Research map / Probabilistic and scientific modeling / Bayesian optimization
Sample-efficient Multi-objective Molecular Optimization with GFlowNets
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a preference-conditioned hypernetwork rather than concatenation or FiLM conditioning (ablation) | GFlowNet on the synthetic molecular optimization task · Synthetic GSK3β + JNK3 | HV: Concat-GFN: 0.534 ± 0.069; FiLM-GFN: 0.431 ± 0.045; HN-GFN: 0.550 ± 0.074 | Helped |
| Use one preference-conditioned HN-GFN instead of separate preference-specific GFlowNets | GFlowNet on the synthetic molecular optimization task · Synthetic GSK3β + JNK3 | HV: PS-GFN: 0.545 ± 0.055; HN-GFN: 0.550 ± 0.074 | No clear effect |
| Use HN-GFN with hindsight-like off-policy sharing (ablation) | HN-GFN within MOBO · MOBO first-round training | average Top-20 reward | Helped |
| Increase the hindsight mixture proportion γ (ablation) | HN-GFN within MOBO · MOBO first-round training | average Top-20 reward | Helped |
| Skew the preference sampling distribution toward harder properties (ablation) | HN-GFN · GSK3β + JNK3 + QED + SA MOBO | HV: 0.312 ± 0.039 -> 0.416 ± 0.023 | Helped |
| Replace weighted-sum scalarization with Tchebycheff scalarization (ablation) | HN-GFN · GSK3β + JNK3 + QED + SA MOBO | HV: 0.416 ± 0.023 -> 0.304 ± 0.075 | Hurt |
| Replace evidential regression with Deep Ensembles (ablation) | HN-GFN MOBO surrogate · GSK3β + JNK3 MOBO | HV: 0.669 ± 0.061 -> 0.583 ± 0.103 | Hurt |
| Replace evidential regression with a Tanimoto-kernel GP (ablation) | HN-GFN MOBO surrogate · GSK3β + JNK3 MOBO | Div: 0.793 ± 0.007 -> 0.739 ± 0.008 | Hurt |
| Replace evidential regression with Deep Ensembles (ablation) | HN-GFN MOBO surrogate · GSK3β + JNK3 + QED + SA MOBO | HV: 0.416 ± 0.023 -> 0.355 ± 0.048 | Hurt |
| Replace evidential regression with a Tanimoto-kernel GP (ablation) | HN-GFN MOBO surrogate · GSK3β + JNK3 + QED + SA MOBO | HV: 0.416 ± 0.023 -> 0.421 ± 0.037 | No clear effect |
| Train HN-GFN using weighted-sum rather than Tchebycheff scalarization (ablation) | HN-GFN · GSK3β + JNK3 + QED + SA MOBO | Div: 0.732 ± 0.014 -> 0.738 ± 0.009 | No clear effect |
Closest papers by meaning
- Generative Flow Networks for Lead Optimization in Drug Design (Student Abstract) (AAAI 2025 · similarity 0.88)
- Pretraining Generative Flow Networks with Inexpensive Rewards for Molecular Graph Generation (ICML 2025 · similarity 0.88)
- Genetic-guided GFlowNets for Sample Efficient Molecular Optimization (NeurIPS 2024 · similarity 0.87)
- Order-Preserving GFlowNets (ICLR 2024 · similarity 0.87)
- SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints (ICLR 2025 · similarity 0.87)
- RGFN: Synthesizable Molecular Generation Using GFlowNets (NeurIPS 2024 · 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.