Intelligent-molecular-systems/completeRXN-benchmark

CompleteRXN Benchmark

206,423 reactions for reaction completion: given a USPTO reaction that is missing molecules, so its atoms don't balance, predict the missing molecules so that the completed reaction is balanced for atoms and charge. The references come from FlowER mechanistic reactions matched to their USPTO records.

Three splits test how well models generalize to increasingly unfamiliar reactions: random, group out-of-distribution and extreme out-of-distribution. Each split has 5 repetitions.

• Code, models and paper analyses: CompleteRxn-Benchmarking
• Paper: CompleteRXN: Toward Completing Open Chemical Reaction Databases
• Fine-tuned CRB model: completeRXN-CRB
• Saved predictions and environments for reproducing the paper: completeRXN-reproduction

Quick start

The completerxn package from the code repository downloads and checks these files for you:

from completerxn import DataLoader

train, val, test = DataLoader().load(split="random", rep=0, population="filtered")
predictions = test.predict(my_model)   # my_model.predict(inputs) returns completed reactions
print(test.evaluate(predictions).summary)

Leave out population="filtered" to evaluate on the full test set. See the code repository's README for the prediction format, command-line tools and examples.

Files

benchmark_final.csv     206,423 reactions with metadata; Unique_identifier links all files
random_splits/          5 files, split column random_split
group_splits/           5 files, split column group_split
extreme_ood_splits/     5 files, split column extreme_ood
eligibility/            IDs of the filtered test set, one file per split and repetition
raw/                    source files for rebuilding the dataset
  master_dataset_mapped.csv
  flower_full_single_reactions_with_main_products.tsv

In every split file, the split column is 0 for train, 1 for validation and 2 for test.

The paper evaluates on the filtered test set: test reactions whose input or reference contains tokens outside the pretrained Molecular Transformer's vocabulary are excluded, so that all models are compared on the same reactions. eligibility/<split>_rep<rep>.csv lists the reactions kept, and eligibility/coverage.csv gives the number kept per split and repetition.

Main columns in benchmark_final.csv

Column	Type	Description
Unique_identifier	string	Reaction ID (e.g. USPTO_US03930836_0), used to join the split and eligibility files
InputReaction	string	Incomplete reaction given to a model, as reactants>reagents>products SMILES
InputReaction_harmonized	string	Default model input, used by all models in the paper: InputReaction with extra copies of a molecule removed where the input contains more copies than the reference (at least one copy is kept)
TargetReaction	string	Reference: complete reaction from FlowER, balanced for atoms and charge
TargetReaction_with_stereo	string	Default reference: TargetReaction with stereochemistry restored from the USPTO record, since FlowER does not include it. Reactions whose stereochemistry could not be restored unambiguously are excluded from the dataset, so this column is always filled
input_target_diff_harmonized	string	The molecules missing from InputReaction_harmonized, i.e. what a model has to add
MissingNrAtoms	int	Number of heavy atoms missing from the product side (the main difficulty measure in the paper)
MissingNrCarbons	int	Number of carbon atoms missing from the product side
MissingMw	float	Molecular weight of the missing molecules
Balance	string	Type and count of the missing atoms
rxn_class	string	Reaction class (1–10, Schneider et al.), predicted by a classifier trained on the two USPTO 50k datasets
1kTPLLabel	string	Reaction template label from USPTO 1k TPL, which groups reactions by the 1,000 most common templates. Empty if the reaction is not in that set
OriginalReaction	string	Original USPTO reaction SMILES before curation (Lowe, 2017)
CanonicalOriginalReaction	string	OriginalReaction in canonical SMILES, without atom mapping
PatentNumber, ParagraphNum, Year	string, string, int	Source patent, paragraph and year
FlowER_ID	string	ID of the matched FlowER reaction

