Research map / Graph learning / Graph self-supervised learning
Learning Condensed Graph via Differentiable Atom Mapping for Reaction Yield Prediction
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 YieldNet, which learns an approximate atom mapping via a differentiable Gumbel/Sinkhorn-style alignment network, builds a continuous condensed graph of reaction as a transition-state surrogate, and predicts yield from multi-step reaction paths trained only on yield values.
Problem
Reaction yield depends on the activation barrier, which requires the transition-state graph and atom mapping between reactants and products; yield datasets lack both atom mappings and transition states, and exact atom mapping is NP-hard to compute.
Approach
Replace the unknown atom mapping with a differentiable Sinkhorn-based node alignment network trained end-to-end from yield supervision alone, use the resulting soft mapping to build a weighted condensed graph of reaction (CGR) as a transition-state surrogate, and encode multi-step reaction paths with a transformer and set aggregators.
Key results
- YieldNet achieves the lowest MAE on seven of eight datasets, e.g., GP MAE 23.152 vs 27.837 for the best baseline DeepReac+ and DF MAE 6.941 vs 12.470 for YieldBERT.
- On GP with ground-truth atom mappings, the learned alignment has average Frobenius error 3.708 versus 5.096 for the attention-based alignment.
- Ablating the CGR embeddings (w/o CGR) raises MAE on NS1 from 9.245 to 11.356.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use YIELDNET instead of the strongest reported baseline | Yield prediction · GP | MAE: 27.837 ± 0.346 -> 23.152 ± 0.393 | Helped |
| Use YIELDNET instead of the strongest reported baseline | Yield prediction · NS1 | MAE: 11.162 ± 0.735 -> 9.245 ± 0.518 | Helped |
| Use YIELDNET instead of the strongest reported baseline | Yield prediction · NS2 | MAE: 9.416 ± 1.045 -> 8.387 ± 0.907 | Helped |
| Use YIELDNET instead of the strongest reported baseline | Yield prediction · NS3 | MAE: 8.780 ± 0.918∗ -> 7.914 ± 0.931 | Helped |
| Use YIELDNET instead of the strongest reported baseline | Yield prediction · NS4 | MAE: 8.603 ± 0.230 -> 7.015 ± 0.495 | Helped |
| Use YIELDNET instead of the strongest reported baseline | Yield prediction · NS5 | MAE: 4.453 ± 0.293∗ -> 4.382 ± 0.249 | Helped |
| Use YIELDNET instead of the strongest reported baseline | Yield prediction · SC | MAE: 10.437 ± 0.328 -> 8.751 ± 0.438 | Helped |
| Use YIELDNET instead of the strongest reported baseline | Yield prediction · DF | MAE: 12.470 ± 0.412 -> 6.941 ± 0.192 | Helped |
| Use the true CGR instead of the learned approximate CGR (ablation) | YIELDNET · GP | MAE: 23.152 ± 0.393 -> 17.396 ± 0.357 | Helped |
| Add CGR and atom-alignment representations instead of directly feeding reactant and product embeddings (ablation) | YIELDNET · NS1, NS2, NS3, SC | MAE: 11.356 -> 9.245; 9.459 -> 8.387; 9.342 -> 7.914; 10.686 -> 8.751 | Helped |
| Use Seq2Seq StepAggr instead of DeepSet (ablation) | YIELDNET reaction path encoder · NS1, NS2, NS3, SC | MAE: 9.643 -> 9.245; 9.008 -> 8.387; 8.283 -> 7.914; 8.551 -> 8.751 | Helped |
| Use Seq2Seq StepAggr instead of sum aggregation (ablation) | YIELDNET reaction path encoder · NS1, NS2, NS3, SC | MAE: 10.192 -> 9.245; 8.858 -> 8.387; 8.349 -> 7.914; 8.853 -> 8.751 | Mixed |
Surprising findings
- The skyline model using the true CGR (GP MAE 17.396) is substantially better than the learned-CGR version, indicating a gap from approximate atom mapping.
Limitations
- Yield depends on factors beyond activation energy (reaction time, purification) that current datasets lack.
- Learned atom mapping is approximate; a gap remains between the learned-CGR model and the true-CGR skyline.
- Temperature and other conditions are excluded from features.
Experimental setup
Datasets: GP (gas-phase isomerization, Grambow et al. 2020) NS1-NS5 (N,S-acetal formation, Zahrt et al. 2019) SC (Suzuki coupling) DF (deoxyfluorination, Nielsen et al. 2018)
Benchmarks: GP NS1 NS2 NS3 NS4 NS5 SC DF
Metrics: MAE Frobenius norm of atom mapping error
Baselines: YieldBERT DeepReac+ GCN HGT TAG GIN RDKit RXNMapper Attention alignment
Builds on: Sinkhorn normalization condensed graph of reaction Set2Set DeepSet
Closest papers by meaning
- Predicting Organic Reaction Outcomes with Weisfeiler-Lehman Network (NeurIPS 2017 · similarity 0.90)
- Uncertainty-Aware Yield Prediction with Multimodal Molecular Features (AAAI 2024 · similarity 0.89)
- Reaction Graph: Towards Reaction-Level Modeling for Chemical Reactions with 3D Structures (ICML 2025 · similarity 0.88)
- Reaction Prediction via Interaction Modeling of Symmetric Difference Shingle Sets (NeurIPS 2025 · similarity 0.86)
- The Catechol Benchmark: Time-series Solvent Selection Data for Few-shot Machine Learning (NeurIPS 2025 · similarity 0.86)
- Neural Atoms: Propagating Long-range Interaction in Molecular Graphs through Efficient Communication Channel (ICLR 2024 · similarity 0.85)
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.