Research map / Learning theory and algorithms / Combinatorial and search optimization
Double-Ended Synthesis Planning with Goal-Constrained Bidirectional Search
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 DESP, a bidirectional AND-OR search for synthesis planning under a specified starting-material constraint, guided by an offline-learned synthetic distance network, and releases Pistachio-based benchmarks.
Problem
Computer-aided synthesis planning usually assumes any building blocks are acceptable, so it fails when a route must use a specific starting material (structure-goal) for a target molecule.
Approach
Runs two AND-OR search graphs at once (top-down retrosynthesis from the target and bottom-up forward synthesis from the starting material), guided by a learned synthetic distance network D(m1, m2) trained offline from reaction-corpus pathways, with F2E and F2F variants.
Key results
- On Pistachio Reachable at 300 expansions, DESP-F2F reaches 97.3% solve rate with 38 average expansions vs 92.7% and 73 for Retro*.
- On USPTO-190 at 500 expansions, DESP-F2E and DESP-F2F both reach 39.5% solve rate vs 35.8% for Retro*, with about 336-340 average expansions vs 351.
- On Pistachio Hard at 500 expansions, DESP-F2E reaches 50.0% solve rate vs 37.0% for Retro*.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add synthetic distance D to Retro* guidance (ablation) | Retro* · USPTO-190, N=500 | solve rate (%): 35.8 -> 37.4 | Helped |
| Add synthetic distance D to Retro* guidance (ablation) | Retro* · Pistachio Reachable, N=300 | solve rate (%): 92.7 -> 96.0 | Helped |
| Add synthetic distance D to Retro* guidance (ablation) | Retro* · Pistachio Hard, N=500 | solve rate (%): 37.0 -> 42.0 | Helped |
| Add bidirectional search to Retro*+D (ablation) | Retro*+D · USPTO-190, N=500 | solve rate (%): 37.4 -> 39.5 | Helped |
| Add bidirectional search to Retro*+D (ablation) | Retro*+D · Pistachio Reachable, N=300 | solve rate (%): 96.0 -> 96.0; 96.0 -> 97.3 | Mixed |
| Add bidirectional search to Retro*+D (ablation) | Retro*+D · Pistachio Hard, N=500 | solve rate (%): 42.0 -> 50.0; 42.0 -> 48.0 | Helped |
| Use DESP-F2E instead of DESP-F2F (ablation) | DESP · USPTO-190 | average number of reactions: 5.95 -> 5.56 | Helped |
| Use DESP-F2E instead of DESP-F2F (ablation) | DESP · Pistachio Easy | average number of reactions: 5.17 -> 4.86 | Helped |
| Use DESP-F2E instead of DESP-F2F (ablation) | DESP · Pistachio Hard | average number of reactions: 4.78 -> 4.67 | Helped |
| Use DESP-F2F instead of DESP-F2E (ablation) | DESP · USPTO-190 | average expansions: 340 -> 336 | Helped |
| Use DESP-F2F instead of DESP-F2E (ablation) | DESP · Pistachio Reachable | average expansions: 41 -> 38 | Helped |
| Use DESP-F2F instead of DESP-F2E (ablation) | DESP · Pistachio Hard | average expansions: 300 -> 293 | Helped |
Surprising findings
- DESP-F2F finds longer routes than Retro* on several benchmarks despite fewer expansions, attributed to not accounting for pathway depth in front-to-front comparisons.
- Standard single-ended searches such as BFS and MCTS perform poorly under the starting-material constraint, with BFS at 24.2% solve rate on USPTO-190 at 500 expansions.
Limitations
- Bottom-up search only performs forward expansions and cannot plan convergent routes with multiple merged branches.
- Requires GPU for k-NN search over about 23 million building blocks and for F2F node comparisons.
- Neither F2E nor F2F guarantees cost-optimal routes; F2F yields longer routes than F2E.
Experimental setup
Datasets: USPTO-Full USPTO-190 Pistachio eMolecules building blocks
Benchmarks: USPTO-190 Pistachio Reachable Pistachio Hard
Metrics: solve rate average number of expansions average number of reactions per route solve rate by SCScore and SAScore complexity
Baselines: Random BFS MCTS Retro* GRASP Bi-BFS Retro*+D
Builds on: Retro* NeuralSym RDChiral GRASP Bi-BFS
Closest papers by meaning
- LLM-Augmented Chemical Synthesis and Design Decision Programs (ICML 2025 · similarity 0.90)
- Retro-fallback: retrosynthetic planning in an uncertain world (ICLR 2024 · similarity 0.89)
- Retrosynthesis Planning via Worst-path Policy Optimisation in Tree-structured MDPs (NeurIPS 2025 · similarity 0.89)
- SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints (ICLR 2025 · similarity 0.87)
- URSA: Chemistry-Aware Benchmark for Utilitarian Retrosynthesis Assessment (NeurIPS 2026 · similarity 0.87)
- RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis (NeurIPS 2025 · similarity 0.87)
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.