Research map / Probabilistic and scientific modeling / Molecules and proteins
Advancing Retrosynthesis with Retrieval-Augmented Graph 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 RARB, a retrieval-augmented diffusion (Markov bridge) molecular graph generator that retrieves similar commercial molecules to prompt reactant generation for template-free single-step retrosynthesis.
Problem
Template-free single-step retrosynthesis with graph diffusion models generates reactants from scratch, ignoring that the product scaffold usually remains unchanged; also weak on out-of-distribution molecules.
Approach
Augment the product-conditioned denoiser with a prompt vector extracted from Morgan-fingerprint retrieved molecules (from a large USPTO-applications reactant library) via a self-attention prompt extractor over Uni-RXN embeddings with rank positional encodings, so the model learns which product substructures stay fixed.
Key results
- RARB (cross-entropy) reaches 58.3% top-1 exact-match accuracy on USPTO-50k vs 50.8% for RetroBridge, a 14.8% relative improvement.
- On the OOD USPTO-50k-cluster split, RARB top-1 accuracy is 54.9% vs 42.9% for RetroBridge (about 28% relative gain).
- Retrieval with dropout and sampling strategies raises diversity from 7.47 (RARB without strategies) to 12.0, compared with 12.9 for RetroBridge.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| augment RetroBridge with retrieval and prompt extraction | RetroBridge · USPTO-50k test | top-1 exact-match accuracy: 50.8 -> 58.3 | Helped |
| use Cross-Entropy instead of VLB (ablation) | RARB · USPTO-50k test | top-1 exact-match accuracy: 56.2 -> 58.3 | Helped |
| use USPTO-applications rather than USPTO-50k as the external retrieval dataset (ablation) | RARB · USPTO-50k | top-1 accuracy: 16.6% | Helped |
| use the self-attention prompt extractor rather than an MLP on concatenated molecule embeddings (ablation) | RARB · USPTO-50k | top-1 accuracy: 65.3% | Helped |
| reduce the base model to 2.8M parameters while retaining retrieval augmentation (ablation) | RARB · USPTO-50k | top-1 exact-match accuracy: 50.8 -> 54.9 | Helped |
| reduce denoising steps from 500 to 200 with retrieval augmentation (ablation) | RARB · USPTO-50k | top-1 exact-match accuracy: 50.8 -> 57.2 | Helped |
| add prompt-extractor dropout=0.5 (ablation) | RARB · USPTO-50k | diversity: 7.47 -> 10.1 | Helped |
| add prompt-extractor dropout=0.5 (ablation) | RARB · USPTO-50k | top-5 exact-match accuracy: 74.1 -> 78.8 | Helped |
| also randomly sample 3 of the top-5 retrieved molecules (ablation) | RARB with prompt-extractor dropout=0.5 · USPTO-50k | diversity: 10.1 -> 12.0 | Helped |
| remove prompt-extractor dropout and random top-5 retrieval selection (ablation) | RARB · USPTO-50k | diversity: 12.9 -> 7.47 | Hurt |
| use retrieval augmentation | RetroBridge · USPTO-50k-cluster test | top-1 exact-match accuracy: 42.9 -> 54.9 | Helped |
| use retrieval augmentation | RetroBridge · USPTO-50k test, least similar retrieval group | relative top-1 accuracy improvement: 0.8 | Hurt |
Surprising findings
- Plain RARB without diversity strategies generates less diverse reactants than the base RetroBridge, reflecting over-reliance on the retrieved prompt.
- Improvements are smallest for complex products with multiple reaction centers (0.4% top-1 gain) and for the least-similar retrieval group before diversity strategies.
Limitations
- Gains shrink for complex products with multiple reaction centers.
- Performance depends on retrieval quality; least-similar retrievals slightly hurt accuracy.
- Plain retrieval reduces diversity of generated reactants.
Experimental setup
Datasets: USPTO-50k USPTO-applications (2001-2016) USPTO-50k-cluster
Benchmarks: USPTO-50k USPTO-50k-cluster
Metrics: top-k exact match accuracy round-trip accuracy round-trip coverage diversity
Baselines: RetroBridge DiffAlign GLN LocalRetro Retrodiff GraphRetro Graph2SMILES Retroformer
Builds on: RetroBridge Uni-RXN DiGress Retrieval-augmented diffusion (Blattmann et al. 2022)
Closest papers by meaning
- GDiffRetro: Retrosynthesis Prediction with Dual Graph Enhanced Molecular Representation and Diffusion Generation (AAAI 2025 · similarity 0.92)
- RetroBridge: Modeling Retrosynthesis with Markov Bridges (ICLR 2024 · similarity 0.92)
- RetroOOD: Understanding Out-of-Distribution Generalization in Retrosynthesis Prediction (AAAI 2024 · similarity 0.90)
- RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis (NeurIPS 2025 · similarity 0.90)
- LLM-Augmented Chemical Synthesis and Design Decision Programs (ICML 2025 · similarity 0.89)
- Retrieval-Retro: Retrieval-based Inorganic Retrosynthesis with Expert Knowledge (NeurIPS 2024 · similarity 0.89)
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.