Utilizing random Forest QSAR models with optimized parameters for target identification and its application to...

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Identifying drug targets is crucial for understanding drug mechanisms like target deconvolution and polypharmacology.
• Traditional target identification methods are time-consuming and expensive, driving the need for in-silico approaches.
• Structure-activity relationships (SARs) offer a feasible in-silico alternative but face data dependency challenges.

Purpose of the Study:

• To develop a robust in-silico model for predicting ligand-target interactions and ranking potential drug targets.
• To overcome the limitations of traditional SAR approaches by improving data handling and model performance.

Main Methods:

• A ligand-based virtual screening model was constructed using 1121 target SAR models.
• The random forest algorithm was employed to build individual target SAR models.
• Model performance was rigorously evaluated using ROC curves, mean scores, cross-validation, and recall rates for top-k targets.

Main Results:

• The benchmark model achieved recall rates of 67.6% for top-11 and 73.9% for top-33 targets in external validation.
• The developed models demonstrate high accuracy in predicting ligand activity and ranking candidate targets.
• A publicly accessible website (http://rfqsar.kaist.ac.kr) was launched for user-friendly target searching.

Conclusions:

• The developed target models offer a unified scoring scheme for predicting ligand activity and ranking targets.
• Scores are probability-fitted, enabling users to estimate the likelihood of active ligand-target interactions.
• The user-friendly web interface provides intuitive access to valuable cross-referenced information.