Chapter 7 Annotation
Once a peak table or feature table has been generated, annotation becomes the key step that connects mass spectrometry signals to chemical meaning. For broader overviews, see(Domingo-Almenara, Montenegro-Burke, Benton, et al. 2018; Chaleckis et al. 2019; Lai et al. 2018; Nash and Dunn 2019; Viant et al. 2017; Allard et al. 2017). The first review proposed five levels for current computational annotation strategies.
Level 1: Peak grouping: MS pseudospectra extraction based on peak shape similarity and peak abundance correlation
Level 2: Peak annotation: adducts, neutral losses, isotopes, and other mass relationships based on mass distances
Level 3: Biochemical knowledge based on putative identification, potential biochemical reactions, and related statistical analysis
Level 4: Use and integration of tandem MS data based on data-dependent or data-independent acquisition mode, or in silico prediction
Level 5: Retention time prediction based on library-available retention indices or quantitative structure-retention relationship (QSRR) models.
Most of the software are at level 1 or 2. If we only have compounds structure, we could guess ions under different ionization method. If we have mass spectrum, we could match the mass spectral by a similarity analysis to the database. In metabolomics, we only have mass spectrum or mass-to-charge ratios. Single mass-to-charge ratio is not enough for identification. That’s the one bottleneck for annotation. So prediction is always performed on MS/MS data.
7.1 A practical annotation workflow
In practice, annotation is easier to understand as a workflow rather than as a list of tools. A useful order is:
- understand annotation limits and confidence levels
- constrain candidate formulas from accurate mass and isotopic information
- collapse redundant MS1 features such as adducts, isotopes, and in-source fragments
- connect representative MS1 features to MS2 or MSn data
- rank candidate structures with MS/MS libraries and in silico tools
- refine the result with retention, biochemical knowledge, and experimental context
- report annotation confidence clearly instead of overclaiming identification
The dark metabolome debate is part of this workflow rather than a side topic: before interpreting unknowns, we need to ask how many features represent real unique compounds and how many are redundant signals from the same chemistry.
7.2 Step 0: understand annotation limits
The major issue in annotation is the redundancy peaks from same metabolite. Unlike genomes, peaks or features from peak selection are not independent with each other. Adducts, in-source fragments and isotopes would lead to wrong annotation. A common solution is that use known adducts, neutral losses, molecular multimers or multiple charged ions to compare mass distances.
Another issue is about the MS/MS database. Only 10% of known metabolites in databases have experimental spectral data. Thus in silico prediction is required. Some works try to fill the gap between experimental data, theoretical values(from chemical database like chemspider) and prediction together. Here is a nice review about MS/MS prediction(Hufsky et al. 2014).
7.2.1 Common sources of peak misidentification
- Isomer
Use separation methods such as chromatography, ion mobility MS, MS/MS. Reversed-phase ion-pairing chromatography and HILIC is useful. Chemical derivatization is another option.
- Interfering compounds
20ppm is the least exact mass accuracy for HRMS.
- In-source degradation products
7.2.2 Annotation vs. Identification
According to the definition from the Chemical Analysis Working Group of the Metabolomics Standards Initiative(Sumner et al. 2007; Viant et al. 2017), four levels of confidence could be assigned to identification:
- Level 1 ‘identified metabolites’
- Level 2 ‘Putatively annotated compounds’
- Level 3 ‘Putatively characterised compound classes’
- Level 4 ‘Unknown’
Schymanski et al. proposed a complementary five-level confidence scheme specifically designed for high-resolution mass spectrometry that has become widely adopted in environmental and exposome studies(Schymanski et al. 2014). The levels range from Level 1 (confirmed structure by reference standard) through Level 5 (exact mass only), with intermediate levels for probable structure (Level 2, diagnostic evidence but no reference standard), tentative candidates (Level 3, evidence for possible structures), and unequivocal molecular formula (Level 4). This framework provides more granularity than the MSI scheme and is especially useful for non-target screening where reference standards are unavailable for most detected features.
In practice, data analysis based annotation could reach Level 2. For Level 1, we need extra methods such as MS/MS, retention time, accurate mass, 2D NMR spectra, and so on to confirm the compounds. However, standards are always required for solid proof.
For specific group of compounds such as PFASs, the communication of confidence level could be slightly different(Charbonnet et al. 2022).
Through MS/MS seemed a required step for identification, recent study found ESI might also generate fragments ions for structure identification (Xue, Guijas, et al. 2020; Xue et al. 2021, 2023; Bernardo-Bermejo et al. 2023).
