Structure-Spectrum Fidelity in Tandem Mass Spectrometry: Quantifying the Limits of Similarity-Based Structural Inference.

Tandem mass spectral similarity is widely used to infer molecular structural relatedness, yet the fidelity of this relationship has not been systematically quantified using authenticated reference spectra. Here, we introduce Structure-Spectrum Fidelity (SSF), describing how faithfully spectral similarity preserves structural relatedness. Using 12,432 authenticated compounds from METLIN Core, we compared structural similarity across 77.3 million pairs with spectral similarity in 4000 pairs spanning the structural similarity distribution at four collision energies in both polarities. Spectral similarity preserved structural information, but SSF was moderate (Spearman rs = 0.37-0.75) and the underlying distributions were not unimodal, causing inference to fail in both directions. Across the full database at 20 eV in positive mode, 42.5% of pairs sharing an exact molecular formula and near-identical structure (ECFP4 Tanimoto ≥ 0.9; n = 15,157) produced modified cosine scores below 0.7, including 24.1% below 0.2. Conversely, when a modified cosine threshold of 0.7 was applied alone to exact-formula candidates, only 27.1% of accepted pairs were close structural analogs, despite a 27-fold enrichment over background. Increasing the threshold to 0.9 raised this proportion only to 31.5%. Discrimination peaked at intermediate collision energy and declined at higher energy as fragmentation increased similarity among unrelated compounds. Spectral similarity is therefore structurally informative but inherently probabilistic: closely related molecules do not invariably produce similar spectra, and highly similar spectra do not uniquely indicate close structural relatedness. SSF provides a quantitative basis for evaluating similarity-based structural inference.