Round-Trip Accuracy

Round-trip accuracy measures whether a retrosynthesis model's predictions are internally consistent. The process takes a target molecule, performs retrosynthesis to generate reactant suggestions, then runs forward reaction prediction on those reactants. If the forward-predicted product matches the original target, the round-trip is considered successful. This metric is critical because it does not require ground-truth reaction data for validation, making it applicable to novel targets where no known synthetic route exists.

The metric is computed as the percentage of retrosynthesis predictions that successfully reconstruct the original target:

• Step 1: Input target molecule T into the retrosynthesis model
• Step 2: Generate top-k sets of reactant predictions {R₁, R₂, ..., Rₖ}
• Step 3: For each reactant set, run forward prediction to obtain product P
• Step 4: Compare P to T using canonical SMILES or InChI string matching
• Step 5: Calculate accuracy as (successful round-trips / total predictions) × 100

Top-k round-trip accuracy reports whether at least one of the k predictions completes the cycle successfully.

While useful, round-trip accuracy has known failure modes:

• Trivial solutions: A model may propose the target molecule itself as a reactant, achieving perfect round-trip accuracy without providing useful disconnections
• Forward model bias: Errors in the forward prediction model can mask retrosynthesis errors or falsely penalize correct disconnections
• Reaction condition blindness: The metric ignores whether the proposed reaction is actually feasible under realistic laboratory conditions
• Multi-step gap: Round-trip accuracy evaluates single-step proposals and does not guarantee that a complete multi-step pathway will succeed

These limitations mean round-trip accuracy should be used alongside other metrics like coverage and synthetic accessibility scores.

Round-trip accuracy creates a symbiotic evaluation loop between retrosynthesis and forward prediction models. The quality of the metric depends entirely on the forward predictor's fidelity. A highly accurate forward model like the Molecular Transformer trained on the USPTO dataset provides a reliable evaluation signal. Conversely, a poor forward predictor introduces noise that undermines the metric's validity. This interdependence has driven research into jointly training retrosynthesis and forward models to maximize mutual consistency.

Round-trip accuracy has become a standard evaluation metric in retrosynthesis literature, appearing alongside top-k exact match accuracy against held-out test sets. Key benchmarking practices include:

• Reporting both top-1 and top-10 round-trip accuracy
• Using canonical SMILES representation to avoid string-matching artifacts
• Comparing against template-based baselines like template-based retrosynthesis systems
• Evaluating on diverse datasets including Pistachio and USPTO to ensure generalizability

Leading template-free models achieve round-trip accuracies exceeding 90% on standard benchmarks.

High single-step round-trip accuracy correlates with but does not guarantee high-quality retrosynthetic trees. When a model recursively applies disconnections, small single-step errors compound exponentially. A model with 95% round-trip accuracy per step will have only approximately 0.95ⁿ probability of a fully valid n-step pathway. This compounding effect makes round-trip accuracy a necessary but insufficient condition for successful multi-step retrosynthetic planning. Researchers often combine it with Monte Carlo Tree Search (MCTS) rollouts that validate entire pathways end-to-end.

The computational task of predicting the major product of a chemical reaction given a set of reactants, reagents, and conditions. In the context of round-trip accuracy, a forward prediction model acts as the verification engine: it takes the reactants proposed by a retrosynthetic model and predicts the product. If this predicted product matches the original target molecule, the round-trip is considered consistent. Modern approaches use Molecular Transformers or Graph Neural Networks to achieve top-1 accuracy exceeding 90% on benchmark datasets like USPTO.

The process of establishing a one-to-one correspondence between atoms in the reactants and atoms in the products of a chemical reaction. Accurate atom mapping is a prerequisite for calculating round-trip accuracy because it defines the reaction center and ensures that the retrosynthetic disconnection is chemically meaningful. Without correct atom mapping, a forward prediction model cannot verify whether the proposed reactants truly reconstruct the target molecule's exact atomic connectivity.

A retrosynthetic strategy that applies a pre-defined library of reaction rules or subgraph patterns to predict disconnections in a target molecule. Round-trip accuracy is often used to filter and validate these templates. A template is considered high-quality if, when applied to a target to generate reactants, a forward predictor can reliably reconstruct the target from those reactants. This metric helps prune noisy or overly generic templates from knowledge bases.

A sequence-to-sequence transformer architecture that treats reaction prediction as a SMILES-to-SMILES translation task. This model can be trained bidirectionally to perform both retrosynthesis and forward prediction. When a single Molecular Transformer is used for both tasks, round-trip accuracy becomes a measure of internal consistency—the model's ability to translate a target to reactants and then back to the identical target string, revealing hallucinations or chemically invalid proposals.

A hybrid approach that first identifies a reaction center using a template, then completes the synthon generation using a template-free generative model. Round-trip accuracy is particularly valuable here because the generative completion step can introduce chemically invalid substructures. By checking that the completed reactants forward-predict to the original target, researchers can quantify how often the template-free completion module generates viable synthons.

The computational task of pinpointing the specific atoms and bonds that are directly involved in bond-breaking and bond-forming during a chemical reaction. Round-trip accuracy depends heavily on correct reaction center identification. If a retrosynthetic model proposes a disconnection at the wrong site, the resulting reactants will be structurally incompatible with the target, and the forward verification step will fail. This metric thus indirectly evaluates the quality of reaction center localization.