RetroVirtSyn: Adaptive Multi‑Center Virtual Synthons for Single‑Step Retrosynthesis

Two‑stage single‑step retrosynthesis is susceptible to error propagation. Reliable reaction‑center guidance is essential, but it should not come at the cost of losing product‑level structural context. Existing approaches face clear limitations: semi‑template methods rely on hard bond disconnections to construct synthons, so an incorrect reaction‑center prediction can disrupt the product context, introduce irreversible errors, and impair downstream generation; template‑free methods, by contrast, lack explicit multi‑center guidance and therefore struggle to represent multiple possible reaction centers. To address error propagation in two‑stage inference, we propose RetroVirtSyn, which abandons hard bond fragmentation and instead constructs confidence‑weighted virtual synthons directly on the intact product graph. By preserving useful reaction‑center information while reducing the effect of incorrect predictions, RetroVirtSyn provides a more stable interface between reaction‑center prediction and reactant generation, thereby reducing cumulative error propagation. RetroVirtSyn encodes candidate reaction centers as probabilistic conditions on the intact product graph, bridging the gap between semi‑template and template‑free approaches. It predicts reaction‑center confidence and employs an adaptive Top‑$m$ strategy to dynamically retain candidate centers: high‑confidence samples keep only the Top‑1 center, whereas low‑confidence samples retain the Top‑3 centers, balancing center precision and candidate coverage. The retained centers are fused into one confidence‑weighted virtual synthon, which, together with the original product graph, conditions a discrete flow matching model. Reactants are generated without explicitly fragmenting the product graph. On USPTO‑50K under the reaction‑class‑unknown setting, the fixed $\(m=1\)$ strategy achieves a Top‑1 accuracy of 63.5%, outperforming the strongest baseline by 3.5 percentage points. The adaptive Top‑$m$ strategy attains Top‑3, Top‑5, and Top‑10 accuracies of 81.3%, 84.1%, and 86.5%, exceeding the corresponding best baselines by 2.2, 1.4, and 0.7 percentage points, respectively. These results show that RetroVirtSyn reduces error propagation and improves retrosynthesis accuracy.