ASCENT: an active transfer learning paradigm for efficient drug-target interaction prediction.

Area of Science:

• Computational chemistry
• Pharmacology
• Machine learning

Background:

• Deep learning advances DTI prediction but is limited by fixed, small datasets.
• Current models struggle with generalizability across vast chemical spaces and unseen drugs/targets.
• This necessitates methods that can adapt to new data and improve predictive performance.

Purpose of the Study:

• Introduce ASCENT, an active transfer learning framework for DTI prediction.
• Enhance model generalizability and reduce annotation costs.
• Facilitate exploration of chemical diversity for drug discovery and repurposing.

Main Methods:

• Utilizes an adaptive active learning strategy for dynamic dataset expansion.
• Incorporates an entropy-based adversarial method to align feature spaces between domains.
• Employs model performance to select representative and uncertain samples for annotation.

Main Results:

• ASCENT demonstrates superior performance in cross-domain DTI prediction.
• Achieves desired predictive accuracy while reducing annotation costs by approximately 20%.
• Case studies highlight potential in novel drug discovery and repurposing.

Conclusions:

• ASCENT is a valuable advancement for DTI prediction, improving efficiency and accuracy.
• The framework effectively captures patterns across large chemical spaces.
• Supports accelerated drug development and offers new perspectives on DTI prediction.