![Author ORCID: We display the ORCID iD icon alongside authors names on our website to acknowledge that the ORCiD has been authenticated when entered by the user. To view the users ORCiD record click the icon. [opens in a new tab]](https://www.cambridge.org/engage/assets/public/coe/logo/orcid.png){kind=logo}
Abstract
MS² spectral similarity is fundamental to interpreting LC-HRMS-NTA data. Beyond the commonly used cosine similarity, a wide range of alternative metrics, including distance-, probability-, and machine-learning–based approaches, provide different perspectives on spectral matching. In this study, we extracted 8,290 subdatasets from publicly available LC-HRMS/MS libraries (MassBank, MoNa, GNPS, and NIST), each containing spectra sharing a precursor m/z within a 5 mDa tolerance. We evaluated 20 similarity metrics by constructing single-generation local molecular networks. Most metrics failed to produce pure networks, achieving complete resolution in only ~8\% of cases at the recommended 0.7 threshold and ~20\% even when individually optimized. The number of fragment ions showed little influence on network resolution. Instead, performance was driven primarily by the similarity metric and thresholding behavior, indicating that library matching alone rarely supports identification confidence above level 3. These findings highlight the need for cumulative strategies that integrate multiple similarity perspectives and orthogonal information, such as retention time or index.
Supplementary materials
Title
Supporting information for: From Spectra to Local Networks: Evaluating MS² Similarity Metrics
Description
The Supporting Information includes distribution of InChIKeys across all subdatasets (Figure S1), heatmaps of true-positive and false-positive counts across thresholds for all similarities (Figure S2), median performance trends for all metrics and similarities (Figure S3), structural information and threshold-dependent network behavior for the example subdataset (m/z 116.143) (Figure S4), PCA score plot and hierarchical clustering of similarity metrics (Figures S5– S6), boxplots of Jaccard score, sensitivity, specificity, and precision at threshold 0.7 (Figures S7– S10), definitions of similarity metrics used in this study (Section S7) and performance metrics calculations (Section S8), schematics illustrating local network construction, central-node selection, and confusion-matrix generation (Figures S11– S13).
Actions
