SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints

TL;DR

SynFlowNet tackles the synthetic accessibility gap in de novo molecular design by integrating a reaction-based action space into a GFlowNet, forcing generation to follow synthesizable pathways. The framework combines a forward policy with a parameterized backward policy and employs masking and fingerprint-based scaling to handle large reaction and building-block spaces, achieving diverse, synthesizable outputs better than RL baselines. Key contributions include a novel backward-policy training regime to maintain MDP-consistent backward trajectories, scaling strategies for large BB libraries, and demonstration that target-specific fragment information can further guide synthesis. The approach promises practical impact for drug discovery by bridging in silico design with real-world synthesis and retrosynthesis planning, while remaining adaptable to multiple targets and space sizes.

Abstract

Generative models see increasing use in computer-aided drug design. However, while performing well at capturing distributions of molecular motifs, they often produce synthetically inaccessible molecules. To address this, we introduce SynFlowNet, a GFlowNet model whose action space uses chemical reactions and purchasable reactants to sequentially build new molecules. By incorporating forward synthesis as an explicit constraint of the generative mechanism, we aim at bridging the gap between in silico molecular generation and real world synthesis capabilities. We evaluate our approach using synthetic accessibility scores and an independent retrosynthesis tool to assess the synthesizability of our compounds, and motivate the choice of GFlowNets through considerable improvement in sample diversity compared to baselines. Additionally, we identify challenges with reaction encodings that can complicate traversal of the MDP in the backward direction. To address this, we introduce various strategies for learning the GFlowNet backward policy and thus demonstrate how additional constraints can be integrated into the GFlowNet MDP framework. This approach enables our model to successfully identify synthesis pathways for previously unseen molecules.

Figures (17)

• Figure 1: SynFlowNet allows for synthesis-aware molecule generation. (A) The state space is induced by combining purchasable building blocks (BBs) and chemical reactions. Every final molecule (rectangle box) is associated with a reward. Training trajectories are constructed by sampling the model forward. Our policy $P_F(a|s)$ is parameterized as a graph transformer which at each timestep processes the current molecular state $s_t$ and outputs a shared embedding which is passed to separate MLP heads to predict the action logits for different action types (5 forward and 3 backward action types). An action $a_t$ is then sampled from this hierarchical distribution to transition to the next state via a reaction. (B) To allow handling large sets of BBs (up to 200k), we represent them using Morgan fingerprints and compute the probability of sampling a particular BB from the normalised dot product between this representation and the MLP output for the current state. The state representation is concatenated with the one-hot encoding of the selected reaction. (C) Finally, when traversing the MDP backwards, to reduce the probability of exiting the MDP defined by our set of reactions and BBs, we train the backward policy to avoid paths that do not terminate in $s_0$.
• Figure 2: Estimated size of state spaces. The full building block set contains 221,181 molecules. $L$ is the maximum trajectory length in the GFlowNet.
• Figure 3: Comparison across MDPs. We evaluate four GFlowNet models: SynFlowNet is trained with an action space of chemical reactions and maximum trajectory lengths (L) of 3 and 4, FragGFN and FragGFN SA are trained with an action space of fragments, however the latter also optimises for synthetic accessibility (SA), besides the sEH binding proxy reward. SynFlowNet molecules are achieving higher binding scores and better synthesizability.
• Figure 4: Comparison between GFlowNet and RL. SynFlowNet (GFlowNet with synthesis actions) discovers more modes compared to entropy-regularised RL trained on the same state & action space.
• Figure 5: SynFlowNet is competitive against other popular models from the literature. SynFlowNet achieves a good balance between reward optimisation, diversity, synthesizability and novelty (assessed by maximum similarity to ChEMBL molecules). REINVENT stays close to its pretraining distribution, harming the novelty of the proposed molecules. SynFlowNet is closest to ideal.