Research map / Probabilistic and scientific modeling / Molecules and proteins
Drug Discovery with Dynamic Goal-aware Fragments
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 GEAM, a fragment-based molecular generator combining goal-aware fragment extraction (FGIB, graph information bottleneck), SAC-based fragment assembly, and GA-based fragment modification with a dynamically updated fragment vocabulary.
Problem
Fragment-based molecular generation relies on fragment vocabularies that ignore target properties or use heuristic rules, and cannot update the vocabulary with goal-aware fragments discovered during generation.
Approach
Extract target-relevant fragments with a fragment-wise graph information bottleneck (FGIB), assemble them with soft actor-critic, explore beyond the vocabulary with a graph genetic algorithm, and feed novel offspring subgraphs back into FGIB to update the vocabulary on-the-fly.
Key results
- On the jak2 novel hit ratio, GEAM reaches 42.95% vs 9.20% for the best baseline MOOD.
- On PMO MPO tasks, GEAM achieves average AUC top-100 of 0.615 vs 0.577 for STONED (top baseline).
- GEAM improves novelty over GEAM-static on all five targets (e.g., 88.6% vs 84.3% on parp1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add dynamic goal-aware vocabulary updates during generation (ablation) | GEAM-static · JAK2 | novel hit ratio (%): 42.056~ ($\pm$~1.855) -> 42.950~ ($\pm$~1.117) | Helped |
| Add dynamic goal-aware vocabulary updates during generation (ablation) | GEAM-static · JAK2 | novelty (%): 86.633~ ($\pm$~1.817) -> 89.222~ ($\pm$~1.824) | Helped |
| Add dynamic goal-aware vocabulary updates during generation (ablation) | GEAM-static · JAK2 | novel top 5% docking score (kcal/mol): -11.812~ ($\pm$~0.055) -> -11.816~ ($\pm$~0.067) | Helped |
| Use FGIB goal-aware fragments instead of FREED fragments (ablation) | FREED · JAK2 ligand generation | optimization performance | Helped |
| Use FGIB goal-aware vocabulary instead of FREED, MiCaM, random BRICS, or property-only scoring (ablation) | GEAM · JAK2 ligand generation | optimization performance | Helped |
| Use FGIB importance-weighted fragment scoring rather than property-only scoring (ablation) | GEAM · JAK2 ligand generation | optimization performance | Helped |
| Add the fragment assembly module (ablation) | GEAM-w/o A · JAK2 ligand generation | optimization performance | Helped |
| Add SAC instead of random fragment assembly (ablation) | GEAM-random A · JAK2 ligand generation | optimization performance | Helped |
| Add the fragment modification module (ablation) | GEAM-w/o M · JAK2 ligand generation | optimization performance | Helped |
| Use a fixed vocabulary of top-1,000 fragments rather than dynamic updates (ablation) | GEAM-static-1000 · JAK2 ligand generation | optimization performance | Hurt |
| Increase the fixed vocabulary from top-300 to top-1,000 fragments (ablation) | GEAM-static · JAK2 ligand generation | diversity | Helped |
Surprising findings
- Fragment vocabulary using only property ranking (GEAM property variant) performs worse than with FGIB, showing the importance of extracting fragments with a model rather than ranking by property alone.
- RationaleRL shows high #Circles on 5ht1b and jak2 but poor optimization and novelty, so diversity metrics alone are misleading.
Limitations
- Fragment extraction relies on BRICS decomposition, which can yield disconnected or limited fragments.
- Evaluation is on docking-based and PMO benchmarks; real-world wet-lab validation is not reported.
- Dynamic vocabulary update adds fragments that could be toxic; the paper only suggests filtering as mitigation.
Experimental setup
Datasets: ZINC250k
Benchmarks: DUD-E docking tasks (parp1, fa7, 5ht1b, braf, jak2) PMO benchmark MPO tasks
Metrics: novel hit ratio novel top 5% docking score novelty #Circles AUC top-100
Baselines: REINVENT Graph GA MORLD HierVAE RationaleRL FREED PS-VAE MOOD STONED GCPN JTVAE GraphAF
Builds on: graph information bottleneck soft actor-critic FREED Graph GA BRICS
Closest papers by meaning
- Molecule Generation with Fragment Retrieval Augmentation (NeurIPS 2024 · similarity 0.89)
- GenMol: A Drug Discovery Generalist with Discrete Diffusion (ICML 2025 · similarity 0.88)
- CombiMOTS: Combinatorial Multi-Objective Tree Search for Dual-Target Molecule Generation (ICML 2025 · similarity 0.85)
- Generating Novel Leads for Drug Discovery Using LLMs with Logical Feedback (AAAI 2024 · similarity 0.85)
- Rethinking the generalization of drug target affinity prediction algorithms via similarity aware evaluation (ICLR 2025 · similarity 0.84)
- MF-LAL: Drug Compound Generation Using Multi-Fidelity Latent Space Active Learning (ICML 2025 · 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.