Data Records

This page summarizes the curated SynRXN records by benchmark family. It is designed as a quick inventory for selecting datasets, checking target columns, and finding the source citation and reuse license for each benchmark table. Dataset-level source and literature citations are collected on the Reference page.

Note

The License column reports the license of the table as distributed in the SynRXN release. Upstream publications and external source repositories should still be cited through the Source column. Users who redistribute modified versions should also check the upstream source terms.

Task families

Reaction rebalancing, atom mapping, classification, property prediction, and synthesis.

Curated tables

Named benchmark records distributed as compressed CSV files.

Citable release

Use the Scientific Data paper and archived data release for reproducible reporting.

Reaction Rebalancing

Data folder: rbl

Reaction rebalancing datasets evaluate whether a method can restore missing species and recover a balanced reaction from an incomplete equation. The MNC, MOS, and MBS subsets are derived from USPTO-50K-style reaction records, whereas the Complex subset is based on manually validated complex-reaction mapping collections. See Reference for complete bibliographic records.

Dataset	Records	Benchmark role	Principal columns	Source	License
complex	1,748	Complex transformations and skeletal rearrangements.	r_id, input rxn, balanced ground_truth	[1, 2]	CC BY 4.0
mbs	491	Missing species on both sides of the reaction.	r_id, input rxn, balanced ground_truth	[3]	CC BY 4.0
mnc	33,147	Missing non-carbon species.	r_id, input rxn, balanced ground_truth	[3]	CC BY 4.0
mos	12,781	Missing species on one side of the reaction.	r_id, input rxn, balanced ground_truth	[3]	CC BY 4.0

Atom-to-Atom Mapping

Data folder: aam

Atom-to-atom mapping records compare predicted mappings with curated or consensus reference mappings across synthetic and biochemical reaction domains. See Reference for source publications and mapper references.

Dataset	Records	Benchmark role	Principal columns	Source	License
ecoli	273	Biochemical mapping benchmark from E. coli reactions.	r_id, ground_truth, mapper outputs, rxn, original_id	[4]	CC BY 4.0
enzyme_map	47,974	Enzymatic atom-mapping benchmark from the EnzymeMap reaction collection.	r_id, ground_truth, rxn, original_id	[5]	CC BY 4.0
golden	1,758	Curated synthetic chemistry mapping benchmark.	r_id, ground_truth, mapper outputs, rxn, original_id	[2, 6]	CC BY 4.0
natcomm	491	Reference mapping set from literature-derived reactions.	r_id, ground_truth, mapper outputs, rxn, original_id	[1]	CC BY 4.0
recon3d	382	Biochemical mapping benchmark from Recon3D reactions.	r_id, ground_truth, mapper outputs, rxn, original_id	[7]	CC BY 4.0
uspto_3k	3,000	USPTO-derived synthetic chemistry mapping benchmark.	r_id, ground_truth, mapper outputs, rxn, original_id	[6]	CC BY 4.0

Reaction Classification

Data folder: classification

Classification datasets evaluate reaction class, template, or enzyme-label prediction. Some corpora are provided in balanced/full and unbalanced/product focused variants. See Reference for label-source and benchmark citations.

Dataset	Records	Benchmark role	Principal columns	Source	License
ecreact	185,734	Enzyme Commission hierarchy classification.	r_id, rxn, ec1, ec2, ec3, split, orig_index	[8]	CC BY 4.0
schneider_b	50,000	Balanced Schneider reaction-class benchmark.	r_id, rxn, label, split	[9, 10]	CC BY 4.0
schneider_u	50,000	Unbalanced/product-focused Schneider variant.	r_id, rxn, label, split	[9]	CC BY 4.0
syntemp	43,441	Hierarchical SynTemp template labels.	orig_index, r_id, rxn, label_0, label_1, label_2	[3, 11]	CC BY 4.0
tpl_b	445,115	Balanced template-label benchmark.	r_id, rxn, label, split	[10, 12]	CC BY 4.0
tpl_u	445,115	Unbalanced/product-focused template-label benchmark.	r_id, rxn, label, split	[12]	CC BY 4.0
uspto_50k_b	50,016	Balanced USPTO-50K class-label benchmark.	r_id, rxn, label, split	[3, 10]	CC BY 4.0
uspto_50k_u	50,016	Unbalanced/product-focused USPTO-50K variant.	r_id, rxn, label, split	[3]	CC BY 4.0

Reaction Property Prediction

Data folder: property

Property datasets evaluate continuous or probabilistic reaction attributes such as activation energy, enthalpy, rate, conversion, or reaction free energies. See Reference for upstream corpus citations.

