Stereochemistry-aware string-based molecular generation.

Area of Science:

• Computational chemistry
• Cheminformatics
• Artificial intelligence in drug discovery

Background:

• Molecular generative modeling is vital for drug discovery and materials design.
• Incorporating stereochemical information is crucial for accurate molecular representation.
• Current models often overlook stereochemistry, limiting their applicability.

Purpose of the Study:

• To compare stereochemistry-aware and unaware molecular generative models.
• To evaluate the impact of stereochemistry on model performance in computational drug discovery.
• To provide guidance on selecting appropriate generative models for specific tasks.

Main Methods:

• Utilized string-based generative approaches with genetic algorithms and reinforcement learning.
• Developed novel benchmarks to specifically assess stereochemistry-aware generative modeling.
• Compared performance across various stereochemistry-sensitive and insensitive tasks.

Main Results:

• Stereochemistry-aware models performed comparably or better than conventional models on stereochemistry-sensitive tasks.
• Conventional models showed advantages in tasks where stereochemistry is less critical.
• Stereochemistry-aware models faced challenges navigating complex chemical spaces in certain scenarios.

Conclusions:

• Stereochemistry-aware models offer significant advantages for tasks where stereochemistry is important.
• The complexity of chemical space navigated by stereochemistry-aware models presents a trade-off.
• Choosing generative models should align with the stereochemical requirements of the specific application.