MOLER: Incorporate Molecule-Level Reward to Enhance Deep Generative Model for Molecule Optimization

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery

Background:

• Deep generative models are successful in molecular optimization but often produce molecules with undesirable deviations in similarity and size.
• Current models focus on substructure generation, neglecting the global molecular context.

Purpose of the Study:

• To introduce Molecule-Level Reward functions (MOLER) for improved molecular optimization.
• To ensure generated molecules maintain similarity and comparable size to the input molecule.

Main Methods:

• MOLER integrates molecule-level similarity and size constraints into deep generative models.
• Policy gradient techniques are employed for efficient optimization of reward-based objectives.
• The method is designed to be compatible with various deep generative architectures.

Main Results:

• MOLER demonstrated significant improvements in molecular optimization tasks.
• Achieved up to 20.2% relative improvement in success rate compared to baseline methods.
• Successfully enhanced properties such as Quantitative Estimate of Drug-likeness (QED), D2 receptor affinity (DRD2), and LogP.

Conclusions:

• MOLER effectively addresses the limitations of substructure-level generation in deep molecular optimization.
• The proposed approach enhances molecular similarity and size control, leading to better chemical properties.
• MOLER offers a computationally efficient and versatile method for advancing drug discovery and molecular design.