Molecular Generation via Synthesis Pathways

36 concepts26 papers1 proposal

In the search for new drugs, artificial intelligence has traditionally acted as a digital architect, "drawing" molecular graphs atom-by-atom to optimize for specific biological targets. However, as Gao and Coley observe, these models often suffer from a "synthesizability gap": they propose "fantasy" molecules that look perfect on a screen but are chemically impossible to build in a laboratory.

A growing body of research addresses this by shifting the generative target. Instead of designing the final structure, these models design the synthesis pathway—the step-by-step recipe of chemical reactions and building blocks required to create a molecule. By making synthesizability an intrinsic constraint rather than a post-hoc filter, this approach ensures that every proposed candidate is "makeable" by construction.

The Shift to Reaction Space

Traditional generative models operate in "atom space," where the action is to add a node or an edge to a graph. Synthesis-constrained models instead operate in "reaction space." The generative process becomes a sequence of chemical transformations applied to a set of starting materials.

As Bradshaw et al. and the authors demonstrate with Molecule Chef, the model selects a "bag of ingredients"—commercially available reactants—and predicts how they will react to form a product. This shift ensures near-perfect chemical validity because the output is always the result of a valid reaction between known ingredients. By navigating a continuous latent space of reactant combinations, these models explore the chemical landscape through the lens of laboratory feasibility.

Synthesis as a Directed Acyclic Graph

To handle complex molecules, the generative process is often framed as the construction of a Synthesis Directed Acyclic Graph (DAG). In this framework, the leaves of the tree are purchasable building blocks, and the internal nodes represent chemical reactions.

Bradshaw et al. propose models that output these DAGs directly. This provides an inductive bias that restricts the AI to the same chemical space navigated by human chemists. Similarly, Noh et al. introduce C-RSVAE, which jointly learns molecular structures and their reaction routes. By encoding the entire sequence of reactions into a latent space, the model guarantees that every generated molecule is associated with a valid, multi-step chemical recipe.

The space of all possible molecules is astronomically large, but the space of synthesizable molecules is a much smaller, structured manifold. Researchers use various search strategies to navigate this manifold:

  • Reinforcement Learning: Gottipati et al. developed PGFS, which uses policy gradients to simulate forward synthesis. By treating molecular assembly like a strategic game, the model learns to pick the best sequence of reactions to reach a high-reward target.
  • GFlowNets: Koziarski and Cretu and Cretu (with SynFlowNet) use Generative Flow Networks to sample diverse synthesis paths. These models are designed to discover a variety of ways to reach high-quality regions of the chemical space, ensuring the "recipe" is not just valid, but also diverse.
  • Monte Carlo Tree Search: Zhao et al. apply MCTS to design ionizable lipids, treating the assembly of building blocks as a search tree where each branch is a validated chemical reaction.

Bridging the Execution Gap

The ultimate goal of this research direction is to close the "execution gap"—the friction between digital design and physical realization identified by Subramanian et al..

By incorporating practical constraints like the cost of building blocks (Gaiński et al.) or focusing on highly reliable "click chemistry" (Wang et al.), these models transition from abstract structure generators to actionable laboratory tools. As Luo and Coley show with PrexSyn, generating the postfix notation of a synthesis pathway rather than just a molecular graph allows for 90%+ reconstruction of massive chemical libraries, proving that AI can finally "think" in the same language as the chemists who must eventually build its creations.

Key Foundational Work

Barking up the right tree: an approach to search over molecule synthesis DAGs (2020) — contributed 4 concept elements to this direction.
Generating Molecules via Chemical Reactions (2019) — contributed 2 concept elements to this direction.
Learning to Navigate The Synthetically Accessible Chemical Space Using Reinforcement Learning (2020) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_synthesizability_constrained_design_8a923ba2,
  title = {synthesizability-constrained design},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/synthesizability-constrained-design/rdr_8a923ba2b62249f78c728d240442b3af},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_synthesizability_constrained_design_8a923ba2_v1,
  title = {synthesizability-constrained design},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/synthesizability-constrained-design/rdr_8a923ba2b62249f78c728d240442b3af/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake