Reaction Class Tokenization

A special classification token (e.g., <RX_1>, <AmideBondFormation>) is prepended to the reactant SMILES string before encoding. This token acts as a global conditioning signal that biases the entire transformer's self-attention mechanism toward reaction-specific patterns. Unlike implicit learning where the model must infer reaction type from reactants alone, this explicit signal reduces ambiguity and improves prediction accuracy for rare reaction classes with limited training examples.

The tokenization strategy enables a single unified model to handle diverse chemical transformations simultaneously. Key architectural benefits include:

• Shared encoder layers learn universal molecular representations
• Class-specific attention heads specialize in distinct reaction mechanisms
• Dynamic routing through the token embedding directs information flow This eliminates the need for separate models per reaction type, reducing deployment complexity and enabling transfer learning between related reaction classes.

Reaction class tokens are derived from systematic classification schemes such as NameRXN or RXNO ontology. The vocabulary is constructed by:

• Clustering reactions by bond changes at the reaction center
• Assigning hierarchical tokens for super-class and sub-class granularity
• Reserving special tokens for unknown or novel reaction types This structured vocabulary allows the model to generalize across similar mechanisms while maintaining specificity for distinct transformations.

During training, each reaction SMILES sequence is prefixed with its corresponding class token. The model learns to:

• Attend to the class token when predicting product atom ordering
• Modulate output distributions based on reaction type priors
• Handle class imbalance through token-frequency weighted sampling This approach has demonstrated 15-20% improvement in top-1 accuracy on the USPTO benchmark compared to unconditioned sequence-to-sequence models, particularly for under-represented reaction classes.

In retrosynthetic workflows, reaction class tokens enable template-free models to generate contextually appropriate precursors. The token conditions the decoder to produce synthons that are:

• Chemically valid for the specified transformation type
• Synthetically accessible with commercially available building blocks
• Stereochemically consistent with the reaction mechanism This bridges the gap between purely data-driven generation and rule-guided synthesis planning, improving the practical utility of AI-generated synthetic routes.

Reaction class tokenization has been validated across multiple standard datasets:

• USPTO-50k: 92.4% top-1 accuracy for product prediction with class tokens vs. 88.7% without
• Pistachio: Robust handling of 1000+ reaction classes with hierarchical token schemes
• Cross-validation: Consistent improvement across template-based and template-free architectures The technique is now a standard component in state-of-the-art molecular transformer implementations, including the Molecular Transformer and Graphormer variants adapted for reaction prediction.