Dataset	Records	Benchmark role	Principal columns	Source	License
b97xd3	16,365	Quantum-chemistry property prediction.	r_id, aam, ea, dh	[13, 14]	CC BY 4.0
cycloadd	5,269	Cycloaddition free-energy prediction.	r_id, aam, G_act, G_r, split	[15]	CC BY 4.0
e2	1,264	E2 activation-energy prediction.	r_id, aam, ea, split	[15]	CC BY 4.0
e2sn2	3,625	E2/SN2 activation-energy prediction.	r_id, aam, ea	[16, 17]	CC BY 4.0
lograte	778	Log-rate prediction.	r_id, aam, lograte	[16, 18]	CC BY 4.0
phosphatase	33,354	Phosphatase conversion prediction.	r_id, aam, Conversion, onehot	[16, 19]	CC BY 4.0
rad6re	31,923	Reaction enthalpy prediction.	r_id, aam, dh	[16, 20]	CC BY 4.0
rdb7	23,852	Activation-energy prediction with predefined splits.	r_id, aam, ea, split	[15]	CC BY 4.0
rgd1	353,984	Large-scale activation-energy prediction.	r_id, aam, ea, split	[15]	CC BY 4.0
sn2	2,361	SN2 activation-energy prediction.	r_id, aam, ea, split	[15]	CC BY 4.0
snar	503	SNAr activation-energy prediction.	r_id, rxn, ea	[21]	CC BY 3.0

Synthesis Prediction

Data folder: synthesis

Synthesis records support single-step forward synthesis, retrosynthesis, reagent/catalyst inference, and reaction-center related workflows. The Scientific Data synthesis table explicitly lists USPTO-50K, USPTO-MIT, and USPTO-500 as the SynRXN synthesis-prediction corpora; the Diels–Alder record is documented here as an additional class-specific synthesis dataset. See Reference for USPTO-derived benchmark citations.

Dataset	Records	Benchmark role	Principal columns	Source	License
da	11,011	Diels–Alder reaction benchmark.	r_id, code, reaction_original, reaction, rsmi	[22]	CC BY 3.0
uspto_500	143,535	Reagent and catalyst inference benchmark.	r_id, rxn, reagent, split	[23]	CC BY 4.0
uspto_50k	50,016	USPTO-50K synthesis/retrosynthesis benchmark.	r_id, aam, split, source	[3, 6]	CC BY 4.0
uspto_mit	479,035	Large-scale USPTO-MIT synthesis prediction benchmark.	r_id, aam, split, rc	[24]	CC BY 4.0

Choosing a dataset

Use the task family first, then choose the dataset based on target type, scale, and whether the benchmark already includes a published split. For example, schneider_b is a compact classification benchmark with a standard reaction class label, b97xd3 is a quantum-chemistry property dataset with activation energy and enthalpy targets, and uspto_50k is a common synthesis or retrosynthesis benchmark.

Practical selection guide

Goal	Start with	Check before training
Compare atom mappers	aam/golden, aam/uspto_3k, or biochemical AAM records.	Whether ground_truth and mapper-output columns match your metric.
Train reaction classifiers	classification/schneider_b or classification/uspto_50k_b.	Label granularity and whether balanced or unbalanced variants are needed.
Predict reaction energies	property/b97xd3, property/rgd1, or mechanism-specific subsets.	Unit conventions, target column name, and split availability.
Recover missing species	rbl/mnc, rbl/mos, rbl/mbs, or rbl/complex.	Whether the missing-species setting matches your method.
Benchmark synthesis tasks	synthesis/uspto_50k, synthesis/uspto_mit, or synthesis/uspto_500.	Whether the task is product prediction, retrosynthesis, reagent inference, or reaction-center prediction.

For a complete reproducibility record, pair the selected dataset with the source mode and version described in Data Concept.