SynRXN#

SynRXN: curated reaction benchmarks for reproducible reaction informatics

SynRXN organizes atom mapping, reaction classification, property prediction, reaction rebalancing, and synthesis datasets into a versioned, citable benchmark resource with a lightweight Python API.

Get started Browse datasets API reference
PyPI Scientific Data Zenodo GitHub

Why SynRXN?#

Benchmarking reaction-informatics methods is difficult when datasets, splits, reaction representations, and provenance metadata are scattered across releases or publications. SynRXN solves this by providing a consistent data layout, version-aware access, documented schema conventions, and reproducible splitting utilities.

Curated datasets

Compressed CSV records grouped by benchmark task, with stable columns and source citations.

Version-aware loading

Use archived Zenodo releases, GitHub tags, exact commits, or development snapshots.

Reproducible splits

Create repeated k-fold or train/validation/test partitions with controlled random seeds.

Accessible API

Load datasets as pandas DataFrames and integrate them directly into ML pipelines.

Framework overview#

SynRXN framework overview

Figure 1. Curated reaction datasets are grouped by benchmark task, distributed through reproducible releases, loaded through a shared API, and evaluated with task-specific workflows.#

The SynRXN pipeline separates the data lifecycle into four practical layers:

  1. Curated assets under Data/<task>/<dataset>.csv.gz.

  2. Versioned distribution through Zenodo records, GitHub releases, or exact Git commits.

  3. Reusable utilities for loading, caching, manifest handling, and splitting.

  4. Task-specific evaluation for mapping, classification, property, rebalancing, and synthesis workflows.

Benchmark collections#

RBL Reaction rebalancing

Recover chemically balanced reactions when reactants, products, solvents, catalysts, or auxiliary species are missing.

 AAM Atom-to-atom mapping

Evaluate predicted atom correspondences against curated, rule-based, or consensus reference mappings.

 CLS Reaction classification

Assign reaction classes, named-reaction labels, template identifiers, or hierarchical enzyme annotations.

 PROP Property prediction

Model kinetic, thermodynamic, and experimental reaction properties such as barriers, enthalpies, rates, yields, and free energies.

 SYN Synthesis prediction

Support forward synthesis, retrosynthesis, reagent prediction, condition recommendation, and reaction-center identification.

 MECH Mechanism prediction

TODO: add datasets for elementary steps, intermediates, and mechanistic pathways.

Quick example#

Install SynRXN, load a released classification benchmark, and inspect the first records:

pip install synrxn

from pathlib import Path
from synrxn.data import DataLoader

loader = DataLoader(
    task="classification",
    source="zenodo",
    version="1.0.0",
    cache_dir=Path("~/.cache/synrxn").expanduser(),
)

print(loader.available_names())
df = loader.load("schneider_b")
print(df.head())

Citation#

If you use SynRXN in published work, cite the primary data descriptor and the exact Zenodo version used for your data archive.

@article{phan2026synrxn,
  title = {SynRXN: An Open Benchmark and Curated Dataset for Computational Reaction Modeling},
  author = {Phan, Tieu-Long and Nguyen Song, Nhu-Ngoc and Stadler, Peter F.},
  journal = {Scientific Data},
  volume = {13},
  pages = {625},
  year = {2026},
  doi = {10.1038/s41597-026-07260-w},
  url = {https://www.nature.com/articles/s41597-026-07260-w}
}