Research map / Probabilistic and scientific modeling / Molecules and proteins
Procedural Synthesis of Synthesizable Molecules
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.
Reframes synthesizable analog generation and molecule design as bi-level program synthesis, using MCMC over syntactic skeletons and amortized policies over fixed-horizon MDPs for the semantics, plus a genetic algorithm over fingerprints and skeletons.
Problem
Generating synthesizable analogs of unsynthesizable molecules and designing synthesizable molecules optimizing black-box oracles, where the combinatorial space of synthesis pathways makes search expensive.
Approach
Decouples the syntactic skeleton (binary tree of reactions) from semantics (reactions and building blocks); runs Metropolis-Hastings over skeletons with an amortized top-down GNN policy as the inner decoder, and a GA over fingerprint x skeleton space with syntactic-mutation sibling generation.
Key results
- On 13 TDC oracles, the method outperforms all synthesis-based baselines on average score (0.670), AUC (0.608), and SA (2.739), ranking 1st on AUC and competitive with graph/string methods.
- On the test set, recovery rate rises to 52.3% versus SynNet's 46.3% for analog generation.
- AutoDock Vina against M-pro reaches top-1 score 9.9 versus SynNet's 8.3 at 5000 oracle calls.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed syntax-guided method instead of SynNet | synthesizable analog generation · Test Set | RR: 46.3% -> 52.3% | Helped |
| use the proposed syntax-guided method instead of SynNet | synthesizable analog generation · Test Set | Diversity Top-5: 0.584 -> 0.653 | Helped |
| use MCMC syntax refinement instead of SynNet decoding | synthesizable analog generation · ChEMBL | RR: 4.9% -> 9.2% | Helped |
| use the proposed method instead of the strongest reported baseline | synthesizable molecule design · 13 TDC oracles average | AUC: 0.607 -> 0.608 | Helped |
| use the proposed method instead of SynNet | molecule design · DRD3 docking, 5000 calls | Top-10 score: 10.30 -> 13.01 | Helped |
| use the proposed method instead of SynNet | molecule design · Mpro docking, 5000 calls | Top-10 score: 8.09 -> 9.54 | Helped |
| use syntax-edit mutations for sibling pool generation rather than τ-predicted skeletons (ablation) | Ours genetic algorithm · JNK3 optimization | Top-10 score: 0.83 -> 0.86 | Helped |
| use syntax-edit mutations rather than random fingerprint bit flips (ablation) | Ours genetic algorithm · JNK3 optimization | Diversity: 0.49 -> 0.61 | Helped |
| add Bayesian optimization acquisition over a beam-search sibling pool (ablation) | SynNet · GSK3β optimization | Top-10 score: 0.907 -> 0.684 | Hurt |
| add Bayesian optimization acquisition over a beam-search sibling pool (ablation) | SynNet · JNK3 optimization | Top-10 score: 0.758 -> 0.241 | Hurt |
| decode the syntax skeleton top-down rather than bottom-up (ablation) | Ours analog generation · Test Set | RR: 40.8 % -> 52.3% | Helped |
| remove about 25% of template classes from policy-network training (ablation) | Ours analog generation · Test Set | RR: 56% -> 52% | No clear effect |
Surprising findings
- Adding the same BO acquisition step to SynNet degraded its performance (e.g., GSK3beta top-1 0.94 to 0.85), suggesting the gains depend on the syntax-driven sibling generation.
Limitations
- Restricted to a finite set of syntactic templates (up to k internal nodes) and 91 reaction templates, limiting coverage of chemistry.
- Recognition MLP alone is overly naive; syntax prediction can overfit to motifs present in building blocks rather than reaction steps.
Experimental setup
Datasets: Enamine Building Blocks (147,505 compounds) 91 reaction templates from prior work ChEMBL
Benchmarks: TDC oracles (13) AutoDock Vina DRD3 AutoDock Vina M-pro ChEMBL analog generation JNK3 GSK3beta DRD2
Metrics: recovery rate average similarity SA score internal diversity top-k average score top-k AUC
Baselines: SynNet DoG-Gen DoG-AE MolPAL REINVENT REINVENT-SELFIES Graph-GA GPBO
Builds on: SynNet Graph-GA syntax-guided synthesis Metropolis-Hastings
Closest papers by meaning
- Projecting Molecules into Synthesizable Chemical Spaces (ICML 2024 · similarity 0.90)
- LLM-Augmented Chemical Synthesis and Design Decision Programs (ICML 2025 · similarity 0.89)
- SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints (ICLR 2025 · similarity 0.89)
- Efficient Evolutionary Search Over Chemical Space with Large Language Models (ICLR 2025 · similarity 0.88)
- Representing Molecules as Random Walks Over Interpretable Grammars (ICML 2024 · similarity 0.87)
- Beyond Chemical QA: Evaluating LLM's Chemical Reasoning with Modular Chemical Operations (NeurIPS 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.