Research map / Probabilistic and scientific modeling / Molecules and proteins
Projecting Molecules into Synthesizable Chemical Spaces
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ChemProjector, a transformer that maps molecular graphs to postfix-notation synthesis pathways over Enamine building blocks and reaction templates, projecting unsynthesizable molecules onto synthesizable analogs.
Problem
Generative models for drug discovery often propose synthetically infeasible molecules, and existing synthesizable design methods have poor coverage or weak optimization of the synthesizable chemical space.
Approach
Represent synthetic pathways as postfix notation (stack-based sequence of building blocks and reactions) and train a graph-transformer encoder with autoregressive transformer decoder to translate molecular graphs into such pathways, so any output is synthesizable by construction and can serve as an analog of an input molecule.
Key results
- On the test set, ChemProjector reaches 97.5% success and 28.4% reconstruction versus SynNet's 84.1% and 10.7%.
- On ChEMBL, reconstruction is 13.3% versus SynNet's 5.4%, with Morgan similarity 0.5978 versus 0.4270.
- In hit expansion for JNK3, 21 of 500 analogs outperform the hit's JNK3 score at average Tanimoto similarity 0.67, while PubChem search found only 7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ChemProjector instead of SynNet | Bottom-up synthesis planning · Test Set | success rate: 84.1% -> 97.5% | Helped |
| Use ChemProjector instead of SynNet | Bottom-up synthesis planning · Test Set | reconstruction rate: 10.7% -> 28.4% | Helped |
| Use ChemProjector instead of SynNet | Bottom-up synthesis planning · ChEMBL | reconstruction rate: 5.4% -> 13.3% | Helped |
| Use ChemProjector instead of SynNet | Bottom-up synthesis planning · ChEMBL | Morgan similarity: 0.4270 -> 0.5978 | Helped |
| Use ChemProjector instead of SynNet | Bottom-up synthesis planning · Stricter Test Set | scaffold similarity: 0.6012 -> 0.7239 | Helped |
| Project unsynthesizable goal-directed molecules into synthesizable analogs | ChemProjector · Goal-directed generation objectives | average objective score: 0.8110 -> 0.4824; 0.9768 -> 0.7296; 0.9316 -> 0.4321; 0.9030 -> 0.5424; 0.7169 -> 0.3552; 0.8889 -> 0.5652; 0.9558 -> 0. | Hurt |
| Project Pocket2Mol molecules into synthesizable analogs | ChemProjector · LIT-PCBA targets | Vina score: -8.70 -> -9.09; -9.10 -> -9.00; -9.46 -> -9.36; -9.77 -> -9.77; -6.54 -> -6.60; -6.96 -> -7.69; -9.50 -> -9.56; -8.07 -> | Mixed |
| Generate 500 synthesizable analogs around the hit | ChemProjector · JNK3 hit expansion | JNK3 score: 21 analogs outperformed the original hit; 0.68 | Helped |
| Use ChemProjector rather than PubChem nearest-neighbor retrieval | JNK3 hit expansion · JNK3 | number of molecules surpassing hit score: 7 out of the retrieved molecules | Helped |
Surprising findings
- Reconstruction of ChEMBL molecules is low (13.3%) even though most generated pathways are valid, showing that exact synthesis-pathway recovery remains hard for unseen chemistry.
Limitations
- Analogs for goal-directed generated molecules are structurally dissimilar (average Morgan similarity 0.43) and often have lower objective scores.
- Exact reconstruction rates are low on unseen molecules (e.g., 13.3% on ChEMBL).
- Limited to a fixed set of 211,220 building blocks and 91 reaction templates.
Experimental setup
Datasets: Enamine US Stock building blocks SynNet reaction templates ChEMBL Release 33 LIT-PCBA GuacaMol multi-property objectives PubChem
Benchmarks: held-out building-block test set ChEMBL 1000 molecules LIT-PCBA Pocket2Mol molecules GuacaMol goal-directed molecules JNK3 hit expansion
Metrics: success rate reconstruction rate Tanimoto similarity (Morgan, scaffold, Gobbi) Vina docking score objective score
Baselines: SynNet PubChem similarity search
Builds on: SynNet graph transformer Pocket2Mol
Closest papers by meaning
- Procedural Synthesis of Synthesizable Molecules (ICLR 2025 · similarity 0.90)
- SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints (ICLR 2025 · similarity 0.89)
- Representing Molecules as Random Walks Over Interpretable Grammars (ICML 2024 · similarity 0.89)
- Navigating Chemical Space with Latent Flows (NeurIPS 2024 · similarity 0.89)
- Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation (ICLR 2024 · similarity 0.88)
- Generative Flows on Synthetic Pathway for Drug Design (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.