Paired Experimental 1H/13C NMR Fingerprint Prediction Under Molecule- and Scaffold-Disjoint Evaluation

Machine-learning approaches for small-molecule nuclear magnetic resonance (NMR) often rely on predicted spectra or evaluation schemes that do not fully control structural overlap between training and test data. Here, we curated 555 paired experimental 1H/13C NMR observations representing 514 canonical structures from nmrshiftdb2. This dataset was used to establish an experimental benchmark for NMR-to-molecular-fingerprint prediction. Assigned chemical shifts were encoded as fixed-length assigned-shift entry-count vectors, and molecular structures as 2048-bit radius-2 Morgan fingerprints. Ridge models were evaluated using 10 molecule-disjoint and 10 Bemis–Murcko scaffold-disjoint partitions. Replicate spectra were grouped by structure, and model selection was based only on validation data. Under molecule-disjoint evaluation, fused 1H + 13C data achieved a mean Tanimoto similarity of 0.327 ± 0.019 and a Top-10 retrieval rate of 39.0%, compared with 0.316 ± 0.021 and 33.5% for 13C, and 0.268 ± 0.013 and 20.5% for 1H. Under scaffold-disjoint evaluation, fused performance decreased to 0.265 ± 0.036 and 23.7% Top-10 retrieval. Repeated validation-selected augmentation with provenance-confirmed simulated structures did not produce a consistent improvement in fingerprint similarity on experimental test data. These results provide a reproducible experimental benchmark, demonstrate complementary information from 1H and 13C peak lists, and show that scaffold novelty and spectral provenance are important factors in evaluating NMR-based molecular fingerprint prediction.