Introduction
Retrosynthesis is a fundamental task in drug discovery, requiring the identification of chemical precursors (reactants) to synthesize a target molecule. Traditional AI models for retrosynthesis are often limited by dataset incompleteness. Training on benchmarks like USPTO-50k, which capture only a fraction of possible chemical space, forces models to prioritize "memorizing" specific dataset entries over discovering chemically valid alternatives. Consequently, creative but valid synthesis routes are often penalized by standard top-k accuracy metrics.
The authors argue for a shift toward prioritizing feasibility and diversity. In pharmaceutical R&D, missing a viable synthesis route (a false negative) is significantly more costly than a failed lab experiment (a false positive). RetroGFN addresses this by focusing on the "hidden" feasible space of chemistry.
Methods
RetroGFN is the first application of Generative Flow Networks (GFlowNets) to retrosynthesis. Unlike standard models that maximize likelihood, GFlowNets sample solutions with a probability proportional to a reward function, enabling the discovery of multiple "modes" or diverse solutions in complex chemical spaces.
The model constructs a reaction template through a structured three-phase generation process:
- Reaction Center Identification: Pinpointing the site of transformation on the product.
- Reactant Pattern Selection: Choosing structural motifs for the required reactants.
- Atom Mapping: Establishing precise correspondences between product and reactant atoms.
A key innovation is the use of Feasibility Guidance. A pre-trained forward reaction model acts as a reward proxy during training. If the forward model predicts that the proposed reactants will successfully yield the target product, RetroGFN receives a high reward. This allows the model to explore and validate reactions beyond the scope of the original training set.
Results
RetroGFN was evaluated on the USPTO-50k and USPTO-MIT datasets. While standard top-k accuracy measures dataset matching, the authors emphasize Round-Trip Accuracy (RTA), which uses a forward model to verify the chemical validity of proposed routes.
On USPTO-50k, RetroGFN achieved a significant lead in RTA for $k > 3$, reaching 63.6% at $k=50$, compared to approximately 54% for the next best model. This performance held steady on the larger USPTO-MIT dataset, demonstrating strong generalization. Furthermore, diversity analysis using molecular scaffolds confirmed that RetroGFN identifies fundamentally different synthetic pathways rather than minor variations, providing more options for medicinal chemists.
Discussion and Conclusion
RetroGFN demonstrates that optimizing for chemical feasibility rather than dataset replication leads to more robust and diverse retrosynthetic tools. While the model excels at finding sets of feasible reactions, future work will focus on improving the ranking of these results at low $k$ values and moving toward fully template-free architectures. Despite a moderate inference time of approximately 3.7 seconds per molecule, its ability to minimize expensive false negatives makes it a powerful asset for real-world pharmaceutical research and development.
Summary updated