Generative Flows on Synthetic Pathway for Drug Design

Seonghwan Seo, Minsu Kim, Tony Shen, Martin Ester, Jinkyoo Park, Sungsoo Ahn, Woo Youn Kim

In the field of drug discovery, a persistent challenge known as the "valley of death" exists: AI models often design molecules that appear ideal in simulations but are impossible or prohibitively expensive to synthesize in a laboratory. To bridge this gap, researchers have introduced RXNFLOW, a generative framework presented at ICLR 2025 that ensures every "invented" molecule is synthesizable by construction.

Bridging the Synthesizability Gap

Traditional AI-driven molecule generation often relies on post-hoc filtering, where molecules are generated first and checked for synthesizability later. This frequently results in the loss of high-performing candidates. Alternatively, reaction-based generation builds molecules step-by-step using chemical reactions. However, these models typically struggle with "combinatorial explosion." When dealing with millions of building blocks and numerous reaction templates, the mathematical action space becomes too vast for standard deep learning models, often limiting researchers to small, unrepresentative libraries.

RXNFLOW addresses these issues by simulating real-world chemical reaction pathways while scaling to industry-standard library sizes, such as the Enamine REAL Space, which contains 1.2 million building blocks.

The RXNFLOW Framework

The core of RXNFLOW is built upon Generative Flow Networks (GFlowNets). Unlike standard Reinforcement Learning, which may converge on a single optimal solution, GFlowNets are designed to sample diverse solutions with probabilities proportional to their rewards. This diversity is critical in drug discovery, where multiple chemical leads are often needed to find a successful candidate.

Overview of the RXNFLOW framework, illustrating action space construction, policy estimation via subsampling, and the iterative generation process.

Methodological Innovations

To handle the massive scale of 1.2 million building blocks, RXNFLOW introduces several key innovations:

  • Action Space Subsampling: Rather than calculating probabilities for every possible building block at every step, RXNFLOW samples a small subset (e.g., 1%) and employs importance sampling to estimate the total flow. This allows the model to explore a massive chemical space without an overwhelming memory footprint.
  • Action Embedding: Building blocks are represented by continuous vectors (embeddings) based on their chemical fingerprints (ECFP4) and physicochemical properties. This allows the model to generalize its knowledge to new, unseen building blocks rather than treating them as static IDs.
  • Non-hierarchical MDP: While most models first select a reaction type and then a building block, RXNFLOW selects them jointly. This non-hierarchical Markov Decision Process (MDP) makes the model significantly more robust. If a specific building block is removed from the library, the model can instantly adjust its strategy without requiring retraining.
Comparison between existing hierarchical MDPs and the proposed non-hierarchical MDP, highlighting adaptability to library modifications.

Experimental Results and Scale

RXNFLOW was evaluated using the Enamine REAL Space (1.2M blocks), the LIT-PCBA benchmark for protein-specific optimization, and CrossDocked2020 for zero-shot pocket-conditional generation.

Key Insights:

  • The Power of Scale: Increasing the library size from 100 to 1 million blocks significantly improved docking scores and molecular diversity. Notably, the runtime only doubled despite a 10,000x increase in the search space.
  • State-of-the-Art Performance: On the CrossDocked2020 benchmark, RXNFLOW achieved a Vina score of -8.85 kcal/mol and a 34.8% synthesizability rate (verified by AiZynthFinder), outperforming both atom-based and fragment-based models.
  • Efficiency and Diversity: The model successfully managed a library of 1.2M blocks—a massive leap from the ~6,000 blocks typical of previous reaction-based GFlowNets—while maintaining a high structural diversity (0.81 Tanimoto distance).

Future Directions

While RXNFLOW represents a significant step forward, future work will focus on making 3D spatial awareness more efficient within the GFlowNet framework to improve 3D interaction modeling. Additionally, researchers suggest that "prioritizing" certain building blocks during the subsampling phase based on chemical properties could further accelerate the discovery of high-reward molecules. Expanding the current 3-step reaction limit to longer, multi-step synthetic routes will also allow for the design of even more complex molecular structures.

Summary updated

BibTeX
@inproceedings{seo2025generative,
  author = {Seonghwan Seo and Minsu Kim and Tony Shen and Martin Ester and Jinkyoo Park and Sungsoo Ahn and Woo Youn Kim},
  title = {{Generative Flows on Synthetic Pathway for Drug Design}},
  booktitle = {International Conference on Learning Representations},
  year = {2025},
  url = {https://openreview.net/forum?id=pB1XSj2y4X},
}
Figure