Research map / Probabilistic and scientific modeling / Molecules and proteins
RetroBridge: Modeling Retrosynthesis with Markov Bridges
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 the Markov Bridge Model, a discrete generative framework that learns the dependency between two intractable discrete distributions, and applies it as RetroBridge for template-free single-step retrosynthesis on USPTO-50k.
Problem
Single-step retrosynthesis requires predicting sets of reactants for a product molecule; the product-reactant dependency is modeled as a mapping between two intractable discrete distributions, which standard diffusion models (designed for one intractable distribution from a tractable noise prior) fi
Approach
Model the product-to-reactant mapping as a Markov bridge pinned at a product start point and a reactant endpoint, using a categorical transition kernel with a learned predictor of the final state trained by a variational lower bound; sampling starts from the product itself rather than from noise, and confidence scores come from the sample frequency of each reactant set.
Key results
- RetroBridge reaches 74.1% top-3 and 80.6% top-5 exact match on USPTO-50k, the best among template-free methods, while top-1 is 50.8%, below GraphRetro's 53.7% among template-based methods.
- Round-trip coverage/accuracy at k=1 of 85.1% for RetroBridge, exceeding template-based GLN (82.5%) and LocalRetro (82.1%).
- On the validation set, RetroBridge-VLB with context reaches 73.04% top-3 vs DiGress with context at 68.56%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the Markov Bridge formulation instead of DiGress with product context | retrosynthesis prediction · USPTO-50k validation | top-5 exact-match accuracy: 73.93 -> 79.44 | Helped |
| add product context at each sampling step (ablation) | RetroBridge-VLB · USPTO-50k validation | top-5 exact-match accuracy: 75.21 -> 79.44 | Helped |
| replace VLB with CE loss (ablation) | RetroBridge with product context · USPTO-50k validation | top-5 exact-match accuracy: 79.44 -> 76.58 | Hurt |
| replace VLB with CE loss (ablation) | RetroBridge without product context · USPTO-50k validation | top-5 exact-match accuracy: 75.21 -> 72.33 | Hurt |
| add molecular weight and atom valency features (ablation) | RetroBridge · USPTO-50k validation | top-5 exact-match accuracy | Hurt |
| reduce sampling steps from T=500 to T=250 (ablation) | RetroBridge with CE loss · USPTO-50k validation | top-k accuracy | No clear effect |
| reduce sampling steps by 5-fold (ablation) | RetroBridge with CE loss · USPTO-50k validation | top-k accuracy | Hurt |
| reduce the number of generated samples by 5-fold (ablation) | RetroBridge with VLB loss · USPTO-50k test | top-k accuracy | No clear effect |
| use one-shot graph prediction instead of iterative sampling (ablation) | graph transformer retrosynthesis model · USPTO-50k validation | exact-match accuracy | Hurt |
| use product context with DiGress (ablation) | DiGress · USPTO-50k validation | top-5 exact-match accuracy | Helped |
| use RetroBridge instead of template-free baselines | retrosynthesis prediction · USPTO-50k test | top-5 exact-match accuracy: 78.1 -> 80.6 | Helped |
| use RetroBridge instead of LocalRetro | retrosynthesis prediction · USPTO-50k test | top-5 round-trip coverage: 94.7 -> 97.1 | Helped |
Surprising findings
- Without product context, DiGress recovers no reactants, while RetroBridge retains reasonable accuracy because large parts of the product structure persist along the sampling trajectory.
Limitations
- Single-step only; multi-step planning not evaluated.
- Does not predict reaction conditions or reagents, nor reaction type.
- Iterative sampling requires hundreds of forward passes, making inference slower than one-shot methods.
Experimental setup
Datasets: USPTO-50k
Benchmarks: USPTO-50k
Metrics: top-k exact match accuracy round-trip accuracy round-trip coverage
Baselines: GLN LocalRetro GraphRetro MEGAN G2G Augmented Transformer SCROP Graph2SMILES Retroformer DiGress
Builds on: DiGress discrete diffusion models Schrödinger bridge Graph Transformer
Closest papers by meaning
- Advancing Retrosynthesis with Retrieval-Augmented Graph Generation (AAAI 2025 · similarity 0.92)
- RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis (NeurIPS 2025 · similarity 0.91)
- Retrosynthesis Planning via Worst-path Policy Optimisation in Tree-structured MDPs (NeurIPS 2025 · similarity 0.90)
- RetroOOD: Understanding Out-of-Distribution Generalization in Retrosynthesis Prediction (AAAI 2024 · similarity 0.90)
- Retro-fallback: retrosynthetic planning in an uncertain world (ICLR 2024 · similarity 0.90)
- GDiffRetro: Retrosynthesis Prediction with Dual Graph Enhanced Molecular Representation and Diffusion Generation (AAAI 2025 · similarity 0.90)
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.