Predicting Regulator-Listed N-Nitrosamines from API Structure: A Rule-Based Method with a Retrospectively Corrected FDA and EMA Evaluation

Regulator lists of N-nitrosamines include assessed and hypothetical impurities as well as detected compounds. We evaluate a deterministic rule-based method that proposes N-nitrosamine structures from an active pharmaceutical ingredient (API) through direct nitrosation, selected nitrosative dealkylation pathways and nitrosation after one degradation step. The engine was refined against the US Food and Drug Administration (FDA) list and frozen. An external European Medicines Agency (EMA) set excludes all FDA API identities using standardized parent connectivity. Source parsing, identity grouping and analysis were corrected after the initial results had been seen; this is a retrospective evaluation. A unique API identity and reference connectivity pair is the evaluation unit. The method recovered 55 of 112 external pairs across 88 APIs, recall 0.491 (Wilson 95% interval 0.400-0.582; API-cluster bootstrap 0.398-0.588). A post hoc ablation limited to direct secondary-amine nitrosation recovered 26 pairs (0.232); the paired gain was 0.259 (cluster bootstrap 0.186-0.336). The method proposed a median of 1 and mean of 1.48 candidates per API, versus 0.33 for the ablation. Degradation routes contributed 21 hits. Development recall was 247/272 (0.908). Of 57 external misses, 29 were positive and 25 negative in a fragmentwise cleavage substructure proxy; separate adjudications identified 2 reference/name disagreements and 1 ring-hydrolysis implementation defect. These cases remain scored misses. Incomplete reference lists prevent a precision estimate, and structural matches do not establish actual impurity formation. Deposited outputs support statistical recomputation and candidate-membership checks; prediction generation requires the proprietary frozen engine.