Uncertainty-aware hybrid deep generative framework for robust and explainable drug discovery.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Machine learning for molecular design

Background:

• Traditional drug discovery is slow and costly due to vast chemical spaces and inherent uncertainties.
• Current AI/ML methods face challenges like high attrition rates and optimizing multiple molecular properties concurrently.
• Need for robust and interpretable methods to improve early-stage drug discovery efficiency.

Purpose of the Study:

• To develop a hybrid computational framework for generating novel drug molecules with desired pharmacological properties.
• To address limitations of existing methods by integrating uncertainty awareness and explainability.
• To enhance the reliability and reduce false positives in AI-driven drug candidate generation.

Main Methods:

• Integration of Graph Convolutional Networks, Variational Autoencoders, and Uncertainty Aware Adaptive Multi-Objective Optimization-based Reinforcement Learning (UAAMOO-RL).
• Framework trained and validated on the ChEMBL dataset (1.5 million bioactive molecules).
• Incorporation of uncertainty awareness and explainability for robust molecular generation.

Main Results:

• Achieved a 88.8% Quantitative Drug-likeness (QED) pass rate on the ChEMBL dataset at a threshold of 0.60.
• Outperformed state-of-the-art techniques in molecule generation while preserving diversity, uniqueness, and validity.
• Demonstrated significant improvements on Zinc250k (82.7%) and PDBbind (85.9%) datasets.

Conclusions:

• The novel hybrid AI framework effectively generates reliable and interpretable drug molecules.
• This approach significantly enhances early-stage drug discovery by reducing the need for extensive experimental screening.
• The method offers a robust solution to key challenges in modern drug development, improving efficiency and success rates.