Other columns

• has_unresolved_stereo, n_unresolved_stereo: whether, and for how many molecules, stereochemistry could not be restored. Kept from the processing pipeline. In this release they are always False and 0, because the 147 affected reactions were excluded. A further 47 reactions were excluded because their input, including reagents and stereochemistry, is not fully contained in the matched FlowER reaction.
• input_harmonized_modified: whether InputReaction_harmonized differs from InputReaction.
• source_not_in_target: whether the input contains molecules that do not occur in the reference.
• rxn_flower, rxn_reagents_identified_flower, moved_metadata: the matched FlowER reaction, the same reaction with unchanged molecules moved to the reagent field, and a record of which molecules were moved.
• InputReaction_canonical, TargetReaction_canonical: canonical SMILES versions of the input and reference.
• TextMinedYield, CalculatedYield, ReactionSmiles: fields from the original USPTO extraction.
• RewrittenReaction, MovedMetadata, MultiProduct, CanonicalReaction, IsBalanced, InvalidReason, IsIn, FlowER_ID_parsed: intermediate fields from curation and from matching USPTO to FlowER, kept for traceability.

Task and metric

Given InputReaction_harmonized, predict the complete reaction, including the molecules that were missing.

The metric is equivalence accuracy: the percentage of test reactions whose prediction is an equivalence match with the reference. A prediction is an equivalence match if it is valid, balanced for atoms and charge, and represents the same reaction as the reference once equivalent ways of writing the molecules are normalized. Stereochemistry is ignored by default. The rules are implemented in src/completerxn/equivalence.py in the code repository.

Rebuilding the dataset

Download the two files in raw/, place them in data/Template_curated/USPTO_FlowER_raw/ in a checkout of the code repository, and run bash data/Template_curated/run_data_processing.sh. See data/Template_curated/README.md there for the individual steps.

Source data

master_dataset_mapped.csv matches the USPTO extraction by Lowe (2017) to USPTO STEREO, USPTO 1k TPL and FlowER by canonical SMILES (column IsIn). The two USPTO 50k datasets are not part of the released files: they were used only to train the classifier behind rxn_class and for the analysis of incompleteness per reaction class in the paper.

Dataset	Reference	Download	License
USPTO (grants, 1976–2016)	Lowe (2017)	figshare	CC0
USPTO 50k (2015)	Schneider et al., JCIM 2015	rxn4chemistry/rxnfp	MIT (rxnfp repository)
USPTO 50k (2016)	Schneider et al., JCIM 2016	ACS supporting information	CC BY-NC 4.0 (ACS figshare)
USPTO STEREO	Schwaller et al., Chem. Sci. 2018	IBM Box	No license stated (subset of Lowe, 2017)
USPTO 1k TPL	Schwaller et al., Nat. Mach. Intell. 2021	IBM Box	No license stated (subset of Lowe, 2017)

These datasets are derived from USPTO patent data.

flower_full_single_reactions_with_main_products.tsv is a preprocessed extract of the FlowER single-step mechanistic reactions: Joung et al., Nature 645, 115–123 (2025), figshare (MIT).

License

This dataset is released under the MIT License, like the code repository. Keep the copyright and license notice when you redistribute it.

It is derived from the sources above, which keep their own terms. The references (TargetReaction, TargetReaction_with_stereo, rxn_flower and derived columns) are derived from FlowER, released under the MIT License by its authors; keep their notice as well. The USPTO extraction by Lowe (2017), the source of all reaction SMILES, is CC0. USPTO STEREO and USPTO 1k TPL state no license; from them, this dataset uses only membership (IsIn) and the template labels in 1kTPLLabel. No records of the two USPTO 50k datasets are included; rxn_class holds predictions of our own classifier trained on them.

Citation

If you use this dataset, please cite:

@article{vogel2026completerxn,
  title={CompleteRXN: Toward Completing Open Chemical Reaction Databases},
  author={Vogel, Gabriel and Noordsij, Minouk and Pidko, Evgeny and Weber, Jana M.},
  journal={Advances in Neural Information Processing Systems},
  volume={39},
  year={2026}
}

Please also cite FlowER (Joung et al., 2025) and the USPTO sources above that are relevant to your use.