neuripssubmission737/uspto_advanced_v1

USPTO-Advanced

A property-stratified single-step retrosynthesis benchmark derived from the public USPTO reaction collection. Each test row is selected to broaden coverage of the long-tailed reaction-property axes (rare reaction classes, large reaction centers, spiro / bridged scaffolds, multi-component reactions, partially-specified stereo, novel Bemis–Murcko scaffolds) while holding the test split product-disjoint from the dominant retrosynthesis training corpora (USPTO-50k, USPTO-Full / GLN-Full, USPTO-MIT, USPTO-STEREO).

The release ships three test tiers — test_2k_hard (challenge subset), test_5k (matched-distribution standard test), and test_10k (comprehensive test) — together with a validation split and a product-disjoint train split.

Splits

Split	Rows	Distinct products	Construction
train	269,746	~260k	Pool minus all val/test sets, product-disjoint.
validation	4,999	4,994+	Stratified the same way as test_5k; disjoint from all test sets.
test_2k_hard	2,000	1,996	Hard challenge subset; concentrated on rare strata. Disjoint from test_10k.
test_5k	5,000	4,987	Standard stratified test (USPTO-50k-test scale, comparable for headline numbers). Subset of test_10k.
test_10k	10,000	9,969	Comprehensive stratified test for tighter confidence intervals on long-tail metrics.

The three tiers parallel PaRoutes' n1 / n5 difficulty axis but apply to single-step retrosynthesis: the hard subset is concentrated on rare patterns (rare reaction classes, RC ≥ 7 atoms, spiro / bridged scaffolds, ≥ 4 reactants, partially-specified stereo, singleton scaffolds), while the 5k and 10k tiers carry a 30 % rare-stratum oversample on top of a 70 % natural-frequency sample.

Stratification axes

The candidate pool was annotated along seven retrosynthesis property axes extracted with rxn-insight and RDKit:

Axis	Stratification target
reaction class (rxn_insight_class)	All 11 valid rxn-insight classes represented; bottom-3 (Protection, Oxidation, Functional Group Addition) oversampled. rxn_insight_failed rows dropped.
reaction center (rc_atom_count, rc_in_aromatic_ring)	Bins {1, 2, 3, 4–6, 7+}; rc ≥ 7 oversampled.
scaffold topology (spiro_flag, bridged_flag, fused_flag)	spiro = True and bridged = True (~1.6 % each in pool) oversampled.
scaffold size (ring_count, ring_system_count)	Bins {0, 1, 2, 3, 4, ≥ 5}; ring_count = 0 and ring_count ≥ 5 oversampled.
scaffold novelty (bm_scaffold frequency)	Singleton Bemis–Murcko scaffolds (seen once in the candidate pool) oversampled.
component count (n_precursors)	Bins {1, 2, 3, ≥ 4}; both n = 1 and n ≥ 4 oversampled. ≥ 5-precursor rows kept but not oversampled (often retain coupling reagents).
stereochemistry (stereo_spec_level)	Four buckets {not_applicable, unspecified, fully_specified, partially_specified}; the rare partially_specified bucket oversampled.

Sampling preserves ≥ 70 % of every rare stratum's mass in train for six of seven strata (rc_atom_count ≥ 7, spiro, bridged, ring_count = 0, partially_specified stereo, rare reaction classes). The seventh, n_precursors ≥ 4, sits at ~64 % because multi-component reactions in USPTO-Full converge on shared products across routes, so the product-disjoint pruning between train and val/test pulls more mass from this stratum than the others. Train still carries ~3 k multi-component examples, so the model has plenty of learning signal.

Loading

from datasets import load_dataset

ds = load_dataset("neuripssubmission737/uspto_advanced_v1")
# Available splits: train, validation, test_2k_hard, test_5k, test_10k

print(ds)
print(ds["test_5k"][0]["reaction_smiles"])
# 'CC(C)(Oc1cccc2c1sc1ccccc12)C(=O)O.CCO>>CCOC(=O)C(C)(C)Oc1cccc2c1sc1ccccc12'

A minimal retrosynthesis evaluation loop:

ds = load_dataset("neuripssubmission737/uspto_advanced_v1", split="test_5k")
for row in ds:
    target_product = row["product_canonical"]
    gold_reactants = row["reactants_canonical"]
    # ... predict precursors from target_product, compare to gold_reactants ...

Schema

Each row carries the reaction strings, identifiers, and the seven feature axes used for stratification.

Identifiers

Field	Type	Description
reaction_id	string	Stable id, format usptofull_raw:all:<patent>:<paragraph>#<dup_idx>.
patent_number	string	US patent identifier (e.g. US03953601).
year	int	Patent year (1976–2016).
source	string	grants or applications.

Reaction strings

Field	Type	Description
reaction_smiles	string	Schwaller-style reactants>>products, atom-mapping stripped, canonical.
reactants_canonical	string	Reactant block, atom-mapping stripped, canonical.
product_canonical	string	Product, atom-mapping stripped, canonical.
mapped_rxn_smi	string	The atom-mapped reaction SMILES (kept verbatim from the source pipeline).

