Research map / Reinforcement learning / Visual and representation RL
Genetic-guided GFlowNets for Sample Efficient Molecular Optimization
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 Genetic GFN, which combines a graph-based genetic algorithm (Graph GA) with GFlowNet trajectory-balance training of a SMILES RNN policy, achieving top PMO benchmark score (16.213) and SARS-CoV-2 inhibitor design with fewer reward calls.
Problem
Sample-efficient de novo molecular optimization, where each reward (oracle) call is computationally expensive, so generative methods must find high-reward molecules with few evaluations.
Approach
Uses a graph-based genetic search to refine policy samples into higher-reward molecules, then fine-tunes a SMILES generative policy with GFlowNet trajectory balance (off-policy, replay with rank-based weighting) plus a KL penalty to the pretrained policy, so the policy samples proportional to reward rather than maximizing likelihood as in GEGL.
Key results
- Achieves total AUC Top-10 score of 16.213 across 23 PMO oracles, best among baselines (Mol GA 15.686, REINVENT 15.185), with best AUC in 14 of 23 tasks.
- On SARS-CoV-2 targets, reaches top-100 average scores of 0.891 (PLPro) and 0.873 (RdRp) at 100 steps, versus 0.772 and 0.854 for the best baselines run for 1000 steps.
- On multi-objective GSK3b+JNK3 hypervolume, scores 0.718 vs HN-GFN 0.669 with 1K evaluations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add genetic search to the GFlowNet policy training (ablation) | Genetic GFN · PMO 23 oracles | AUC Top-10: 15.738 -> 16.213 | Helped |
| use genetic search instead of epsilon-greedy exploration (ablation) | Genetic GFN · PMO 23 oracles | AUC Top-10: 15.626 -> 16.213 | Helped |
| add a KL-divergence penalty to the training loss (ablation) | Genetic GFN · PMO 23 oracles | AUC Top-10: 15.928 -> 16.213 | Helped |
| increase rank-based sampling weight-shifting factor k from 0.01 to 0.1 (ablation) | Genetic GFN · PMO 23 oracles | AUC Top-10: 16.213 -> 14.652 | Hurt |
| increase rank-based sampling weight-shifting factor k from 0.01 to 0.1 (ablation) | Genetic GFN · PMO 23 oracles | diversity: 0.432 -> 0.581 | Helped |
| increase rank-based sampling weight-shifting factor k from 0.01 to 0.05 (ablation) | Genetic GFN · PMO 23 oracles | AUC Top-10: 16.213 -> 15.330 | Hurt |
| increase rank-based sampling weight-shifting factor k from 0.01 to 0.05 (ablation) | Genetic GFN · PMO 23 oracles | diversity: 0.432 -> 0.515 | Helped |
| use Genetic GFN rather than Mol GA | PMO molecular optimization · PMO 23 oracles | AUC Top-10 sum: 15.686 -> 16.213 | Helped |
| use Genetic GFN rather than MolRL-MGPT | SARS-CoV-2 inhibitor optimization · SARS-CoV-2 PLPro | Top-100 average score: 0.772 -> 0.891 | Helped |
| use Genetic GFN rather than MolRL-MGPT | SARS-CoV-2 inhibitor optimization · SARS-CoV-2 RdRp | Top-100 average score: 0.854 -> 0.873 | Helped |
| use Genetic GFN rather than LS-GFN | GFlowNet molecular optimization · PMO 23 oracles | AUC Top-10: 15.230 -> 16.213 | Helped |
Surprising findings
- Fragment-based GFlowNet baselines score far below SMILES-based methods on the PMO benchmark (sum ~9.9 vs ~16.2), suggesting representation choice matters a lot for GFlowNets here.
Limitations
- Relies on availability of effective domain-specific genetic operators, which are hard to design for other domains.
- Reported diversity is lower than some baselines (0.432 vs Mol GA 0.465 and Graph GA 0.661), trading diversity for score.
- Not designed for multi-objective optimization; multi-objective results use scalarized rewards.
Experimental setup
Datasets: ZINC 250K PMO benchmark oracles
Benchmarks: PMO benchmark (23 oracles) GSK3b+JNK3 multi-objective SARS-CoV-2 PLPro_7JIR and RdRp_6YYT docking
Metrics: AUC Top-10 AUC Top-1 AUC Top-100 sum of scores diversity hypervolume top-100 average score
Baselines: Mol GA REINVENT GEGL GP BO Fragment GFN Fragment GFN-AL LS-GFN HN-GFN
Builds on: GFlowNets Graph GA GEGL REINVENT PMO benchmark
Closest papers by meaning
- Generative Flow Networks for Lead Optimization in Drug Design (Student Abstract) (AAAI 2025 · similarity 0.86)
- Scalable and Cost-Efficient de Novo Template-Based Molecular Generation (NeurIPS 2025 · similarity 0.86)
- RGFN: Synthesizable Molecular Generation Using GFlowNets (NeurIPS 2024 · similarity 0.85)
- SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints (ICLR 2025 · similarity 0.85)
- Pretraining Generative Flow Networks with Inexpensive Rewards for Molecular Graph Generation (ICML 2025 · similarity 0.85)
- Pre-Training and Fine-Tuning Generative Flow Networks (ICLR 2024 · similarity 0.84)
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.