Improved Ranking of Putative Candidates Through a Hybrid In Silico/ Real Fragmentation Technique

Importance of the topic

Accurate structural identification of unknown small molecules remains a major bottleneck in metabolomics, environmental analysis and other fields relying on mass spectrometry. Reference spectral libraries give high-confidence identifications when matches exist, but library incompleteness forces routine reliance on chemical databases that return large numbers of putative candidates. Combining experimentally observed fragmentation evidence with database-derived structures improves candidate prioritization and reduces manual interpretation time. The work summarized here describes a practical, hybrid algorithmic approach that leverages real spectral library fragmentation data together with chemical database searching to produce a ranked list of likely structures for unknown compounds.

Objectives and study overview

The primary objective was to demonstrate an algorithm (mzLogic) that ranks putative chemical-database candidates for unknown compounds by integrating: (1) chemical database search results (ChemSpider), (2) spectral library similarity search (mzCloud) in a similarity mode, and (3) mapping of real fragment structures from the library onto candidate structures to compute explained substructure coverage. The algorithm was tested on compounds absent from the reference library (examples include glycyl-prolyl-glutamic acid and rosmarinic acid) to evaluate its ability to reduce candidate lists and prioritize correct or closely related structures.

Methodology

High-resolution accurate-mass (HRAM) LC–MSn data were acquired for single standards and small mixes. Key steps of mzLogic processing are:

• Derive molecular weight and/or elemental composition from high-resolution MS1 data, using isotopic fine structure and fragment coverage to refine composition candidates.
• Search chemical databases (ChemSpider) by molecular weight or elemental composition to assemble a list of putative structures.
• Run a spectral similarity search against mzCloud without constraining precursor mass; retrieve similarity hits from any level of the MSn tree and evaluate both forward and reverse similarity.
• Map annotated fragment substructures from the mzCloud similarity hits onto each chemical-database candidate to determine the maximum explained substructure and the proportion of the candidate explained by real observed fragments.
• Compute combined scores that integrate spectral similarity and structural/explained-substructure coverage to produce a final ranked list of candidates.

Instrumentation used

The experimental workflow and acquisition parameters reported in the study included:

• Mass spectrometer: Thermo Scientific Orbitrap Fusion Tribrid MS.
• LC system: Thermo Scientific Vanquish UHPLC with a Hypersil GOLD 100 x 5 mm, 3 µm C18 column at 35 °C.
• Ionization: Electrospray ionization in both positive and negative modes (separate injections).
• Acquisition: Full MS1 at 60,000 FWHM (m/z 200); data-dependent MS2 by HCD with stepped collision energy 40% ± 20% at 30,000 resolution; MS3 on top 3 MS2 ions by trap CID at 30% NCE.
• Sample prep: Standards dissolved in DMSO or MeOH to 0.1–0.5 mM stocks, diluted to ~50 nM for injection.
• Software: mzLogic implemented in Thermo Scientific Mass Frontier 8.0 and Compound Discoverer 3.0; spectral library used: mzCloud; chemical database: ChemSpider.

Main results and discussion

Key findings demonstrated the ability of mzLogic to meaningfully prioritize plausible candidates by combining spectral-similarity information and real fragmentation mapping:

• Explained-substructure ranking: By mapping annotated fragments from mzCloud similarity hits onto each chemical-database candidate, mzLogic quantifies how much of a candidate structure is supported by real observed fragments. Candidates with higher proportions of their structure explained by library-derived fragments receive higher ranks.
• Elemental composition refinement: High-resolution MS1 fine isotope patterns plus MS/MS fragment composition constraints were combined to refine elemental composition candidates and reduce the database search space before structural ranking.
• Example — glycyl-prolyl-glutamic acid (MW 302.1344): From 36 database hits the algorithm ranked the correct peptide-like structure highly; the top hits included structurally related compounds and the algorithm deprioritized less peptide-like candidates despite similar molecular weight.
• Example — rosmarinic acid: From over 250 initial database candidates, mzLogic substantially reduced complexity; the correct structure was returned as the second-ranked candidate while structurally similar compounds populated the top ranks.
• Improved reliability vs. purely in silico fragmentation: Using annotated, real fragmentation patterns from mzCloud reduces errors and biases inherent to purely predictive fragmentation models and provides experimentally validated fragment-to-substructure links.

Benefits and practical applications

The mzLogic hybrid approach offers several practical advantages for analytical laboratories:

• Significant reduction in candidate lists, decreasing manual review time and accelerating identification workflows in metabolomics, environmental analysis, and forensic screening.
• Improved prioritization of chemically plausible and fragment-supported candidates, delivering higher-confidence leads even when exact library matches are absent.
• Compatibility with MSn data: leveraging higher-order fragmentation enables finer substructure mapping and better discrimination among isomeric or closely related structures.
• Integration into existing software ecosystems (Mass Frontier, Compound Discoverer) enables adoption without major workflow re-engineering.

Future trends and potential applications

Opportunities and likely directions for development include:

• Broader library coverage and richer fragment annotations will further enhance the power of hybrid approaches; community curation of fragment annotations could be beneficial.
• Integration of complementary in silico fragmentation predictors with confidence weighting could help when no good similarity hits exist, creating a more seamless hybrid predictive/empirical framework.
• Machine-learning models trained on library fragment-to-substructure mappings may improve automated mapping and scoring.
• Application to large-scale non-targeted studies: automated ranking can be embedded in high-throughput pipelines to flag high-priority features for targeted follow-up.
• Extension to adduct-aware and stereoisomer-aware ranking strategies could further improve identification specificity.

Conclusion

The mzLogic algorithm demonstrates a pragmatic, hybrid strategy to improve candidate ranking for unknown small molecules by combining chemical-database searching with spectral-library similarity and real fragmentation mapping. By using experimentally observed fragment annotations to validate and quantify substructure coverage across database candidates, mzLogic reduces ambiguity, elevates chemically plausible structures, and mitigates risks associated with purely in silico fragmentation prediction. The approach is practically implemented in common data-analysis tools and is applicable to workflows that employ HRAM MSn data.

References

1. Ridder L, van der Hooft J, Verhoeven S, de Vos R, van Schaik R, Vervoort J. Substructure-based annotation of high-resolution multistage MSn spectral trees. Rapid Communications in Mass Spectrometry, 2012. doi:10.1002/rcm.6364
2. Allen F, Pon A, Wilson M, Greiner R, Wishart D. CFM-ID: a web server for annotation, spectrum prediction and metabolite identification from tandem mass spectra. Nucleic Acids Research, June 2014.