Retrosynthesis Model Predictions and Analysis for Extrapolation Assessment

Two datasets derived from the Lowe USPTO collection, containing 50k and 480k chemical reactions, support the evaluation of template-free retrosynthesis models. Jonghwi Choe created this repository to provide raw data, model predictions, and analysis scripts for reproducing results from the associated 2026 paper. It includes predictions from five baseline models and enables calculation of exact-match accuracy, template distribution, and round-trip accuracy metrics.

Use Cases

• Benchmark exact-match accuracy (Top-1 to Top-10) of Molecular Transformer, MEGAN, Graph Retrosynthesis, Chemformer, and LocalRetro models on provided test sets.
• Analyze template distribution of model predictions to categorize out-of-distribution (OOD) reaction templates.
• Evaluate round-trip prediction accuracy by processing forward-reaction data generated by LocalTransform.
• Reproduce the train/valid/test splits using the provided datasplit scripts on the raw reaction data with and without atom mapping.

Strengths

• Provides two distinct dataset scales (50k and 480k reactions) for analysis.
• Includes raw prediction outputs from five established baseline models for direct comparison.
• Contains complete preprocessing and analysis scripts (Jupyter notebooks) to fully reproduce the paper's results.

Limitations

• The dataset size (1.5 GB) and dependency on specific Python packages (rdkit, localmapper) may pose computational and setup barriers.
• The core reaction data is derived from the Lowe USPTO dataset, which may inherit any biases or limitations of that source.

Provenance

Source

Derived from the Lowe USPTO dataset.

Collection Method

Processed and augmented with model predictions and analysis for the specific research paper.

Freshness

Last updated in April 2026.