TL;DR
This paper introduces an "in-house" de novo drug design workflow that aligns AI molecule generation with the actual, limited chemical inventory of a specific laboratory. By training a fast machine learning model to predict local synthesizability, the researchers successfully generated, synthesized, and experimentally validated a novel active inhibitor for monoglyceride lipase (MGLL) using only available lab resources.
What Problem Does This Paper Address?
Artificial intelligence has revolutionized de novo drug design, allowing algorithms to propose thousands of novel molecules optimized for specific biological targets. However, a glaring disconnect remains between computational generation and experimental reality: AI often designs "virtual" molecules that are exceptionally difficult or impossible to synthesize in a standard lab.
While Computer-Aided Synthesis Planning (CASP) tools help evaluate whether a molecule can be made, they typically assume access to "infinite" commercial libraries (e.g., 17+ million building blocks). For small academic or industrial laboratories operating on limited budgets and strict timelines, relying on massive external vendor catalogs is impractical. Furthermore, running full CASP simulations inside an AI optimization loop is computationally prohibitive. The authors address this by asking: Can we train an AI to generate molecules tailored specifically to what a local lab can actually make right now?
Methods
The authors developed an end-to-end workflow that shifts synthesizability from a global assumption to a local constraint:
- In-House CASP Evaluation: They used
AiZynthFinder, an open-source retrosynthesis tool, to benchmark how well a small laboratory inventory could replicate the performance of a massive commercial library. - Fast Synthesizability Scoring: Because full CASP runs take minutes to hours per molecule, they trained a rapid proxy model. Using an
XGBoostbinary classifier trained on Morgan fingerprints and basic chemical properties, they created a score capable of predicting whether a molecule is synthesizable using only the local inventory in milliseconds. - Multi-Objective De Novo Generation: They integrated this local synthesizability score into
DrugEx v3, a reinforcement learning-based molecular generator. The generator optimized molecules simultaneously for two objectives: predicted biological activity against MGLL (via a QSAR classification model) and in-house synthesizability. - Experimental Validation: To prove the workflow's real-world utility, domain experts selected three generated candidates, synthesized them using the AI-suggested routes, and evaluated their inhibitory activity in vitro.
Datasets
The study leveraged several key datasets for training and evaluation:
- Led3: A local inventory of 5,955 building blocks from Leiden University, representing the resource-limited lab setting.
- Zinc: A baseline library of 17.4 million commercially available building blocks representing "infinite" availability.
- Caspyrus (1k, 10k, 20k, 50k subsets): Clustered subsets derived from the Papyrus bioactivity dataset, used to benchmark CASP performance and train the synthesizability scores.
- ChEMBL200k: A sample of 200,000 drug-like molecules from ChEMBL used as an independent test set to evaluate model generalization.
- MGLL Ligands: 700 known monoglyceride lipase ligands extracted from Papyrus, used to fine-tune the molecular generator and train the QSAR activity model.
Key Insights and Ideas
- The 3000x Library Reduction Myth: Shrinking the building block library by a factor of 3,000 (from 17.4 million down to ~6,000) only resulted in a 12% drop in synthesis success rate (solvability) across drug-like spaces. However, the trade-off is that the local synthesis routes are, on average, two reaction steps longer.
- Local vs. Global Chemical Spaces: Optimizing a generator using a general synthesizability score versus an in-house score drives the AI into completely different chemical neighborhoods. As shown in Figure 5, you cannot simply optimize for general synthesizability and expect the results to easily map to your local inventory later.
- Data Efficiency: An accurate local synthesizability score requires a dataset of only 10,000 molecules for training. This means a lab can completely retrain its custom synthesizability model within a single day whenever its physical inventory changes.
Key Results
- CASP Performance: The local
Led3inventory achieved a ~60% solvability rate on drug-like datasets compared to ~70% for the massiveZinclibrary (Figure 1).
- Score Accuracy: The trained XGBoost model achieved an F1-score above 0.8 on independent test sets when trained on 10,000 or more molecules, proving highly predictive of full CASP outcomes (Figure 3).
- In Vitro Validation: Out of the three generated molecules synthesized via the algorithm's exact 1-to-2 step local pathways, Compound 1 demonstrated clear, confirmed activity against MGLL with an $IC_{50}$ of $1\ \mu\text{M}$.
Limitations and Future Work
While the proof-of-concept was successful, the authors noted a common pitfall in AI drug design: limited scaffold novelty. Because the generator relied on a QSAR model trained on known active compounds, it tended to reuse familiar core scaffolds and primarily modify side chains.
Additionally, to make experimental validation feasible, the researchers filtered for very short synthesis routes (1-2 steps), which likely excluded more complex, potentially more potent molecules.
Future Directions: The authors suggest replacing the QSAR model with structure-based methods like molecular docking or shape-based pharmacophores to encourage the generation of more creative core structures. They also propose using this framework to algorithmically optimize the ideal composition of a lab's physical building block collection to maximize the accessible chemical space.
Summary updated