Bridging the Synthesizability Gap in Molecular Generation

902 concepts500 papers2 proposals

Deep generative models have fundamentally changed how we explore chemical space. Whether using variational autoencoders, diffusion models, or instruction-tuned large language models, algorithms can now rapidly propose novel molecular structures optimized for specific target properties, such as binding affinity or narrow emission bands.

Yet, a glaring disconnect remains between computational optimization and laboratory reality. Models frequently generate "chemical hallucinations"—structures that look pristine on a computer screen and satisfy basic rules of valency, but are practically impossible, prohibitively expensive, or highly unstable to synthesize in a physical laboratory. This discrepancy is known as the synthesizability gap.

The Root of the Disconnect

Why do highly sophisticated models generate unbuildable molecules? The issue lies in how molecular generation is traditionally framed. Most algorithms treat molecules as discrete graphs (nodes as atoms, edges as bonds) or text strings (such as SMILES). The model's task is to modify these graphs atom-by-atom or fragment-by-fragment to maximize a property score.

As noted by Kisoo Kwon, ensuring that a generated molecule satisfies basic chemical valency does not mean the structure can be physically assembled. Traditional de novo design methods optimize molecular graphs for target properties without considering the underlying chemical reactions required to produce them. This creates what John Bradshaw terms the molecular recipe problem: the model delivers a final product design but provides no instructions on how to make it.

Furthermore, computational metrics often abstract away the messy physical realities of organic chemistry. As Abraham Heifets et al. point out, many models focus solely on qualitative connectivity, ignoring critical physical factors such as reaction yields, temperature constraints, and stereochemistry. A molecule that appears geometrically perfect inside a simulated protein pocket may require a synthetic pathway that violates thermodynamic stability or relies on non-existent precursors, rendering it useless for practical applications.

Traditional Approach:
[Target Properties] ──(Optimization)──> [Molecular Graph] ──> (Synthesis Bottleneck: No Recipe)

Reaction-Driven Approach:
[Available Reactants] ──(Simulated Reactions)──> [Synthesizable Product] ──> (Property Evaluation)

Shifting from Graph Space to Reaction Space

To close this gap, researchers have proposed flipping the generative paradigm on its head: instead of generating a structure and hoping a retrosynthetic route exists, models can generate molecules by explicitly simulating the assembly process.

In Generating Molecules via Chemical Reactions, John Bradshaw et al. introduced Molecule Chef. Instead of drawing a molecule atom-by-atom, the model operates over a continuous latent space of reactant combinations. It selects a "bag of ingredients" (commercially available starting materials) and predicts how they will react using a reaction predictor. Because the generation process itself mimics forward synthesis, every proposed molecule inherently arrives with a realistic, valid recipe.

Differentiable Rewards and Constrained Optimization

For models that do not generate molecules via forward reactions, the synthesizability gap must be managed by embedding synthetic constraints directly into the model's optimization loop.

A common strategy is to incorporate synthetic accessibility scores into multi-objective reinforcement learning reward functions, as demonstrated by Aditya Malusare. However, optimizing for chemical synthesis is notoriously difficult because it is often an "all-or-nothing" discrete problem—a single misplaced atom can render an entire synthetic route unviable.

To address this, Philippe Schwaller et al. developed TANGO (Tanimoto Group Overlap), a smooth, differentiable reward function. TANGO evaluates how well a generated molecule's structural fragments overlap with a pre-defined set of available chemical building blocks. By converting the rigid, discrete challenge of chemical synthesis into a continuous, learnable signal, the framework allows generative models to smoothly navigate toward high-performing spaces that chemists can actually construct.

Grounding in Literature and Local Inventories

Another dimension of the synthesizability gap is data alignment. Computational metrics for ease-of-manufacture often fail to capture real-world laboratory complexities. To anchor models in historical precedent, Edward Kim demonstrated that literature-trained neural networks can serve as effective screening tools. By training a model on millions of past reactions, the system can evaluate whether a newly proposed material aligns with established laboratory practices and known precursor behaviors.

Finally, the definition of "synthesizable" is often highly contextual. A molecule might be theoretically possible to make, but practically unfeasible if the required precursors take weeks to ship. Alan Kai Hassen et al. argue for a pragmatic shift toward in-house synthesizability. In Generate what you can make: achieving in-house synthesizability with readily available resources in de novo drug design, the authors emphasize constraining generative models to the exact inventory of bottles available on a specific laboratory's shelves. This constraint eliminates supply-chain bottlenecks and ensures that the automated discovery cycle can move immediately from digital design to physical validation.

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 2, frozen text):

@misc{lacuna_synthetic_accessibility_gap_1437044e_v2,
  title = {synthetic accessibility gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/synthetic-accessibility-gap/rdr_1437044edb614506b3936ba2894f3d3b/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake