TL;DR
This paper introduces SynTwins, a novel framework that bridges the gap between AI-designed molecules and real-world synthesis. By mimicking how chemists find accessible alternatives, SynTwins generates molecules that are not only structurally similar to desired targets but also practically synthesizable, outperforming current AI methods.
What Problem Does This Paper Address?
The dream of AI-driven drug and materials discovery is often hampered by a critical bottleneck: synthetic feasibility. While generative AI can rapidly propose novel molecules with excellent predicted properties, many of these proposed structures are incredibly difficult, if not impossible, to synthesize using known chemical reactions and available building blocks. This disconnect means that promising AI-generated candidates often never make it to the lab, slowing down the discovery process. Existing methods for predicting synthetic accessibility can help, but they don't fundamentally solve the problem of generating synthesizable analogs that are also structurally similar to the original, desirable molecule.
Methods
SynTwins tackles this challenge by emulating the intuitive, three-step process that experienced chemists use when faced with a difficult-to-synthesize molecule:
- Retrosynthesis: The target molecule is broken down step-by-step using a set of pre-defined "retro-reaction templates" (essentially, the reverse of known chemical reactions). This creates a "synthesis tree" exploring potential precursors.
- Building Block Search: For each precursor identified in the retrosynthesis, SynTwins searches a database of commercially available molecules for those that are structurally similar. This search uses a k-nearest neighbor (kNN) algorithm based on chemical fingerprints (ECFP).
- Virtual Synthesis: Using "forward-reaction templates" (the original reactions), SynTwins virtually combines the identified similar building blocks to create potential molecular analogs.
Crucially, SynTwins uses a search-based algorithm rather than a purely data-driven, stochastic generator, allowing for more interpretable and controlled analog generation.
Datasets
- Building Blocks: A dataset of 150,560 commercially available molecules from the Enamine Building Block catalogue was used.
- Reaction Templates: 101 forward-reaction templates were compiled from existing literature (Hartenfeller et al. and Button et al.), from which 1,163 retro-reaction templates were derived.
- Evaluation Datasets:
- Virtually synthesized products (1,000)
- ChEMBL molecules (1,000)
- USPTO dataset (170)
- FDA-approved drugs (100)
Key Insights and Ideas
- Mimicking Chemists is Powerful: The core idea that emulating human chemists' strategies (retrosynthesis, then finding similar building blocks) leads to better, more practical results than purely data-driven AI is a significant insight.
- Interpretability Matters: Unlike black-box ML models, SynTwins' retrosynthesis tree provides a clear, interpretable pathway. This allows researchers to understand why certain analogs are generated and, importantly, identify missing building blocks or reactions, guiding future resource acquisition.
- Hybrid Approaches are Key: Integrating SynTwins into existing molecule optimization pipelines (like REINVENT and GraphGA) shows promise. While it might slightly reduce the "ideal" property scores, it ensures the final optimized molecules are actually synthesizable, a crucial step for real-world application.
- Synthetic Accessibility as a Constraint: The framework demonstrates that explicitly incorporating synthetic constraints early in the design process can lead to more robust and ultimately useful molecular discoveries.
Key Results
- Superior Performance: SynTwins consistently outperformed state-of-the-art ML-based models (ChemProjector and SynFormer) across all tested datasets in both exact-match rate and top-k average structural similarity.
- High Structural Similarity: Even when an exact match wasn't achieved, SynTwins generated analogs that were highly structurally similar to the target molecules. For example, on the USPTO dataset, SynTwins achieved an 1.8% exact-match rate but a top-5 average similarity of 0.4941, significantly higher than baselines.
- Real-World Data Advantage: SynTwins showed particular strength with real-world datasets like USPTO and FDA-approved drugs, where other methods struggled to recover original molecules.
- Improved Synthesizability: When integrated into optimization workflows, SynTwins-based optimizers produced molecules with significantly lower synthetic accessibility scores (easier to synthesize) with only a minimal drop in property scores compared to unconstrained optimizers.
Limitations and Future Work
- Dependency on Available Resources: SynTwins' performance is inherently limited by the provided set of reaction templates and building blocks. If a target molecule requires entirely novel chemistry or unavailable precursors, SynTwins may struggle.
- Retrosynthesis Depth: The default retrosynthesis depth (dmax=3) might not be sufficient for highly complex molecules.
- Future Work: The authors suggest that SynTwins' interpretability can guide the strategic expansion of chemistry resources (building blocks and reactions) to improve coverage. Integrating SynTwins into more advanced molecular optimization frameworks is also a promising direction. The framework's ability to identify missing chemistry is a key advantage for iterative discovery workflows.
Summary updated