Enhancing reaction-based de novo design using a multi-label reaction class recommender.

Area of Science:

• Computational chemistry
• Drug discovery
• Chemical synthesis

Background:

• Reaction-based de novo design generates novel molecules using known reaction rules.
• Existing methods using reaction vectors overlook competing functional groups, limiting synthetic accessibility.
• There is a need to improve the prediction of synthesizable molecules in computational design.

Purpose of the Study:

• To develop a Reaction Class Recommender to enhance reaction-based de novo design.
• To filter reaction vectors and reduce the search space of computationally designed molecules.
• To improve the synthetic accessibility of generated chemical structures.

Main Methods:

• Developed a Reaction Class Recommender as a filter for reaction vectors.
• Integrated the recommender into a de novo molecular design framework.
• Validated the approach using external medicinal chemistry literature and simulated design experiments.

Main Results:

• The Reaction Class Recommender significantly reduced the number of explored solutions in de novo design.
• The method preserved the ability to find relevant molecular solutions.
• The global synthetic accessibility of designed molecules was demonstrably increased.

Conclusions:

• The Reaction Class Recommender effectively enhances reaction-based de novo design by improving synthetic accessibility.
• This tool addresses limitations of previous reaction vector methods by considering reaction context.
• The approach offers a promising strategy for more efficient and successful in silico drug discovery.