Evaluation of reinforcement learning in transformer-based molecular design

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery

Background:

• Designing novel drug compounds with specific properties is crucial in early drug discovery.
• Transformer models can generate molecules similar to a starting compound but lack flexibility for user-defined properties.

Purpose of the Study:

• To evaluate the impact of reinforcement learning on transformer-based molecular generative models.
• To assess the ability of reinforcement learning to steer molecular generation towards user-specific property profiles.

Main Methods:

• Utilized transformer-based deep learning models pre-trained on molecular data.
• Applied reinforcement learning as a tuning phase to guide the generative model.
• Evaluated performance on molecular optimization and scaffold discovery tasks.

Main Results:

• Reinforcement learning guided transformer models to generate more compounds of interest.
• Demonstrated improved flexibility for optimizing user-specific property profiles.
• Investigated the influence of pre-trained models, learning steps, and learning rates.

Conclusions:

• Reinforcement learning effectively enhances transformer-based generative models for drug discovery.
• This approach facilitates multi-parameter optimization and expands the chemical space exploration for novel compounds.