Quantum Chemistry-Driven Molecular Inverse Design of Stable Isomers with Data-Free Reinforcement Learning.

Area of Science:

• Computational Chemistry
• Artificial Intelligence in Chemistry
• Molecular Modeling

Background:

• Molecular inverse design (ID) is a significant challenge in chemistry.
• Current machine learning (ML) and artificial intelligence (AI) methods often require large pre-trained datasets, introducing bias.
• A data-free approach is needed for unbiased molecular generation.

Purpose of the Study:

• To present PROTEUS, a data-free generative AI model for de novo molecular design.
• To enable unbiased molecular design from first principles using quantum mechanical calculations.
• To optimize specific chemical properties of generated molecules.

Main Methods:

• Integration of reinforcement learning with on-the-fly quantum mechanical calculations.
• Utilizing a custom syntax and hierarchical learning architecture for chemical space navigation.
• A data-free approach, avoiding reliance on pre-existing large datasets.

Main Results:

• Demonstrated efficiency in solving complex molecular design tasks, such as maximizing isomerization energy gaps in styrene derivatives.
• Showcased robustness and flexibility in exploring chemical spaces and exploiting chemical rewards.
• Successfully performed inverse design for problems with known solutions.

Conclusions:

• PROTEUS offers a quantum chemistry-driven, unbiased strategy for molecular design.
• The framework is flexible and scalable for addressing diverse chemical design challenges.
• Opens new avenues for de novo molecular discovery without prior data bias.