Research map / Probabilistic and scientific modeling / Molecules and proteins
Scalable and Cost-Efficient de Novo Template-Based Molecular Generation
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 SCENT, a template-based GFlowNet that uses recursive cost guidance on the backward policy, an exploitation penalty, and a dynamic building-block library to cut synthesis cost and scale to large fragment libraries.
Problem
Template-based molecular generation with GFlowNets suffers from high synthesis cost, poor scaling to large building block libraries, and weak performance with small fragment sets.
Approach
Recursive Cost Guidance trains a learned surrogate to approximate the recursive synthesis cost (and decomposability) of backward transitions, steering the backward policy toward cheap, viable precursors; an exploitation penalty counters the resulting exploitation, and a dynamic library adds high-reward intermediates as reusable building blocks to enable full synthesis trees.
Key results
- On sEH MEDIUM (threshold 8.0), SCENT (C+D+P) reports 32761 modes and 86081 scaffolds vs RGFN's 4755 modes and 10638 scaffolds, and SCENT (C+P) reaches 37714 modes.
- In SMALL sEH, SCENT (C+D+P) finds 715 modes vs 538 for RGFN and reduces cost of high-reward paths to 20.7 vs 37.7 for RGFN.
- SCENT (C+D+P) achieves the highest reward in SMALL (7.66) and LARGE (7.78) sEH settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Synthesis Cost Guidance (ablation) | SCENT without cost guidance · sEH SMALL | synthesis cost: 37.1 \std{1.0} -> 23.7 \std{4.0} | Helped |
| Add Synthesis Cost Guidance (ablation) | SCENT without cost guidance · sEH MEDIUM | scaffolds > 8.0: 11478 \std{823} -> 52340 \std{4303} | Helped |
| Add Synthesis Cost Guidance (ablation) | SCENT without cost guidance · sEH SMALL, reward threshold 7.2 | modes: 13862 \std{422} -> 11978 \std{197} | Hurt |
| Add Exploitation Penalty to cost-guided SCENT (ablation) | SCENT (C) · sEH SMALL, reward threshold 7.2 | modes: 11978 \std{197} -> 14631 \std{206} | Helped |
| Add Dynamic Library to cost-guided SCENT (ablation) | SCENT (C) · sEH SMALL | modes > 8.0: 478 \std{11} -> 596 \std{19} | Helped |
| Add Decomposability Guidance to cost-guided SCENT (ablation) | SCENT (C) · sEH MEDIUM | scaffolds > 8.0: 52340 \std{4303} -> 52607 \std{1352} | No clear effect |
| Set decomposability guidance temperature alpha_D to 2 instead of 5 (ablation) | SCENT (C+D+P) · sEH MEDIUM | modes > 8.0: 16373 ± 1827 -> 20023 ± 1484 | Helped |
| Set decomposability guidance temperature alpha_D to 10 instead of 5 (ablation) | SCENT (C+D+P) · sEH MEDIUM | modes > 8.0: 16373 ± 1827 -> 20444 ± 2055 | Helped |
| Set synthesis cost guidance temperature alpha_S to 10 instead of 5 (ablation) | SCENT (C+D+P) · sEH MEDIUM | synthesis cost: 1163 ± 81 -> 898 ± 42 | Helped |
| Set dynamic library size L to 200 instead of 400 (ablation) | SCENT (C+D+P) · sEH MEDIUM | modes > 8.0: 16373 ± 1827 -> 19700 ± 1919 | Helped |
| Set exploitation penalty initial gamma_0 to 0.5 instead of 1.0 (ablation) | SCENT (C+D+P) · sEH MEDIUM | modes > 8.0: 16373 ± 1827 -> 20690 ± 1512 | Helped |
Surprising findings
- Synthesis Cost Guidance improved mode discovery as well as cost, which the authors initially found counterintuitive.
- RxnFlow and SynFlowNet perform far below RGFN and SCENT in most settings under the unified benchmark, and SynFlowNet could not match its reported 7.2-threshold results.
Limitations
- Synthesis cost model is a simplified first-order heuristic omitting solvents, purification, and conditions; yields and prices are approximate.
- Performance varies across biological targets, and the number of reaction templates scaling was not assessed.
- No toxicity or ADMET consideration and no wet-lab validation.
Experimental setup
Datasets: Enamine REAL building blocks (64k MEDIUM, 128k LARGE subsets) curated SMALL fragment set SynFlowNet reaction templates
Benchmarks: sEH proxy GSK3beta proxy JNK3 proxy
Metrics: modes (ECFP6 Tanimoto < 0.5) above threshold unique Bemis-Murcko scaffolds above threshold average reward synthesis cost of high-reward paths AiZynthFinder success rate SA score
Baselines: RGFN SynFlowNet RxnFlow FGFN FGFN (SA)
Builds on: RGFN trajectory balance Guided Trajectory Balance SynFlowNet ActionPiece
Closest papers by meaning
- SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints (ICLR 2025 · similarity 0.90)
- RGFN: Synthesizable Molecular Generation Using GFlowNets (NeurIPS 2024 · similarity 0.90)
- Generative Flows on Synthetic Pathway for Drug Design (ICLR 2025 · similarity 0.89)
- Pretraining Generative Flow Networks with Inexpensive Rewards for Molecular Graph Generation (ICML 2025 · similarity 0.86)
- Generative Flow Networks for Lead Optimization in Drug Design (Student Abstract) (AAAI 2025 · similarity 0.86)
- Compositional Flows for 3D Molecule and Synthesis Pathway Co-design (ICML 2025 · similarity 0.86)
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.