Systematic validation of graph neural network explanations against adverse outcome pathway reactive centers for skin sensitization

Whether graph neural network (GNN) attributions capture the reactive chemistry encoded in adverse outcome pathway (AOP) annotations for skin sensitization has not been tested under a label-permutation control. We trained AttentiveFP on a 436-molecule LLNA-labeled subset (81 sensitizers) of a curated skin-sensitization dataset and extracted atom-level attributions from seven methods (integrated gradients, GradCAM, attention, GNNExplainer, PGExplainer, GraphMask, and a rank-aggregated ensemble); GCN and GIN were retrained as architecture controls. Attributions were evaluated against 50 sensitization-specific SMARTS-based reactive-center annotations spanning six molecular initiating event (MIE) mechanism classes. The LLNA-trained model reached test AUC 0.73, compared with AUC 0.49 for a shuffled-label retraining control. Across all seven attribution methods, paired real-vs-shuffled atom-AUC differences crossed zero, indicating that most apparent attribution–mechanism correlation reflects structural priors of the attribution methods rather than learned sensitization chemistry. Mechanism-class stratification surfaced one candidate exception—pre-hapten autoxidation chemistry, with a Δ = +0.26 atom-AUC gap and non-overlapping confidence intervals—on a stratum of approximately two unique molecules, reported as a hypothesis pending replication on dedicated pre-hapten benchmarks. Per-element residual analysis converges on the same interpretation: alignment for gradient- and attention-based methods drops to chance once elemental identity is regressed out. Scientific contribution This work pairs a reproducible evaluation framework (50 skin-sensitization-specific SMARTS patterns scored against atom-level GNN attributions) with a label-permutation negative control that separates learned chemistry from structural priors of the attribution method.