Retrosynthesis Model Predictions and Reaction Data for Extrapolation Analysis

Two benchmark datasets derived from the Lowe USPTO dataset, containing 50,000 and 480,000 chemical reactions, support the evaluation of template-free retrosynthesis models. Jonghwi Choe published this collection in 2026 to provide raw model outputs and analysis scripts for the associated research paper. The repository includes ground-truth test sets, predictions from five baseline models, and tools for computing exact-match and round-trip accuracy metrics.

Use Cases

• Benchmark exact-match accuracy (Top-1 to Top-10) of Molecular Transformer predictions against the ground-truth test sets.
• Analyze the distribution of reaction templates in Graph Retrosynthesis model outputs to assess out-of-distribution generalization.
• Evaluate round-trip prediction accuracy by processing forward-reaction data generated by LocalTransform.
• Reproduce the paper's analysis of extrapolation capability using the provided preprocessed prediction files and Jupyter notebooks.
• Compare the performance of five baseline models (MT, MG, GR, CF, LR) on the two dataset scales (50k and 480k).

Strengths

• Includes raw predictions from five distinct baseline retrosynthesis models (MT, MEGAN, GR, Chemformer, LocalRetro).
• Provides two dataset scales (50k and 480k reactions) derived from the Lowe USPTO dataset for comparative analysis.
• Contains complete analysis scripts to reproduce the paper's three key evaluation metrics.

Limitations

• The total number of individual data rows (reactions) within the provided files is not explicitly stated.
• Dataset scope is limited to chemical reactions from the USPTO patent corpus, which may introduce domain-specific biases.
• Requires specific Python dependencies (e.g., rdkit, localmapper) and version constraints (NumPy 1.x) for full reproducibility.

Provenance

Source

Derived from the Lowe USPTO reaction dataset.

Collection Method

Processed and split for benchmarking template-free retrosynthesis models, with predictions generated by published baseline models.

Freshness

Last updated in April 2026.