RLMolLM: Reinforcement Learning-Enhanced Language Model Framework for Inverse Molecular Design.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery

Background:

• Inverse molecular design is challenged by vast chemical space and complex property requirements.
• Existing language models struggle with molecular validity, multi-property optimization, and structural constraints.

Purpose of the Study:

• To present RLMolLM, a reinforcement learning framework integrating Proximal Policy Optimization (PPO) with genetic algorithms.
• To address limitations in molecular generation, including validity, multi-property optimization, and scaffold preservation.

Main Methods:

• RLMolLM framework combines Proximal Policy Optimization (PPO) with genetic algorithms.
• Optimizes quantitative estimates of drug-likeness (QED), synthetic accessibility (SA), and ADMET properties.
• Enables scaffold-constrained generation, preserving specific substructures.

Main Results:

• Achieved superior QED scores across GDB13, Moses, and Zinc datasets, with up to 31% improvement.
• Demonstrated substantial ADMET improvements, including a 4.5-fold reduction in hERG toxicity and enhanced Caco-2 permeability.
• Significantly improved molecular validity and property optimization under structural constraints while preserving scaffolds.

Conclusions:

• RLMolLM offers a versatile solution for pharmaceutical and materials molecular design.
• The framework effectively integrates reinforcement learning and genetic algorithms for multi-property optimization and scaffold preservation.
• Advances inverse molecular design by overcoming key limitations of current generative models.