insilicomedicine/URSA-benchmarking-sets

URSA benchmarking sets

Benchmark targets for single-step retrosynthesis from Zagribelnyy et al. (2026). Each file is a CSV of product molecules to predict precursors for.

File	Targets	Description
URSA-expert-2026.csv	100	Expert benchmark for challenging medicinal-chemistry targets
USPTO-50K-test.csv	4,972	Curated USPTO-50K test split
USPTO-50K-test-mini.csv	497	Fixed random 10% subsample of USPTO-50K-test

CSV format: Structure ID, SMILES — product SMILES (stereochemistry preserved where applicable).

---

License and data provenance

This release is distributed under the Apache License 2.0. The USPTO-50K-test subset is based on the publicly available USPTO-50K test split from the seq2seq repository, including the original test_sources and test_targets files, which are released under the same license. We redistribute only the curated product targets, while the original ground-truth reactants remain available from the upstream source. The curation procedure and removed entries are documented below for reproducibility.

URSA-expert-2026

Expert-curated benchmark of 100 product structures selected to stress-test retrosynthesis models on medicinally relevant scaffolds: higher complexity, diverse chemotypes, and stereochemistry typical of drug-like molecules. SMILES are in URSA-expert-2026.csv; 2D depictions are in the paper (Table 9, arXiv:2602.03554).

Identifiers use internal structure codes (e.g. X404-223-1421).

Dataset diversity

Reported dataset diversity: 0.85886.

Diversity is defined as the average pairwise dissimilarity over all distinct molecule pairs.

Usage

from datasets import load_dataset

ds = load_dataset(
    "insilicomedicine/URSA-benchmarking-sets",
    data_files="URSA-expert-2026.csv",
)

---

USPTO-50K-test

Conventional holdout benchmark for single-step retrosynthesis, derived from the USPTO-50K test split introduced by Liu et al. (2017), with a conservative curation pass by InSilico Medicine to remove entries that undermine benchmark semantics.

Original source

The canonical test products and reference reactants come from the seq2seq preprocessing pipeline:

• Products (sources): test_sources
• Reactants (targets): test_targets

The public release here contains product SMILES only (USPTO-50K-test.csv). Ground-truth reactants from the original split are not redistributed in this file; use the links above or the original USPTO-50K publication if you need reference answers for method development.

Size

Split	Pairs
Original USPTO-50K test	5,004
This release (USPTO-50K-test.csv)	4,972
Removed	32

Stable identifiers: USPTO-50K-test-0 … USPTO-50K-test-4971 (row order after curation; not the same as 0-based indices in the original 5,004-entry file).

Curation rationale

We removed 32 entries (0-based indices in the original 5,004-pair test file) that are unambiguous dataset artifacts or ill-posed as retrosynthesis targets:

1. Role-assignment errors — reagents or solvents incorrectly labeled as the product (e.g. trifluoroacetic acid, maleic/fumaric acid).
2. Very small products — molecular weight of the product < 100 Da (rarely meaningful synthetic targets in practice).
3. Other processing artifacts — obvious corruption from the original reaction-processing pipeline.

This reduces evaluation noise and keeps the benchmark focused on chemically meaningful target molecules.

Removed entries

Indices refer to the original Liu et al. test ordering (5,004 pairs).

Reason	Count	Removed indices
Trifluoroacetic acid (TFA) as product (reaction-processing artifact)	20	144, 161, 200, 204, 295, 350, 1644, 1647, 1666, 2664, 3260, 3281, 3321, 3753, 3990, 4086, 4482, 4605, 4985, 4992
Maleic or fumaric acid as product (reaction-processing artifact)	4	727, 2250, 2566, 4380
Other artifacts of reaction processing	5	433, 497, 2390, 3084, 4453
Product MW < 100 Da	3	645, 1526, 3929
Total	32

Usage

from datasets import load_dataset

ds = load_dataset(
    "insilicomedicine/URSA-benchmarking-sets",
    data_files="USPTO-50K-test.csv",
)

Reference (USPTO-50K original split)

If you compare to the canonical USPTO-50K test split, also cite Liu et al. and note our curation (32 removed entries):

@article{liu2017retrosynthetic,
  author  = {Liu, Bowen and Ramsundar, Bharath and Kawthekar, Prasad and Shi, Jade and Gomes, Joseph and Luu Nguyen, Quang and Ho, Stephen and Sloane, Jack and Wender, Paul and Pande, Vijay},
  title   = {Retrosynthetic Reaction Prediction Using Neural Sequence-to-Sequence Models},
  journal = {ACS Central Science},
  volume  = {3},
  number  = {10},
  pages   = {1103--1113},
  year    = {2017},
  doi     = {10.1021/acscentsci.7b00303},
  note    = {PMID: 29104927},
  url     = {https://doi.org/10.1021/acscentsci.7b00303},
}

---

USPTO-50K-test-mini

A smaller, fixed benchmarking set for fast iteration and low-cost evaluation (~10% of the full set).

Construction

497 product targets drawn uniformly at random, without replacement from the 4,972 curated targets in USPTO-50K-test.

Identifiers: USPTO-50K-test-mini-0 … USPTO-50K-test-mini-496.

Dataset diversity

Reported dataset diversity: 0.85793, computed with the same definition as URSA-expert-2026.

Usage example

from datasets import load_dataset

ds = load_dataset(
    "insilicomedicine/URSA-benchmarking-sets",
    data_files="USPTO-50K-test.csv",
)

Use this set when full USPTO-50K-test evaluation is too expensive; report metrics on 497 targets and note the mini split when comparing to other work.

---

Citation

If you use these benchmarking sets, please cite:

@misc{zagribelnyy2026singleanswerenoughrethinking,
  title={When Single Answer Is Not Enough: Rethinking Single-Step Retrosynthesis Benchmarks for LLMs},
  author={Bogdan Zagribelnyy and Ivan Ilin and Maksim Kuznetsov and Nikita Bondarev and Roman Schutski
          and Thomas MacDougall and Rim Shayakhmetov and Zulfat Miftakhutdinov
          and Mikolaj Mizera and Vladimir Aladinskiy and Alex Aliper and Alex Zhavoronkov},
  year={2026},
  eprint={2602.03554},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2602.03554},
}