DrugProtAI: A machine learning-driven approach for predicting protein druggability through feature engineering and...

Area of Science:

• Biochemistry
• Computational Biology
• Pharmacology

Background:

• Drug development faces high failure rates, often due to poor target selection.
• Current computational methods for predicting protein druggability are limited in scope and features.
• Precision and sensitivity in identifying viable drug targets are critical for successful clinical research.

Purpose of the Study:

• To develop and validate DrugProtAI, a novel computational tool for predicting human protein druggability.
• To improve the accuracy and scope of drug target identification in early-stage drug discovery.
• To provide a freely accessible resource for researchers to assess protein druggability.

Main Methods:

• Implemented a partitioning-based method trained on the entire human proteome.
• Utilized a comprehensive feature set including 183 biophysical, sequence- and non-sequence-derived properties.
• Evaluated machine learning algorithms, with Random Forest and XGBoost showing optimal performance.

Main Results:

• Achieved a median Area Under Precision-Recall Curve (AUC) of 0.87 for target prediction.
• Validated the model on a blinded dataset of recently approved drug targets.
• Identified key predictors contributing to protein druggability, aiding target selection.

Conclusions:

• DrugProtAI offers a significant advancement in predicting protein druggability for human proteins.
• The tool's comprehensive approach and high performance can help mitigate drug development failures.
• DrugProtAI is accessible online, empowering researchers with data-driven target selection capabilities.