and expensive.
• Computational approaches offer a feasible alternative for predicting drug combinations.

Purpose of the Study:

• To develop a novel computational method for predicting effective drug combinations.
• To integrate network topological features with individual drug features for enhanced prediction.
• To build an optimal machine learning model for drug combination prediction.

Main Methods:

• Constructed a drug network using chemical-chemical interaction data from STITCH.
• Employed a novel network embedding method to extract topological features of drug combinations.
• Utilized synthetic minority oversampling technique (SMOTE), minimum redundancy maximum relevance (mRMR), and incremental feature selection (IFS).
• Developed an optimal support vector machine (SVM) classifier.

Main Results:

• The optimal SVM classifier achieved a Matthews correlation coefficient (MCC) of 0.806.
• The proposed method, incorporating topological and individual features, outperformed models using only individual features.
• Combining topological and essential features significantly improved prediction performance.

Conclusions:

• The novel network embedding method effectively captures crucial topological features for drug combination prediction.
• The developed computational approach provides an efficient and accurate tool for identifying effective drug combinations.
• This method has the potential to accelerate the discovery of treatments for complex diseases.