Faster and more diverse de novo molecular optimization with double-loop reinforcement learning using augmented SMILES

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular generation

Background:

• Generative deep learning and reinforcement learning (RL) models can design novel molecules with specific properties.
• Current methods face challenges due to computationally intensive scoring functions, slowing down the RL optimization process.
• Efficient molecule generation is crucial for accelerating drug discovery and material science applications.

Purpose of the Study:

• To enhance the efficiency and speed of molecule generation using generative deep learning and RL.
• To address the computational bottlenecks associated with scoring functions in RL-based molecule design.
• To improve the diversity and similarity of generated molecules to known ligands.

Main Methods:

• Proposed a double-loop reinforcement learning framework.
• Incorporated simplified molecular line entry system (SMILES) augmentation within an inner loop.
• Reused scoring calculations and introduced non-canonical SMILES for additional RL rounds.

Main Results:

• The double-loop RL approach significantly speeds up the molecule generation and optimization process.
• Optimal performance was achieved with 5-10 augmentation repetitions.
• The method led to increased compound diversity, improved sampling reproducibility, and generation of molecules similar to known ligands.

Conclusions:

• Double-loop reinforcement learning with SMILES augmentation offers a more efficient and robust method for de novo molecule design.
• This approach effectively mitigates computational costs and enhances the quality of generated molecules for drug discovery.
• The strategy provides protection against mode collapse and increases the overall utility of generative models in chemical research.