SynRXN: An Open Benchmark and Curated Dataset for Computational Reaction Modeling.

Area of Science:

• Chemistry
• Computer Science
• Data Science

Background:

• Computer-aided synthesis planning (CASP) lacks standardized benchmarks, hindering reproducible research and method comparison.
• Existing datasets are often heterogeneous and lack clear provenance, leading to inconsistencies in evaluation.

Purpose of the Study:

• To introduce SynRXN, a unified benchmark dataset resource for CASP.
• To provide standardized datasets and evaluation metrics for five key CASP task families.
• To enable fair, reproducible, and rigorous benchmarking of CASP methods.

Main Methods:

• Assembled curated reaction corpora from diverse public sources into a harmonized representation.
• Packaged versioned datasets for reaction rebalancing, atom-to-atom mapping, reaction classification, reaction property prediction, and synthesis prediction.
• Implemented leakage-aware partitions, standardized metrics, and reproducible build recipes.

Main Results:

• SynRXN provides versioned datasets with explicit metadata, licenses, and manifests.
• Predefined, leakage-aware partitions and tailored evaluation metrics are available for each task.
• Held-out test sets and restricted distribution of sensitive tasks ensure robust benchmarking.

Conclusions:

• SynRXN enables fair longitudinal comparison and rigorous stress testing of CASP methods.
• The resource lowers the barrier for practitioners seeking reliable performance estimates for synthesis planning.
• Open licenses promote reuse and extension, advancing the field of reaction informatics.