Streamlining pipeline efficiency: a novel model-agnostic technique for accelerating conditional generative and...

Area of Science:

• Computational chemistry
• Drug discovery and development
• Machine learning in pharmacology

Background:

• Identifying novel therapeutic agents for critical diseases is a major challenge.
• Current methods like high-throughput screening and virtual screening are often inefficient and time-consuming.
• Machine learning models can emulate docking functions for faster drug candidate identification but face data scarcity issues.

Purpose of the Study:

• To propose an active learning-based model to enhance molecule generation architectures.
• To address model degradation in generative pipelines when using pre-trained models.
• To improve the efficiency and accuracy of identifying drug-like molecules with high binding affinity.

Main Methods:

• Developed an active learning framework integrated with molecule generation architectures.
• Utilized uncertainty sampling to dynamically learn from generated molecules.
• Focused on molecules sampled from diverse regions of the chemical space.

Main Results:

• The proposed active learning model significantly improves molecule generation.
• Achieved a 70% runtime improvement compared to baseline models.
• Required labeling of only 30% of molecules compared to the baseline oracle, while maintaining high binding affinity.

Conclusions:

• Active learning effectively supplements generative models for drug discovery.
• The approach mitigates model degradation in data-scarce regions of chemical space.
• Offers a faster and more efficient method for identifying potential therapeutic agents.