Synthesizability Constrained Generative Molecular Design

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2524 papers758 conceptsfrozen 2026-10-04

In the quest for new medicines, researchers have turned to generative AI to navigate the astronomical "chemical space"—the $10^{60}$ possible drug-like molecules. However, a fundamental friction has emerged: while models like Variational Autoencoders (VAEs) and Diffusion models are excellent at "hallucinating" molecules that fit a protein pocket or satisfy a mathematical property, they often produce structures that are impossible to actually build in a laboratory.

This research direction focuses on bridging the gap between computational creativity and experimental feasibility, ensuring that AI-designed molecules are not just "valid" on a screen, but synthesizable at the bench.

The Chemical Hallucination Problem

Traditional generative models often treat molecules as strings (SMILES) or graphs, building them atom-by-atom or character-by-character. As Daniel C. Elton and colleagues note in their state-of-the-art review, this approach frequently leads to "fantasy" molecules. These structures might have high predicted binding scores but contain unstable rings, impossible bond angles, or complex macrocycles that no known chemical reaction can produce.

The problem is one of grounding. A model might learn the syntax of chemistry without understanding the physics of synthesis. As Qiaoyu Hu explains in Target-aware 3D molecular generation, even geometrically plausible designs often lack practical utility because they violate the fundamental rules of valence and structural stability that human chemists take for granted.

From Atoms to Motifs

One primary strategy to solve this is shifting the generative unit from individual atoms to motifs or fragments. By using pre-validated chemical building blocks, the model is physically constrained to stay within the bounds of known chemistry.

Yonghong Tian and the team behind DM-Assembler demonstrate that assembling molecules from domain-specific motifs—rather than atoms—inherently respects the complex rules of functional group compatibility. Similarly, Michael Sun introduced Foundation Molecular Grammar (FMG), which uses graph grammars to ensure that every step of the generative process follows a "design story" rooted in realistic chemical assembly.

Retrosynthesis in the Loop

If the goal is to make a molecule, the model should perhaps know how to take it apart. Researchers are increasingly integrating retrosynthesis engines directly into the generative loop. These engines act as a "sanity check," attempting to find a step-by-step recipe for a proposed molecule using commercially available starting materials.

Maksym Korablyov and Marwin H. S. Segler developed RetroGNN, which uses Graph Neural Networks to rapidly approximate expensive retrosynthetic planning. This allows the generative agent to be biased toward compounds that are not just potent, but manufacture-ready. More recently, the Saturn model demonstrated that including a full retrosynthesis engine in the optimization loop can achieve state-of-the-art results with significantly higher sample efficiency, ensuring that the "dream" molecules are actually reachable.

Generating the Recipe, Not the Product

A more radical shift involves generating the reaction pathway itself rather than the final structure. In this paradigm, the model doesn't "draw" a molecule; it "writes" a sequence of chemical reactions.

Mike Tyers and colleagues introduced RGFN (Reaction-GFlowNet), which operates entirely in "reaction space." By navigating a sequence of simulated chemical reactions, the model guarantees that every candidate it produces is synthesizable by definition. This approach effectively collapses the gap between design and production, as the output of the AI is the synthesis plan itself.

LLMs as Chemical Reasoning Agents

The rise of Large Language Models (LLMs) has introduced a new layer of "intuition" to this field. Unlike specialized graph models, LLMs can be prompted to reason about chemical stability and synthetic feasibility using natural language.

Systems like the OSDA Agent by Zhaolin Hu use LLMs as a "reasoning brain" to evaluate and reflect on generated structures. By integrating a "Self-reflector" that critiques designs based on chemical grounding, these agents avoid the common pitfalls of pure optimization. Similarly, DrugAssist by Geyan Ye allows researchers to iteratively refine molecules through dialogue, leveraging the LLM's broad domain knowledge to steer away from unsynthesizable "hallucinations" toward practical, drug-like leads.

Papers in this version

Showing 60 of 524 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 501 papers
  2. Version 1 · 2026-10-04 · 524 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_synthetic_accessibility_gap_1437044e,
  title = {synthetic accessibility gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/synthetic-accessibility-gap/rdr_1437044edb614506b3936ba2894f3d3b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_synthesizability_constrained_generative__1437044e_v1,
  title = {Synthesizability Constrained Generative Molecular Design},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/synthesizability-constrained-generative-molecular-design/rdr_1437044edb614506b3936ba2894f3d3b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}