Research map / Probabilistic and scientific modeling / Molecules and proteins
Molecule Generation with Fragment Retrieval Augmentation
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 f-RAG, a fragment-level retrieval-augmented generation framework that augments frozen SAFE-GPT with hard fragments (explicitly included), soft fragments (injected via a trainable cross-attention module), an iteratively updated fragment vocabulary, and genetic-algorithm fragment modification for molecular optimization.
Problem
Fragment-based molecule generation is limited to recombining fragments already in the database, restricting exploration of novel high-quality chemical space for drug discovery.
Approach
Retrieve two kinds of fragments from a property-scored fragment vocabulary: hard fragments that are fixed into the output molecule and soft fragments that guide new-fragment generation through a lightweight cross-attention fragment injection module inserted into a frozen SAFE-GPT; the vocabulary is continually refreshed with generated fragments and combined with Graph GA-style genetic modification
Key results
- On the 23 PMO tasks, f-RAG reaches a sum AUC top-10 of 16.928, versus 16.213 for Genetic GFN and 14.708 for Mol GA, and is best on 12 of 23 tasks.
- Average top-100 SA score of 2.026 is the best among compared methods (Genetic GFN 3.770, Mol GA 4.605).
- On novel top 5% docking score across five protein targets, f-RAG obtains the best result on all five, e.g. parp1 -12.945 vs GEAM -12.891.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add hard fragment retrieval to the soft+GA variant (ablation) | f-RAG on PMO benchmark · PMO benchmark (23 tasks) | sum AUC top-10: 15.231 -> 16.928 | Helped |
| add soft fragment retrieval to the hard+GA variant (ablation) | f-RAG on PMO benchmark · PMO benchmark (23 tasks) | sum AUC top-10: 16.890 -> 16.928 | Helped |
| add genetic fragment modification to the hard+soft variant (ablation) | f-RAG on PMO benchmark · PMO benchmark (23 tasks) | sum AUC top-10: 14.048 -> 16.928 | Helped |
| add hard and soft fragment retrieval to GA-only generation (ablation) | f-RAG on PMO benchmark · PMO benchmark (23 tasks) | sum AUC top-10: 15.395 -> 16.928 | Helped |
| add soft fragment retrieval to hard+GA generation (ablation) | f-RAG on PMO benchmark · PMO benchmark (23 tasks) | average diversity: 0.497 -> 0.532 | Helped |
| add soft fragment retrieval to hard+GA generation (ablation) | f-RAG on PMO benchmark · PMO benchmark (23 tasks) | average novelty: 0.720 -> 0.800 | Helped |
| add genetic fragment modification to hard+soft generation (ablation) | f-RAG on PMO benchmark · PMO benchmark (23 tasks) | average diversity: 0.744 -> 0.532 | Hurt |
| add genetic fragment modification to hard+soft generation (ablation) | f-RAG on PMO benchmark · PMO benchmark (23 tasks) | average SA score: 3.443 -> 2.026 | Helped |
| lower the similarity-filter threshold from the full method setting to δ=0.4 (ablation) | f-RAG on PMO benchmark · PMO benchmark (23 tasks) | sum AUC: 16.928 -> 15.765 | Hurt |
| lower the similarity-filter threshold from the full method setting to δ=0.4 (ablation) | f-RAG on PMO benchmark · PMO benchmark (23 tasks) | average diversity: 0.532 -> 0.724 | Helped |
| insert the fragment injection module at layer 1 rather than layer 6 (ablation) | f-RAG on PMO benchmark · PMO benchmark (23 tasks) | AUC top-10 | Mixed |
| use dynamic fragment vocabulary updates during generation | f-RAG on jak2 · jak2 | docking score | Helped |
Surprising findings
- Ablation shows that removing soft fragment retrieval lowers AUC top-10 sum from 16.928 to 16.890 only when combined with GA-only settings, but removing hard retrieval drops it to 15.231, showing hard retrieval matters most for optimization.
Limitations
- Performance depends heavily on the generation quality of the frozen pre-trained backbone SAFE-GPT.
- Diversity, novelty and synthesizability gains can conflict with optimization performance; the similarity-based fragment filter trades optimization for diversity.
Experimental setup
Datasets: ZINC250k
Benchmarks: PMO benchmark
Metrics: AUC top-10 top-100 diversity top-100 novelty top-100 SA score novel top 5% docking score
Baselines: Graph GA Mol GA Genetic GFN REINVENT SELFIES-REINVENT GP BO STONED LSTM HC SMILES GA JT-VAE GEAM MOOD
Builds on: SAFE-GPT Graph GA RetMol PMO benchmark
Closest papers by meaning
- Drug Discovery with Dynamic Goal-aware Fragments (ICML 2024 · similarity 0.89)
- GenMol: A Drug Discovery Generalist with Discrete Diffusion (ICML 2025 · similarity 0.88)
- Advancing Retrosynthesis with Retrieval-Augmented Graph Generation (AAAI 2025 · similarity 0.84)
- Beam Enumeration: Probabilistic Explainability For Sample Efficient Self-conditioned Molecular Design (ICLR 2024 · similarity 0.83)
- MADGEN: Mass-Spec attends to De Novo Molecular generation (ICLR 2025 · similarity 0.83)
- Retrieval Augmented Zero-Shot Enzyme Generation for Specified Substrate (ICML 2025 · similarity 0.83)
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.