USPTO Dataset

The dataset contains millions of chemical reactions, making it one of the largest publicly available resources for reaction informatics. Its scale is critical for training deep learning models that require vast amounts of data to generalize effectively.

• Total Reactions: Over 3 million unique reactions in recent filtered versions.
• Source: Extracted from U.S. patent grants and applications from 1976 onwards.
• Accessibility: Freely available, fostering reproducible research and lowering the barrier to entry for computational chemistry.

Reactions are canonically represented as Reaction SMILES strings, a line notation that encodes reactants, agents, and products separated by the '>' symbol. This format is directly consumable by sequence-based models like the Molecular Transformer.

• Format: reactant1.reactant2>agent1>product1
• Atom Mapping: High-quality versions include atom-to-atom mapping, which is essential for identifying the reaction center and training template-extraction algorithms.
• Utility: Enables direct translation tasks where the model learns to map a product SMILES string back to its precursor reactants.

Standardized data splits are crucial for fair model comparison. The USPTO dataset is commonly divided into subsets to benchmark different generalization capabilities.

• USPTO-50K: A curated subset of 50,000 reactions across 10 reaction classes, widely used for single-step retrosynthesis benchmarking.
• USPTO-MIT: A popular split using 80% for training, 10% for validation, and 10% for testing, often filtered to remove duplicate and erroneous reactions.
• USPTO-Full: The complete dataset used for pre-training large models, testing scalability, and evaluating performance on rare reaction types.

Raw patent data is noisy and requires rigorous preprocessing to be useful for machine learning. Key cleaning steps directly impact model performance.

• Atom Mapping: Algorithms like RXNMapper or Indigo are used to establish atom correspondence, a prerequisite for extracting reaction templates.
• Reagent Removal: Reagents and solvents are often stripped to focus the model on the core structural transformation between reactants and products.
• Canonicalization: SMILES strings are canonicalized to ensure a one-to-one representation of molecular structures, preventing data leakage from different tautomeric or stereochemical forms.

Reactions in the USPTO dataset can be categorized into specific types based on the structural transformation at the reaction center. This enables the development of specialized, class-conditioned models.

• Top Classes: Include heteroatom alkylation, acylation, C-C bond formation, deprotection, and functional group interconversion.
• Class Tokenization: A special token representing the reaction class is often prepended to the input sequence, conditioning the model to generate context-specific predictions.
• Imbalanced Distribution: The dataset has a highly skewed class distribution, requiring techniques like oversampling or weighted loss functions to prevent model bias towards common reactions.

Despite its scale, the USPTO dataset has known biases that affect the real-world applicability of trained models.

• Patent Bias: Reactions are reported to demonstrate novelty, leading to an overrepresentation of complex, late-stage functionalizations and an underrepresentation of simple, robust transformations.
• Positive Result Bias: Failed reactions are rarely reported, creating a skewed view of chemical space that lacks negative examples for training classifiers.
• Yield Reporting: Reaction yields are inconsistently recorded and often omitted, making it difficult to train accurate yield prediction models directly from this source.

The process of establishing a one-to-one correspondence between atoms in reactants and products. The USPTO dataset's value is significantly enhanced by high-quality atom mapping, which explicitly labels the structural transformation at the reaction center.

• Essential for extracting reaction templates
• Enables precise reaction center identification
• Underpins the training of forward reaction prediction models

A retrosynthetic strategy that applies a pre-defined library of reaction rules or subgraph patterns extracted directly from reaction databases like USPTO. The model identifies a matching template for the target molecule and applies the corresponding disconnection.

• Relies on the diversity and coverage of the USPTO training corpus
• Produces chemically interpretable and valid precursors
• Limited by the scope of the extracted template library

A sequence-to-sequence transformer architecture that treats reaction prediction as a SMILES-to-SMILES translation task. Trained extensively on the USPTO dataset, it learns to map reactant strings to product strings without explicit rule extraction.

• Achieved state-of-the-art top-1 accuracy on USPTO benchmarks
• Captures implicit reaction knowledge beyond hand-coded rules
• Uses self-attention to model long-range dependencies in molecular sequences

The task of classifying each molecule in a reaction record as a reactant, reagent, solvent, catalyst, or product. The raw USPTO dataset often lacks this distinction, requiring preprocessing to separate molecules that actively participate in bond changes from spectators.

• Critical for training models on clean reaction data
• Prevents models from learning spurious correlations with solvents
• The Pistachio dataset is a commercially curated alternative with expert-labeled roles

A validation metric that measures the consistency of a retrosynthesis model. The predicted reactants are fed into a forward prediction model, and the output is checked against the original target molecule.

• A high score indicates that the retrosynthetic disconnection is chemically reversible
• Used as a primary benchmark for models trained on USPTO data
• Exposes models that propose invalid or unstable synthons

The computational task of predicting the major product of a chemical reaction given a set of reactants and reagents. This is the inverse of retrosynthesis and is often trained on the same USPTO data to validate retrosynthetic proposals.

• Used to filter out unfeasible retrosynthetic suggestions
• Models learn to rank product distributions for reactions with multiple outcomes
• Essential for building a complete closed-loop synthesis planning system