ReaDISH

Data and codes for the paper "Reaction Prediction via Interaction Modeling of Symmetric Difference Shingle Sets" (NeurIPS 2025).

Overview

ReaDISH is a new reaction prediction framework designed to overcome two persistent limitations in machine-learning-based organic reaction modeling: sensitivity to input order and the lack of explicit interaction modeling between key molecular substructures. It encodes reactions using symmetric-difference shingle fragments, which remove order sensitivity and precisely capture structural changes between reactants and products. To further model the mechanistic essence of reactivity, ReaDISH introduces a structural-geometric interaction attention mechanism that integrates geometric distance, structural similarity, and chemical connectivity at the substructure level.

Requirements

We implement our model on Python 3.10. These packages are mainly used:

rdkit                2024.3.3
torch                2.3.1
tensorboard          2.17.0
lightning            2.3.3
pytorch-lightning    2.3.3
unicore              0.0.1
unimol_tools         0.1.0.post1
rxnfp                0.1.0

Datasets

Pre-training dataset

We filtered reactions from USPTO (https://figshare.com/articles/dataset/Chemical_reactions_from_US_patents_1976-Sep2016_/5104873) and CJHIF (https://github.com/jshmjs45/data_for_chem). The code to cluster reactions based on DRFP fingerprints and Kmeans is stored in /data/cluster.py.

Downstream dataset

The downstream datasets (7 sources) and the generated coordinates (/ds_confs) are stored in /downstream_datasets.

Raw datasets can be downloaded from:

• BH and SM datasets: https://github.com/rxn4chemistry/rxn_yields/tree/master/data.
• NiCOlit dataset: https://github.com/truejulosdu13/NiCOlit/tree/master/data.
• ELN dataset: https://github.com/nsf-c-cas/yield-rxn/tree/master/data/az.
• N,S-acetal and C-heteroatom datasets: https://zivgitlab.uni-muenster.de/m_kueh11/fp-dm-tool.
• USPTO_TPL dataset: https://github.com/rxn4chemistry/rxnfp/tree/master/data. Please download the USPTO dataset from the link provided, as it is too large to include in the GitHub repository.

Shingle generation

The code for shingle generation is stored in /model/generate_shingling.py (for the pre-training dataset) and /model/utils.py. The pipeline is adopted from https://github.com/reymond-group/drfp.

mappings = ShinglingEncoder.encode([one_reaction_smiles], radius=3, show_progress_bar=False, atom_index_mapping=True, root_central_atom=False, symmetric_id=symmetric_id)[0]
# symmetric_id, 0: symmetric difference, 1: reactant shingles only, 2: union shingles

Experiments

Pre-training

Run pretraining.py to pre-train ReaDISH. For example,

python pretraining.py --max_epochs 3 --batch_size 8 --init_lr 5e-5 --min_lr 5e-6 --devices 0,1,2,3,4,5,6,7

The pre-trained model is stored in /checkpoint.

Fine-tuning

Run run_BH.py to fine-tune ReaDISH on the BH dataset. We also provide an example to fine-tune on another given downstream dataset. For example,

python run_BH.py

python downstream.py --devices 0,1,2,3 --batch_size 4 --accumulate_grad_batches 4 --ds_name SM --repeat_times 10 --max_epochs 150 --check_val_every_n_epoch 1 --num_workers 8 --init_lr 1e-3 --min_lr 1e-4 --warmup_lr 1e-6 --warmup_steps 500 --pred_type regression --init_checkpoint checkpoint/last.ckpt --norm

Citation

@inproceedings{shi2025reaction,
  title={Reaction Prediction via Interaction Modeling of Symmetric Difference Shingle Sets},
  author={Shi, Runhan and Chen, Letian and Yu, Gufeng and Yang, Yang},
  booktitle={Proceedings of the Thirty-ninth Annual Conference on Neural Information Processing Systems (NeurIPS)},
  year={2025}
}