Integrating synthetic accessibility with AI-based generative drug design.

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Artificial intelligence in drug discovery

Background:

• Generative models are crucial for de novo drug design, but often fail to consider molecular synthesizability.
• Ensuring synthesized molecules is a practical necessity for generative methods in drug discovery.
• Existing methods for estimating molecular synthesizability lack consensus.

Purpose of the Study:

• Introduce the Retro-Score (RScore) for assessing molecular synthetic accessibility.
• Validate RScore's reliability against expert chemist assessments.
• Develop a pipeline for generating synthesizable drug molecules.
• Investigate the use of RScore and its neural network-learned variant (RSPred) as constraints in molecular generation.

Main Methods:

• Developed the Retro-Score (RScore) using data-driven retrosynthetic analysis via Spaya software and API.
• Compared RScore against a "chemist score" for validation.
• Implemented a generation pipeline incorporating RScore as a constraint.
• Trained a Neural Network to learn RScore, creating RSPred.
• Conducted experiments comparing molecular generator outputs under various constraints.

Main Results:

• RScore demonstrated reliability as a synthetic accessibility score.
• Using RScore or RSPred as a constraint improved molecular generator outputs.
• Generated molecules exhibited enhanced synthesizability and diversity.
• The open-source Python code for scores and experiments is available.

Conclusions:

• The Retro-Score provides a reliable measure of molecular synthetic accessibility.
• Integrating RScore or RSPred into generative models significantly enhances the production of synthesizable and diverse drug candidates.
• This approach addresses a critical gap in practical de novo drug design.