Diversity oriented Deep Reinforcement Learning for targeted molecule generation.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Artificial Intelligence

Background:

• Identifying new drug candidates is a complex and time-consuming process.
• Deep learning offers potential for accelerating drug discovery through computational methods.

Purpose of the Study:

• To develop a deep learning framework for the computational generation of molecules with specific biological properties.
• To introduce a novel exploratory strategy for reinforcement learning to enhance novelty in generated compounds.

Main Methods:

• A two-part deep neural network framework: a Generator for molecule construction using SMILES notation and a Predictor for evaluating biological affinity.
• Reinforcement learning optimization of the Generator, incorporating an innovative strategy with two interchangeable Generators to balance exploration and exploitation.
• Training the model to optimize partition coefficient and inhibitory power against Adenosine [Formula: see text] and [Formula: see text] opioid receptors.

Main Results:

• The deep learning model successfully generated molecules with desired properties, adjusting them towards the target direction.
• The novel training strategy yielded promising sets of unique and diverse molecules.
• Demonstrated effectiveness in designing molecules with optimized partition coefficients and high inhibitory power.

Conclusions:

• Deep learning, particularly with the novel exploratory strategy, can effectively streamline drug discovery by generating novel and diverse molecular candidates.
• The proposed method shows promise for identifying unique compounds with specific biological activities, addressing a key goal in drug development.