Reaction-Center-Guided Reranking of Transformer Product Predictions with a Product-Conditioned Graph Neural Network

Forward reaction prediction is commonly formulated as sequence generation from reactant representations, but a purely sequence-based score does not explicitly assess whether a proposed product is locally consistent with the structural changes expected at a reaction center. We present a hybrid product-prediction framework that couples a compact SMILES Transformer with a product-conditioned graph neural network (GNN) and a learned candidate reranker. A high-confidence atom-mapped USPTO-50K subset containing 47,517 reactions was partitioned into record-disjoint training, validation, and held-out test sets (38,013/4,751/4,753). Reaction-center labels were enhanced to capture both bond changes and atom-state changes, eliminating 7,503 otherwise unlabeled state-only transformations. The GNN uses four GINEConv layers (hidden dimension 256) to score reactant atoms while conditioning on a product graph. In a diagnostic held-out reaction-center localization evaluation, where the recorded product graph defines the known transformation context, the GNN achieved ROC-AUC 0.9971, average precision 0.9556, F1 0.8911, and MCC 0.8860 using a validation-frozen threshold of 0.9574. For product prediction, the oracle product is not supplied: the frozen GNN is evaluated independently on each Transformer-generated candidate. A 5-beam Transformer baseline obtained 32.99% Top-1 and 45.34% Top-3 exact product accuracy. Combining Transformer ranking features with candidate-conditioned GNN reaction-center consistency descriptors increased held-out Top-1 accuracy to 35.37% and Top-3 accuracy to 46.52%. The Top-1 gain was 2.377 percentage points (7.21% relative), with 127 reactions recovered and only 14 harmed. The improvement was significant by exact McNemar testing (p = 5.82 × 10 − 24) and paired bootstrap analysis (95% CI for the gain, 1.894–2.882 percentage points). Complexity-stratified analysis showed significant gains across product-size and reactant-component strata, but no measurable advantage for reactions involving three or more reaction-center atoms. These results show that reaction-center consistency provides information complementary to sequence likelihood and can improve a compact Transformer without retraining the generator.