Circumventing the synthesizability problem in generative molecular design.

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Drug discovery

Background:

• Generative structure-based drug design (SBDD) models accelerate the discovery of novel drug candidates.
• A key limitation of current SBDD models is the generation of compounds with poor synthesizability, hindering practical drug design.
• Existing virtual ligand screening (VLS) methods face challenges with increasing chemical space and computational demands.

Purpose of the Study:

• To address the synthesizability challenge in generative SBDD.
• To introduce a novel model-guided virtual screening (MGVS) pipeline.
• To improve the efficiency and practicality of drug candidate identification.

Main Methods:

• Developed an MGVS pipeline integrating generative SBDD models with chemical similarity search.
• Utilized MGVS to identify synthesizable analogs of generated compounds within ultra-large compound databases.
• Evaluated MGVS performance against standard VLS across diverse protein targets and SBDD models.

Main Results:

• MGVS reliably identified synthesizable analogs with comparable or superior docking scores and binding poses.
• MGVS demonstrated a consistent 25x improvement in screening efficiency compared to standard VLS.
• The approach proved effective across multiple SBDD models and various protein targets.

Conclusions:

• The MGVS pipeline effectively overcomes the synthesizability limitations of generative SBDD.
• MGVS offers a significant advancement in screening efficiency, crucial for navigating expanding chemical spaces.
• This approach is vital for accelerating practical drug discovery in the face of growing computational challenges.