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

TL;DR

The paper tackles the challenge of designing synthesizable de novo molecules by introducing S3-GFN, a soft-constrained GFlowNet that leverages a pretrained SMILES prior to steer generation toward synthesizable chemical spaces. It decouples synthesizability constraints from rewards via a distributional regularization implemented through two replay buffers and a contrastive auxiliary loss, enabling flexible, off-policy learning. Empirical results show synthesizability rates exceeding 95% and improved task rewards across both target-fold and structure-based drug discovery tasks, often outperforming reaction-based GFlowNets. The approach also demonstrates rapid realignment under changing constraints and robustness in sample-limited settings, underscoring its practical potential for scalable, feasible molecular generation. Overall, S3-GFN provides a flexible, scalable pathway to integrate rich chemical priors with synthesizability constraints in sequence-based molecular generation.

Abstract

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.

Figures (11)

• Figure 1: Overview of Synthesizable SMILES via Soft-constrained GFN (S3-GFN). A pretrained SMILES prior provides chemical plausibility and is continuously referenced through RTB. On-policy updates apply RTB using positive samples only, while replay updates introduce a contrastive auxiliary loss that separates positive and negative samples, while preserving shared substructures.
• Figure 2: Deceptive 2D grid world with feasibility constraints. (a) Target distribution, where black cells denote infeasible states and colors indicate reward levels. (b) Learned sampling distribution using feasible-only training, and (c) learned sampling distribution with the auxiliary contrastive loss $\mathcal{L}_{\text{aux}}$.
• Figure 3: Comparison over different MDPs on sEH. While SFN guarantees 100% validity on the generation constraints (Positive Ratio), S3-GFN achieves higher success on AiZynthFinder and discovers candidates with consistently higher sEH scores.
• Figure 4: Example of synthetic pathway of Top-2 candidates under given reaction $\mathcal{R}$ and building block $\mathcal{M}$. With a high positive ratio, our generated molecules have valid synthetic pathways.
• Figure 5: ALDH1 docking results (Top-3). The molecules demonstrate shape complementarity with the target site, achieving strong predicted binding affinities (Vina scores $< -12.3$) and high drug-likeness (QED $> 0.87$) within the synthesizable space.