Research map / Probabilistic and scientific modeling / Molecules and proteins
Reaction Prediction via Interaction Modeling of Symmetric Difference Shingle Sets
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 ReaDISH, a permutation-invariant reaction property predictor that encodes symmetric-difference molecular shingles with interaction-aware (geometric, structural, connectivity) attention, pre-trained on pseudo-reaction-type clustering labels.
Problem
Reaction property prediction models are sensitive to molecule and SMILES atom orderings and fail to model substructural interactions that govern reactivity, leading to poor out-of-sample generalization.
Approach
Encode the symmetric difference of reactant and product shingle sets (from DRFP) as continuous embeddings, and add a pairwise attention bias built from 3D geometric distance, Tanimoto structural distance, and same-molecule connectivity via Gaussian kernels with pair types.
Key results
- ReaDISH improves R2 by 11.11% on BH and 6.40% on SM out-of-sample splits under permutation perturbations (average 8.76%).
- On BH out-of-sample, ReaDISH reaches R2 0.480 vs 0.479 for ReaMVP and 0.458 for UAM; on SM out-of-sample R2 0.549 vs 0.534 for ReaMVP.
- Random split BH R2 0.976 and USPTO_TPL accuracy 0.992 are top or tied-top among compared models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add pairwise interaction modeling to self-attention (ablation) | ReaDISH · BH reactant-based and SM ligand-based out-of-sample | R²: 0.430 ± 0.038 -> 0.480 ± 0.032; 0.506 ± 0.026 -> 0.549 ± 0.019 | Helped |
| add geometric distance features (ablation) | ReaDISH · BH reactant-based and SM ligand-based out-of-sample | R²: 0.457 ± 0.040 -> 0.480 ± 0.032; 0.535 ± 0.028 -> 0.549 ± 0.019 | Helped |
| add structural distance features (ablation) | ReaDISH · BH reactant-based and SM ligand-based out-of-sample | R²: 0.464 ± 0.034 -> 0.480 ± 0.032; 0.530 ± 0.023 -> 0.549 ± 0.019 | Helped |
| use symmetric-difference shingles instead of all reactant and product shingles (ablation) | ReaDISH · BH reactant-based and SM ligand-based out-of-sample | R²: 0.465 ± 0.039 -> 0.480 ± 0.032; 0.538 ± 0.017 -> 0.549 ± 0.019 | Helped |
| add pseudo-reaction-type pre-training (ablation) | ReaDISH · BH reactant-based and SM ligand-based out-of-sample | R²: 0.472 ± 0.035 -> 0.480 ± 0.032; 0.540 ± 0.020 -> 0.549 ± 0.019 | Helped |
| set maximum shingle radius to 2 rather than 3 (ablation) | ReaDISH · BH reactant-based and SM ligand-based out-of-sample | R²: 0.480 ± 0.032 -> 0.465 ± 0.014; 0.549 ± 0.019 -> 0.523±0.228 | Hurt |
| set maximum shingle radius to 4 rather than 3 (ablation) | ReaDISH · BH reactant-based and SM ligand-based out-of-sample | R²: 0.480 ± 0.032 -> 0.421 ± 0.006; 0.549 ± 0.019 -> 0.505±0.043 | Hurt |
| use ReaDISH instead of the strongest reported baseline on each benchmark | reaction prediction models · BH, SM, NiCOlit, ELN, N,S-acetal, and C-heteroatom out-of-sample | R²: BH: 0.479 ± 0.056 -> 0.480 ± 0.032; SM: 0.534 ± 0.018 -> 0.549 ± 0.019; NiCOlit: 0.563 -> 0.574; ELN: 0.212 ± 0.057 -> 0 | Helped |
Surprising findings
- On the ELN dataset all methods show R2 below 0.3, indicating poor generalization on real-world data.
- A simple cross-attention baseline underperforms ReaDISH but the gap on some sets is small relative to variance.
Limitations
- Computational overhead: shingle count can exceed atom count, increasing memory and training time; mitigated by caps.
- Poor performance on the real-world ELN dataset (R2 below 0.3 for all methods).
- Limited fine-tuning data constrains adaptation to diverse chemical contexts.
Experimental setup
Datasets: USPTO CJHIF BH SM NiCOlit ELN N,S-acetal C-heteroatom USPTO_TPL
Benchmarks: BH SM NiCOlit ELN N,S-acetal C-heteroatom USPTO_TPL
Metrics: MAE RMSE R2 ACC MCC CEN
Baselines: YieldBERT YieldBERT-DA UA-GNN UAM ReaMVP Cross-attention
Builds on: DRFP Uni-Mol Gaussian Kernel with Pair Type ECFP
Closest papers by meaning
- Reaction Graph: Towards Reaction-Level Modeling for Chemical Reactions with 3D Structures (ICML 2025 · similarity 0.88)
- Uncertainty-Aware Yield Prediction with Multimodal Molecular Features (AAAI 2024 · similarity 0.87)
- Learning Condensed Graph via Differentiable Atom Mapping for Reaction Yield Prediction (ICML 2025 · similarity 0.86)
- RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis (NeurIPS 2025 · similarity 0.85)
- Beyond Chemical QA: Evaluating LLM's Chemical Reasoning with Modular Chemical Operations (NeurIPS 2025 · similarity 0.85)
- ReactGPT: Understanding of Chemical Reactions via In-Context Tuning (AAAI 2025 · 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.