Improving drug discovery with a hybrid deep generative model using reinforcement learning trained on a Bayesian...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Generative models are key for designing new pharmaceuticals.
• Limited experimental data often hinders generative model performance, leading to poor results or overfitting.
• Existing methods struggle to generate diverse, novel chemotypes.

Purpose of the Study:

• To reduce data dependency in generative molecular design for new chemotypes.
• To enhance the diversity and performance of generated molecules using reinforcement learning.
• To accelerate the discovery of potential therapeutic agents.

Main Methods:

• Incorporated docking scores into the reward function of a deep generative model.
• Used a machine learning-based Bayesian regression model for approximate docking scores.
• Combined limited drug activity data with approximate docking scores for initial training.
• Final evaluation of high-scoring compounds via full docking simulations.
• Employed reinforcement learning to infer molecular-receptor interactions.

Main Results:

• Achieved 10-20% improvement in docking scores for generated molecules compared to similar-sized molecules.
• Demonstrated a 130x speed increase over docking-only approaches on a GPU workstation.
• Showcased correlation between higher docking scores and poses similar to known inhibitors.
• Observed MM-GBSA binding energies comparable to known DDR1 inhibitors.

Conclusions:

• The developed method effectively reduces data dependency for generating novel chemotypes.
• The combination of learned molecular representations and feature-based docking regression enables efficient learning of receptor interactions.
• This approach is a powerful tool for discovering new chemotypes with therapeutic potential.