Projecting Molecules into Synthesizable Chemical Spaces

Shitong Luo, Wenhao Gao, Zuofan Wu, Jian Peng, Connor W. Coley, Jianzhu Ma

Introduction

In the field of drug discovery, generative AI models have demonstrated a remarkable ability to design molecules that theoretically optimize for specific biological targets. However, a persistent bottleneck remains: the "synthesizability gap." Many AI-generated candidates, while high-performing in simulations, are often "fantasy molecules"—structures that are chemically impossible to synthesize or prohibitively difficult to manufacture. Current approaches to mitigate this issue typically fall into two extremes: restrictive virtual libraries that limit chemical diversity, or de novo models that ignore reaction rules entirely.

ChemProjector is a novel framework designed to bridge this gap. By treating chemical synthesis as a sequence of stack-based instructions, it "projects" theoretically promising but unsynthesizable molecules into a space of makeable compounds, ensuring that every proposed drug candidate comes with a viable synthetic recipe.

The ChemProjector Framework

Synthesis as Postfix Notation

The core innovation of ChemProjector is its representation of molecular synthesis. Rather than using static graphs or SMILES strings, the framework utilizes a postfix notation (similar to Reverse Polish Notation) to describe a synthetic pathway. This stack-based instruction set allows the model to represent complex, branched synthesis pathways through two primary operations:

  1. Push: Add a specific building block from a library onto a memory stack.
  2. Apply: Execute a specific chemical reaction template on the top items of the stack.

This approach allows for multi-reactant reactions and complex branching that traditional linear models struggle to represent, effectively framing chemistry as a sequence of discrete, executable code.

Model Architecture

ChemProjector employs an encoder-decoder architecture to translate molecular structures into synthetic recipes:

  • Encoder: A Graph Transformer that processes the target molecule's structure to create a latent representation.
  • Decoder: An autoregressive Transformer that generates the postfix recipe. It predicts the next action (Push vs. Apply), identifies the necessary building blocks via fingerprint prediction and nearest-neighbor search, and selects the appropriate reaction templates.

The Projection Concept

The "projection" mechanism allows researchers to take an existing, potentially unsynthesizable molecule and find its closest synthesizable analog. By generating a recipe that results in a structurally similar product, ChemProjector acts as a "fixer" for other generative models (like Pocket2Mol), preserving the essential functional characteristics of a lead compound while ensuring its manufacturability.

Experimental Setup

Datasets and Resources

To ground the model in real-world chemistry, the authors utilized the Enamine US Stock catalog, comprising approximately 211,220 commercially available building blocks. The model's "vocabulary" of reactions consists of 91 expert-defined templates (e.g., amide coupling, Suzuki reaction) curated from the SynNet dataset.

Evaluation Benchmarks

The framework was evaluated across several standard benchmarks:

  • ChEMBL: 1,000 molecules used to test the reconstruction of known bioactive compounds.
  • LIT-PCBA: 15 protein targets used to evaluate structure-based drug design.
  • GuacaMol: A standard benchmark for goal-directed molecular generation.

Key Results and Insights

Superior Performance and Reconstruction

ChemProjector demonstrated significant improvements over existing baselines. It achieved a 97.5% success rate in finding valid synthetic pathways for a test set, compared to 84.1% for the baseline model, SynNet. Furthermore, the model successfully reconstructed 28.4% of test molecules exactly—nearly triple the performance of previous state-of-the-art methods (10.7%).

Practical Utility in Drug Design

In structure-based drug design tests, ChemProjector generated analogs that maintained similar docking poses to the original "fantasy" molecules while ensuring they could be synthesized from the Enamine catalog. In a specific "hit expansion" test for a JNK3 inhibitor, the model generated 500 analogs; notably, 21 of these analogs outperformed the original hit in predicted binding scores (Vina score) while remaining fully synthesizable.

Preserving Function through Structure

The results suggest that projecting a molecule into synthesizable space does not necessarily degrade its efficacy. In many cases, the predicted binding affinity was maintained or even improved, indicating that the model successfully identifies and preserves the pharmacophores essential for biological activity.

Limitations and Future Work

While ChemProjector represents a significant step forward, several challenges remain:

  • Template Dependency: The model is currently restricted to its 91 predefined reaction templates and cannot propose chemistry outside this set.
  • Catalog Constraints: The reliance on the Enamine catalog limits the diversity of starting materials; future iterations could benefit from more dynamic or broader building block libraries.
  • Experimental Conditions: The model provides the synthetic steps but does not yet predict specific experimental parameters such as temperature, solvents, or catalysts.
  • Stereochemistry: Explicit control over stereoisomers during synthesis generation remains an area for further refinement to ensure high-fidelity chemical production.

Conclusion

ChemProjector offers a robust solution to the synthesizability problem in AI-driven drug discovery. By framing synthesis as a sequence of stack-based instructions, it provides a practical pathway for turning computational "dreams" into laboratory realities, ensuring that the next generation of drug candidates is not just effective on screen, but also makeable at the bench.

Summary updated

BibTeX
@inproceedings{luo2024projecting,
  author = {Shitong Luo and Wenhao Gao and Zuofan Wu and Jian Peng and Connor W. Coley and Jianzhu Ma},
  title = {{Projecting Molecules into Synthesizable Chemical Spaces}},
  booktitle = {International Conference on Machine Learning},
  year = {2024},
  url = {https://openreview.net/forum?id=scFlbJQdm1},
}
Figure