Synthesizable Molecular Generation via Soft-constrained GFlowNets with Rich Chemical Priors

The application of generative models for experimental drug discovery campaigns is severely limited by the difficulty of designing molecules de novo that can be synthesized in practice. Previous works have leveraged Generative Flow Networks (GFlowNets) to impose hard synthesizability constraints through the design of state and action spaces based on predefined reaction templates and building blocks. Despite the promising prospects of this approach, it currently lacks flexibility and scalability. As an alternative, we propose S3-GFN, which generates synthesizable SMILES molecules via simple soft regularization of a sequence-based GFlowNet. Our approach leverages rich molecular priors learned from large-scale SMILES corpora to steer molecular generation towards high-reward, synthesizable chemical spaces. The model induces constraints through off-policy replay training with a contrastive learning signal based on separate buffers of synthesizable and unsynthesizable samples. Our experiments show that S3-GFN learns to generate synthesizable molecules ($\geq 95$%) with higher rewards in diverse tasks.

Deep generative models can help propose new molecules for drug discovery, but promising candidates also need to be practical to make and test in the lab. Existing methods ensure this by building molecules from predefined chemical reactions, which improves synthesizability but ties the generator to a fixed reaction library. We propose S3-GFN, a more flexible approach that generates molecules as text-like chemical strings using a model initialized from large collections of known molecules. Rather than forcing the model to follow predefined synthesis routes, we train it to prefer molecules that are likely to be synthesizable and avoid those that are not. This preserves the flexibility of string-based generation while steering the model toward candidates that are both high-scoring for the target property and more likely to be made in practice. Our experiments show that this is especially useful when synthesizability requirements change or when training samples are limited.