Abstract
Current de novo drug design often produces molecules that are theoretically synthesizable but practically inaccessible due to limited local building block inventories. This study introduces a "fast-to-adapt" workflow that prioritizes in-house synthesizability by training generative AI on specific, limited chemical stocks. Using AiZynthFinder to benchmark synthesis success, the authors developed an XGBoost-based proxy scorer to guide the DrugEx v3 reinforcement learning generator. The approach was validated by identifying a novel monoglyceride lipase (MGLL) inhibitor with an IC50 of 1 µM, synthesized using only 6,000 locally available building blocks. This method bridges the gap between AI-driven design and laboratory reality, offering a pragmatic solution for small-to-medium-sized research facilities.
Introduction
The "Design-Make-Test-Analyze" (DMTA) cycle is a cornerstone of drug discovery. While AI has significantly accelerated the "Design" phase, a persistent "retrosynthesis gap" remains. Most generative models rely on general synthesizability scores (e.g., SAScore, RAscore) that assume access to the millions of building blocks available globally. However, practical constraints—such as budget, shipping delays, and existing inventory—often limit researchers to a much smaller subset of "in-house" materials. When AI proposes molecules requiring unavailable precursors, the DMTA cycle stalls. This work addresses this challenge by forcing generative models to "generate what you can make" based on local availability.
Methods
The researchers implemented a multi-stage pipeline to align AI generation with local synthesis capabilities. First, they used AiZynthFinder, an open-source Computer-Aided Synthesis Planning (CASP) tool, to evaluate synthesis success rates. They compared a massive commercial library (Zinc, 17.4M compounds) against a university stock (Led3, 5,955 compounds).
To enable real-time evaluation during molecule generation, they trained an XGBoost classifier as a "proxy" for the computationally expensive CASP process. This model predicts whether a molecule can be synthesized using only the 6,000 in-house blocks. This scorer was then integrated into DrugEx v3, a reinforcement learning-based generator, for multi-objective optimization targeting Monoglyceride Lipase (MGLL) activity and in-house synthesizability.
Results
A key finding of the study was the resilience of synthesis planning to library size. Reducing the building block library by a factor of 3,000 (from Zinc to Led3) resulted in only a ~12% decrease in the fraction of molecules successfully solved by the CASP tool across various datasets.
However, this reduction in precursor diversity necessitated more complex synthetic routes. On average, molecules synthesized with the smaller Led3 library required 6.6 reaction steps, compared to 4.5 steps for the Zinc library.
The workflow successfully generated over 20,000 candidates predicted to be both active against MGLL and synthesizable in-house. Experimental validation of three top candidates led to the discovery of Compound 1, which exhibited an IC50 of 1 µM. Two other compounds showed lower potency (~100 µM), but all three were successfully synthesized following the AI-suggested routes.
Discussion
The ability to retrain the in-house XGBoost scorer in less than a day allows for rapid adaptation to changing inventories. While the current model tends to favor known molecular scaffolds, it ensures that generated leads are immediately actionable. Future work may involve 3D activity modeling (e.g., docking) to improve lead quality and using the tool to identify optimal building blocks for strategic inventory expansion. This approach represents a significant step toward making AI-driven drug design a practical tool for every laboratory.
Summary updated