Research map / Probabilistic and scientific modeling / Molecules and proteins
Preference Optimization for Molecule Synthesis with Conditional Residual Energy-based Models
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 conditional residual energy-based models (CREBM) trained with a Bradley-Terry preference loss to rerank synthetic routes from existing retrosynthesis planning strategies, improving top-1 route accuracy in a plug-and-play way.
Problem
Retrosynthetic planning generates synthetic routes greedily via locally normalized one-step models, which ignores long-range route quality and criteria such as starting-material feasibility, cost, and step count.
Approach
Add a conditional residual energy function on top of a frozen retrosynthesis strategy's route distribution, trained with a pairwise preference (Bradley-Terry) loss over routes ranked by a heuristic feasibility reward (forward-model similarity plus similarity to the reference starting materials), so it can rerank routes from any strategy without retraining it.
Key results
- Neuralsym+CREBM with Retro*-0 reaches 44.5% top-1 vs 42.0% for Neuralsym baseline (+2.5), the claimed state of the art.
- Consistent gains across strategies, e.g. Transformer+CREBM top-1 35.0 vs 31.3 under Retro*; GraphRetro+CREBM 16.3 vs 15.3.
- Gains grow on longer routes, e.g. GLN+CREBM +14.5 top-1 at depth 6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CREBM to Neuralsym | Neuralsym+Retro*-0 · RetroBench test | Top-1 exact match accuracy (%): 42.0 -> 44.5 | Helped |
| add CREBM to GLN | GLN+Retro*-0 · RetroBench test | Top-1 exact match accuracy (%): 39.5 -> 43.2 | Helped |
| add CREBM to GraphRetro | GraphRetro+Retro*-0 · RetroBench test | Top-1 exact match accuracy (%): 15.3 -> 16.3 | Helped |
| add CREBM to Transformer | Transformer+Retro*-0 · RetroBench test | Top-1 exact match accuracy (%): 31.2 -> 34.9 | Helped |
| add CREBM to FusionRetro | FusionRetro+Retro*-0 · RetroBench test | Top-1 exact match accuracy (%): 37.5 -> 39.6 | Helped |
| add CREBM to Neuralsym | Neuralsym+Retro*-0 · RetroBench test | Top-5 exact match accuracy (%): 54.3 -> 55.2 | Helped |
| add E to -log P_Retro in route ranking (ablation) | Neuralsym+Retro*-0 · RetroBench test | Top-1 exact match accuracy (%): 42.0 -> 44.5 | Helped |
| add E to -log P_Retro in route ranking (ablation) | Neuralsym+Retro*-0 · RetroBench test | Top-5 exact match accuracy (%): 54.3 -> 55.2 | Helped |
| rank using E alone, removing -log P_Retro from ranking (ablation) | Neuralsym+Retro*-0 · RetroBench test | Top-1 exact match accuracy (%): 42.0 -> 34.1 | Hurt |
| rank using E alone, removing -log P_Retro from ranking (ablation) | Neuralsym+Retro*-0 · RetroBench test | Top-5 exact match accuracy (%): 54.3 -> 53.1 | Hurt |
| subtract E from -log P_Retro instead of adding it (ablation) | Neuralsym+Retro*-0 · RetroBench test | Top-1 exact match accuracy (%): 42.0 -> 19.2 | Hurt |
| subtract E from -log P_Retro instead of adding it (ablation) | Neuralsym+Retro*-0 · RetroBench test | Top-5 exact match accuracy (%): 54.3 -> 44.6 | Hurt |
Surprising findings
- Using the energy alone for ranking drops top-1 from 42.0 to 34.1, and subtracting the energy from the log-probability drops it to 19.2.
- An RL variant using the same reward caused a significant accuracy drop versus the base model, per the authors' attempt.
Limitations
- Reward relies on a heuristic feasibility score (forward model similarity), not true yield, cost or expert ranking.
- Energy ignores intermediate molecules; only starting-material feasibility is optimized.
- Only one strategy (Neuralsym+Retro*-0) used for sampling preference routes.
Experimental setup
Datasets: RetroBench (USPTO-full reaction network)
Benchmarks: RetroBench
Metrics: top-1 to top-5 set-wise exact match accuracy top-k accuracy by route depth
Baselines: Retrosim Neuralsym GLN G2Gs GraphRetro Transformer Megan FusionRetro Retro* Retro*-0
Builds on: Energy-based models Bradley-Terry model Retro* Neuralsym FusionRetro
Closest papers by meaning
- LLM-Augmented Chemical Synthesis and Design Decision Programs (ICML 2025 · similarity 0.88)
- Enhancing Ligand Validity and Affinity in Structure-Based Drug Design with Multi-Reward Optimization (ICML 2025 · similarity 0.86)
- Aligning Transformers with Continuous Feedback via Energy Rank Alignment (NeurIPS 2025 · similarity 0.86)
- RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis (NeurIPS 2025 · similarity 0.86)
- Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.86)
- Aligning Target-Aware Molecule Diffusion Models with Exact Energy Optimization (NeurIPS 2024 · similarity 0.86)
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.