Drug Target Identification with Machine Learning: How to Choose Negative Examples.

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Identifying protein targets for drug molecules is crucial in drug discovery.
• Machine learning accelerates target prediction but suffers from biased training data, leading to false positives.
• High false positive rates increase experimental costs and time.

Purpose of the Study:

• To develop a new scheme for selecting negative examples in drug-target interaction databases.
• To minimize false positives in machine learning-based target prediction.
• To improve the efficiency of experimental validation campaigns.

Main Methods:

• Proposed a novel negative example selection scheme ensuring equal representation of proteins and drugs in positive and negative sets.
• Artificially reproduced target identification processes for three specific drugs and 200 approved drugs.
• Trained machine learning models using the proposed scheme.

Main Results:

• The new negative example selection scheme significantly reduced false positive predictions.
• Improved the ranking of true drug targets among predicted targets.
• Demonstrated enhanced prediction accuracy for both specific drug examples and a larger set of approved drugs.

Conclusions:

• The proposed method effectively corrects statistical bias in Drug-Target Interactions databases.
• Reduces the number of false positive predictions, leading to more efficient and cost-effective drug discovery.
• Enhances the reliability of machine learning models for predicting drug-protein interactions.