Feature axes

reaction_type: rxn_insight_class (str), rxn_insight_name (str).

component_count: n_precursors (int), n_products (int), n_spectator_reactants (int), atom_mapping_source (str).

reaction_center: rc_atom_count (int), rc_in_aromatic_ring (bool), n_bond_changes (int), bond_change_type (str), bond_change_heavy_pair (list), rc_env_smarts (str), rc_env_morgan_r1 (str), rc_env_morgan_r2 (str).

scaffold (product side, then reactant side mirrors with _reactants suffix): bm_scaffold (str), generic_scaffold (str), ring_count (int), ring_system_count (int), largest_ring_size (int), spiro_flag / bridged_flag / fused_flag / aromaticity_flag (bool).

stereochemistry: stereo_spec_level (str), stereo_preservation (str), stereo_present / stereo_present_product / stereo_present_reactants (bool), stereo_count_product (int), stereo_potential_count_product (int), n_potential_stereo_bonds (int).

functional_group: fg_count_product / fg_count_reactants / fg_preserved_count (int), fg_product_set / fg_product_multiset / fg_reactants_set / fg_reactants_multiset / fg_changed_in_product / fg_changed_in_reactants / fg_preserved / fg_at_rc (list, IFG vocabulary).

shard: construction-time provenance string (the upstream filter shard the row came from); kept for reproducibility but not a feature.

Construction

1. Source. Daniel M. Lowe's USPTO reaction deposit on Figshare (DOI 10.6084/m9.figshare.5104873), covering US patent grants and applications 1976 – Sep 2016 (~3.75 M raw reactions).
2. Filtering. Quality-filtered down to ~1.29 M reactions using a modified RetroFilter pipeline (atom-mapping resolution, valence checks, spectator removal, tier-based audit).
3. Exclusion set. Reaction-level dedup-canonicalised join (strip atom mapping + stereo, canonical SMILES) against the union of the four dominant retrosynthesis training splits — USPTO-50k, USPTO-Full / GLN-Full, USPTO-MIT (Jin et al. 2017), USPTO-STEREO (Schwaller et al.
   2018. — yielding 999 k training reactions / 922 k distinct products.
4. Candidate pool. Anti-join the filtered USPTO-Full against the exclusion set's product set → 291,036 candidate reactions disjoint at the product level from all four training corpora.
5. Stratified sampling. 30 % rare-bin quotas + 70 % natural-frequency draw; product-disjoint pruning between train and val/test.

The full construction code will be released alongside the paper after the double-blind review period.

Caveats

• Atom mapping is upstream-derived. The mapped_rxn_smi column carries Lowe's mappings as resolved by the RetroFilter pipeline. It is not re-mapped on release; downstream consumers needing different mapping conventions should re-run their preferred mapper on reaction_smiles.
• 0.013 % canonicalisation failures dropped. 37 / 291,782 reactions failed canonical SMILES generation (exotic valences such as [PH2] or [CH2+]) and were excluded from the release: 36 from train, 1 from validation, none from any test split. Downstream consumers needing the dropped rows can recover them from the upstream filter outputs and parse mapped_rxn_smi with a relaxed parser.
• rxn_insight_failed rows excluded. 2,847 reactions for which rxn-insight failed to assign a reaction class were dropped from the candidate pool before sampling — they cannot be stratified.
• Multi-component reactions ≥ 5 precursors. Genuine Ugi / Hantzsch / Passerini reactions appear in the long tail, but ≥ 5-precursor rows are also enriched in coupling-reagent contamination (EDC, HOBt, explicit water, counterions) that survived spectator removal. We keep these rows in the pool but do not oversample them.

Citation

If you use USPTO-Advanced, please cite both the source data and our construction paper:

@misc{lowe2017uspto,
  author       = {Lowe, Daniel Mark},
  title        = {Chemical reactions from US patents (1976--Sep 2016)},
  year         = 2017,
  doi          = {10.6084/m9.figshare.5104873},
  url          = {https://doi.org/10.6084/m9.figshare.5104873}
}

@inproceedings{anonymous2026usptoadvanced,
  title     = {USPTO-Advanced: a property-stratified single-step retrosynthesis benchmark},
  author    = {Anonymous},
  year      = {2026},
  booktitle = {Submitted to NeurIPS Datasets and Benchmarks Track},
  note      = {Author list withheld for double-blind review}
}

Double-blind notice. This dataset accompanies a paper currently under double-blind review. The construction paper's author list and code-repository URL will be added after the review period.

License

CC-BY-4.0 — same as Lowe's USPTO Figshare deposit. Attribution to both Lowe (2017) and the construction paper is required.

Acknowledgements

• Source data: Daniel M. Lowe (USPTO Figshare deposit, 2017).
• Reaction-class and functional-group annotation: rxn-insight (Schwaller lab).
• Filtering pipeline: derived from RetroFilter (Coley lab).
• USPTO-50k / GLN-Full / USPTO-MIT / USPTO-STEREO reference splits used for the exclusion set: respective dataset authors (Jin et al. 2017; Schwaller et al. 2018; Dai et al. 2019).