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

Atom-to-atom Mapping

Atom-to-atom mapping (AAM) establishes the structural lineage that reveals microscopic changes defining each transformation. Accurate AAM is essential for identifying reaction centers, extracting mechanistic templates, and supervising models that reason about bond changes. SynRXN stratifies the reaction corpus into two distinct domains: synthetic chemical reactions and biochemical transformations for comprehensive benchmarking.

Key Datasets for AAM

Mapping Accuracy Benchmark (Table 5)

Reaction Classification

Reaction classification maps raw reaction inputs to predefined classes based on their structural or functional signatures. SynRXN provides a benchmark suite spanning multiple levels of granularity, from fine-grained SMARTS templates to high-level hierarchical ontologies. It includes datasets from USPTO patents and biochemical transformations (ECREACT).

Key Datasets for Classification

Classification Performance (Weighted F1, Table 6 Excerpt)

Reaction Property Prediction

This task targets the quantification of continuous chemical attributes, such as yields, activation barriers, and transition-state features. SynRXN aggregates data from public repositories and the literature, encompassing ab initio kinetics datasets, specific mechanistic classes, and high-throughput experimental results.

Key Datasets for Property Prediction

Property Prediction Performance (MAE, Table 7 Excerpt)

Synthesis Prediction

The synthesis prediction task consolidates essential benchmarks for algorithmic single-step reaction prediction, combining forward and retrosynthesis. SynRXN provides standardized, deterministic splits to resolve prevalent issues with benchmark comparability, focusing on conventional top-k accuracy alongside structural similarity metrics.

Key Datasets for Synthesis Prediction