Atom Mapping

Atom mapping explicitly identifies the reaction center—the specific atoms and bonds that undergo transformation. This pinpoints the bond-breaking and bond-forming events, distinguishing reacting atoms from the inert scaffold. Accurate mapping is a prerequisite for extracting reaction templates and training template-based retrosynthesis models.

Modern atom mapping relies on several computational strategies:

• Maximum Common Substructure (MCS): Identifies the largest unchanged subgraph between reactants and products, then maps the remaining atoms via optimization.
• Indigo Toolkit: A widely-used open-source implementation using a modified MCS algorithm with heuristic scoring.
• RXNMapper: A transformer-based model trained on manually curated reaction data that predicts atom mappings with high accuracy, particularly for complex rearrangements.

Atom-mapped reaction datasets are the ground truth for supervised learning in chemical AI. Key resources include:

• USPTO Dataset: Contains millions of reactions extracted from US patents, with mappings generated algorithmically.
• Pistachio Dataset: A commercial database curated by NextMove Software, prized for its high-quality, manually verified atom mappings and reaction role labeling.
• Golden Datasets: Small, expert-curated sets used for benchmarking mapper accuracy.

The fidelity of atom mapping is evaluated through:

• Mapping Accuracy: The percentage of atoms correctly mapped compared to a ground-truth reference.
• Round-Trip Consistency: A validation check where the forward reaction product is predicted from mapped reactants; the output must match the original product.
• Valence Compliance: Ensures mapped atoms do not violate standard valence rules, flagging chemically impossible transformations.

Atom mapping is the critical preprocessing step for template-based retrosynthesis. By comparing the mapped reactant and product graphs, algorithms extract the reaction template—a subgraph isomorphism pattern encoding the local structural change. This template is then stored in a library and applied to new target molecules to propose disconnections.

A key challenge in atom mapping is accounting for reagents, solvents, and catalysts that participate in the mechanism but are not stoichiometrically incorporated into the product. Advanced mappers perform reaction role labeling to classify molecules as reactants, reagents, or solvents, ensuring only true reactants are mapped to product atoms.

The computational task of pinpointing the specific atoms and bonds directly involved in bond-breaking and bond-forming. Atom mapping is a prerequisite for this task; once atoms are mapped, the reaction center is identified as the subset of atoms whose connectivity or bond order changes between reactants and products. Algorithms often use graph edit distance to isolate these changing subgraphs.

A validation metric measuring the consistency of a reaction prediction model. It works by:

• Taking a target product and predicting retrosynthetic reactants
• Feeding those reactants into a forward prediction model
• Checking if the predicted product matches the original target High round-trip accuracy indicates that the underlying atom mapping is logically coherent and reversible.

A fixed-length vector representation encoding the structural transformation occurring at the reaction center. Atom mapping defines which atoms participate in the transformation, and the fingerprint captures the difference in their local environments. Common implementations include DRFP (differential reaction fingerprint) and rxnfp, which use mapped atoms to generate a compact, machine-readable reaction signature for similarity searching.

A commercial chemical reaction database curated by NextMove Software, extracted from patent literature. It is widely considered the gold standard for training atom mapping models because it provides expert-validated, high-quality atom mappings and explicit reaction role labeling (reactant, reagent, solvent, product). Models trained on Pistachio consistently outperform those trained on automatically mapped datasets.

A retrosynthetic strategy applying a pre-defined library of reaction rules to predict disconnections. Each rule is essentially a stored atom mapping—a subgraph transformation pattern extracted from known reactions. When a target molecule matches the product side of a template, the rule applies the reverse mapping to generate the corresponding reactant precursors. The quality of the template library depends entirely on the accuracy of the original atom mappings.

The computational step where a disconnected bond is converted into valid, synthetically equivalent molecular fragments. After atom mapping identifies which bond breaks, synthon generation completes the fragments by adding appropriate leaving groups or functional group interconversions. This step transforms an abstract disconnection into chemically realistic precursor molecules that can be sourced or further planned.