Peak Risk Score: A Peak-Space Verification Layer for AI Scientists in NMR-Guided Molecular Discovery

AI systems for molecular discovery increasingly generate candidate structures together with simulated evidence, but existing evaluation methods struggle when observations contain missing peaks, spurious peaks, and uncertainty in peak cardinality. In AI-driven discovery pipelines, models generate hypotheses, simulate observations, and must decide which candidates are consistent with experimental measurements; errors at this stage can lead to overconfident but incorrect selections. In NMR-guided structure elucidation, standard metrics such as shift MAE or RMSE can reward low average error while missing chemically meaningful peak-pattern mismatches. We introduce the Peak Risk Score (PRS), a model-agnostic probabilistic scoring framework for spectroscopy-grounded molecular hypotheses. PRS treats simulated spectra as approximate forecasts, constructs predictive ensembles that capture realistic variability, and evaluates them in peak space using proper scoring-rule components for peak existence, position, and integral/amplitude, yielding interpretable peak-wise contributions. On a synthetic point-process benchmark, PRS detects over-predictive models that add spurious events while preserving low matched-event error, identifying the correct forecaster in 99.6% of cases compared with 28–93% for standard baselines. On 1,849 paired simulated/experimental 13C NMR spectra, PRS achieves 95.5% top-1 accuracy in a structure-identification task with one correct candidate and four decoys, outperforming the best classical shift-based metric, 62.5%, and Wasserstein-1 distance, 93.3%, while matching DP4 and DP5 probabilities, 95.0% and 95.5%. Combining PRS with DP5 further improves accuracy to 95.8%. PRS provides a domain-grounded verification layer for AI-scientist workflows by explicitly modeling uncertainty, penalizing structural mismatches, and handling variable-cardinality experimental observations.