Programmable Exploration of Synthesizable Chemical Space

Shitong Luo, Connor W. ColeyOpenReview 2026 · Submission to ICLR 2026 · Rejected · OpenReview

TL;DR

PrexSyn is a new generative framework that ensures every molecule it designs can actually be made in a lab. By treating molecular design as a "programmable" query task, it allows researchers to combine multiple property constraints using logical operators (AND, OR, NOT) and achieves record-breaking efficiency in finding synthesizable analogs for known drugs.

What Problem Does This Paper Address?

In the world of AI-driven drug discovery, we have a "paper molecule" problem. Most generative models represent molecules as strings or graphs without considering how to actually build them. Consequently, an AI might propose a "miracle drug" that is chemically impossible to synthesize.

Recent "synthesis-aware" models try to fix this by generating the "recipe" (synthetic pathway) instead of just the final structure. However, these models face three major hurdles:

  1. Massive Search Space: Building block libraries (like Enamine REAL) contain hundreds of thousands of chemicals, making it hard for models to learn which ones to pick.
  2. Sparsity: Most combinations of chemicals and reactions don't result in a valid molecule.
  3. Structural Blindness: Many models focus so much on the "recipe" that they lose track of the final molecule's actual shape and properties.

Methods

The authors introduce PrexSyn (Programmable Exploration of Synthesizable Chemical Space), a decoder-only Transformer model.

1. Postfix Notation of Synthesis

Instead of drawing a graph, the model generates a sequence of tokens called postfix notation. Think of this as a stack-based recipe: "Take Building Block A, take Building Block B, apply Reaction 1." This format guarantees that every output corresponds to a valid synthetic pathway.

2. Two-Level Token Prediction

To handle the massive library of building blocks (over 223,000), PrexSyn uses a two-step approach. It first predicts the class of the next step (e.g., "pick a building block") and then uses a specialized classifier to pick the specific chemical. This is much more efficient than searching the entire library at every single step.

3. Programmable Logic (Product of Experts)

This is the "programmable" part. If you want a molecule that is soluble AND binds to a specific protein BUT NOT toxic, PrexSyn treats each requirement as an "expert" distribution. It combines these distributions mathematically during sampling, allowing users to "program" complex multi-objective goals without retraining the model.

4. Query Space Optimization

Instead of tweaking the molecule's atoms directly, the authors optimize the query sent to the model. By iteratively refining the property prompts based on feedback from a "black-box" simulator (an oracle), the model explores the chemical space much faster than traditional methods.

Datasets

  • Training Data: Randomly constructed synthetic pathways using 223,244 building blocks from the Enamine US in-stock set and 115 reaction templates.
  • Enamine Testset: 1,000 molecules used to test how well the model can "reconstruct" known synthesizable molecules.
  • ChEMBL Testset: 1,000 molecules from a general medicinal chemistry database to test the model's ability to find synthesizable analogs for arbitrary structures.
  • GuacaMol Benchmarks: A standard suite for evaluating multi-objective molecular optimization.
  • Biological Targets: Docking simulations for sEH (soluble epoxide hydrolase) and Mpro2 (a SARS-CoV-2 protease).

Key Insights and Ideas

  • Structure Determines Function: The model is trained on easy-to-calculate structural properties (like fingerprints and molecular weight). Because structure dictates function, the authors show that the model can generalize to complex biological functions (like protein binding) at inference time without having seen that specific data during training.
  • Bridging the Gap: By conditioning the synthesis recipe on structural prompts, the model creates a direct link between "how we make it" and "what it does," solving the structural blindness of previous synthesis-based models.
  • Efficiency through Abstraction: Optimizing in "query space" (the prompts) rather than "action space" (the specific building blocks) allows the model to bypass the combinatorial explosion that bogs down other methods.

Key Results

  • Record Coverage: PrexSyn achieved a 92.9% reconstruction rate on the Enamine testset, shattering the previous state-of-the-art of 69.1%.
  • Speed: It is remarkably fast, taking only 0.18 seconds to find a synthesizable analog for a target molecule.
  • Optimization Power: In the GuacaMol benchmarks, PrexSyn outperformed both synthesis-agnostic and synthesis-based baselines in 6 out of 8 tasks.
  • Biological Accuracy: In docking tasks for the SARS-CoV-2 protease, the model generated molecules from scratch that achieved better binding scores than known inhibitors, using 30x fewer "oracle calls" (simulations) than competing models like SynFlowNet.

Limitations and Future Work

  • Independence Assumption: The "Product of Experts" logic assumes that different molecular properties are independent. In reality, properties like molecular weight and rotatable bonds are often correlated. While the model works well despite this, a more nuanced composition rule could improve results.
  • Reaction Scope: The model is currently limited to the 115 reaction templates provided in the training set. Expanding this to a broader range of chemical transformations would further increase the diversity of the generated molecules.
  • Secondary Building Blocks: The authors noted that using branched pathways (intermediates as building blocks) improved capacity, suggesting that further refining the "data pipeline" for synthetic pathways is a promising direction.

Summary updated

Abstract

The constrained nature of synthesizable chemical space poses a significant challenge for sampling molecules that are both synthetically accessible and possess desired properties. In this work, we present a programmable model for the discovery of molecules within synthesizable space. The model can generate synthesizable molecules subject to complex logical queries of molecular properties. By leveraging this querying capability, molecular optimization with respect to black-box oracle functions can be performed through iterative refinement of the queries, which achieves high optimization efficiency while preserving synthesizability. We demonstrate the model's high coverage of the synthesizable chemical space, achieving a record-high 92\% reconstruction rate on a chemical space projection test set curated from the Enamine REAL database. We then evaluate its capability for complex query-based molecular design through a series of multi-objective molecular discovery benchmarks. Finally, we show that the query-based molecular optimization technique achieves substantially higher sampling efficiency than both synthesis-based and synthesis-agnostic methods.
BibTeX
@misc{luo2026programmable,
  author = {Shitong Luo and Connor W. Coley},
  title = {{Programmable Exploration of Synthesizable Chemical Space}},
  year = {2026},
  url = {https://openreview.net/forum?id=SlVkjuq5Ig},
  howpublished = {OpenReview},
  note = {Submitted to ICLR 2026},
}
Figure