7.2.3 The Dark Metabolome
In a typical untargeted LC-MS experiment, thousands of features are detected, but only a small fraction can be annotated with known metabolite identities. The vast majority of detected features remain unidentified, a phenomenon often referred to as the “dark matter” of the metabolome(Silva et al. 2015). Even in well-studied matrices like blood plasma, if a feature is detected in over 50% of samples, there is roughly a 50% chance it has a known annotation; for low-abundance features, annotation rates often fall below 5%. This annotation gap is one of the central challenges in metabolomics and has sparked active debate about its causes and solutions(Petras et al. 2018; Koelmel et al. 2025).
7.2.3.1 What makes up the dark matter?
The detected features in an untargeted experiment are not all independent metabolites. A single compound can generate multiple features through several mechanisms:
In-source fragmentation (ISF): Molecules can fragment during the electrospray ionization process before entering the mass analyzer, generating fragment ions that appear as separate features. The extent of ISF is a subject of active debate (see below).
Adducts and multimers: Beyond the protonated or deprotonated molecular ion, compounds readily form sodium, potassium, ammonium adducts, as well as dimers and trimers, each appearing as a separate feature.
Isotope peaks: Natural isotope distributions (especially 13C, 34S, 37Cl) generate satellite peaks for every compound.
Multiple charge states: Some compounds, particularly larger ones, can carry multiple charges.
After accounting for these redundancies, the number of unique compounds in a sample is substantially smaller than the number of detected features. However, even after deduplication, the majority of unique compounds remain unannotated because they are absent from existing spectral databases.
7.2.3.2 The in-source fragmentation debate
In 2024, Giera et al. analyzed the METLIN database of 931,000 molecular standards and reported that in-source fragmentation could account for over 70% of the peaks observed in typical LC-MS metabolomic datasets(Giera et al. 2024). This finding suggested that the dark metabolome might be largely a measurement artifact rather than representing genuine molecular diversity. A follow-up study further examined the relationship between ISF and the dark metabolome/lipidome(Uritboonthai et al. 2025).
However, this conclusion was challenged. Li and Mahieu performed a systematic analysis of 61 representative public LC-MS datasets and found that in-source fragments contribute to less than 10% of features in real experimental data(Li and Mahieu 2025). Their khipu-based pre-annotation approach showed that the majority of abundant features have identifiable ion patterns (adducts, isotopes, etc.), and that the dark matter is explainable in an abundance-dependent manner: most features come from real compounds, but the number of unique compounds is much smaller than the number of features. A separate perspective from de la Briere et al. further examined the impact of unintentional fragments on molecular networking and public repository-scale analysis(Briere et al. 2025).
This debate reflects a key tension in the field: the extent of ISF depends heavily on the instrument, ionization source design, mobile phase composition and applied voltages, making it difficult to generalize from standard databases to real experimental conditions. What is clear is that both ISF and genuine molecular diversity contribute to the dark metabolome, and the relative contribution varies by sample type and analytical conditions.
7.2.3.3 A missing perspective: data-driven deduplication
Notably, the above debate has largely focused on either standard-library-based estimation (Giera/Siuzdak) or ion-pattern-based pre-annotation (Li/Mahieu). Neither side has systematically addressed the question from a mass distance statistics perspective. The Paired Mass Distance (PMD) approach(M. Yu et al. 2019) offers a complementary, data-driven strategy: by analyzing the frequency distribution of mass differences between all feature pairs in a dataset, high-frequency PMDs that correspond to known chemical relationships (adducts, in-source fragments, biotransformations) can be identified directly from the data without relying on a predefined adduct list or a standard compound library.
The GlobalStd algorithm(M. Yu et al. 2019), built on PMD analysis, takes this further by selecting independent “precursor ions” from the feature list — features whose mass differences to other features cannot be explained by high-frequency PMDs are more likely to represent unique parent compounds. This effectively collapses a feature list of thousands into a much smaller set of candidate independent compounds, providing a quantitative and data-driven estimate of the true chemical complexity in a sample. Such an approach could provide additional evidence for the dark matter debate: by applying GlobalStd to the same public datasets used by either side, one could independently estimate the ratio of redundant features to unique compounds without assumptions about ISF rates from standard libraries.
Furthermore, the PMDDA workflow(Yu et al. 2022) extends this logic to connect MS1 features with MS2 data acquisition: by using GlobalStd to prioritize truly independent precursor ions for targeted MS2 collection, it addresses both the redundancy problem and the annotation gap simultaneously.
7.2.3.4 Strategies to address the dark matter
Several complementary approaches are being developed to reduce the annotation gap:
Pre-annotation and feature deduplication: Tools like khipu(Li and Mahieu 2025) and CAMERA(Kuhl et al. 2012) group related features (adducts, isotopes, fragments) based on ion pattern recognition and co-elution. The PMD/GlobalStd approach(M. Yu et al. 2019) provides an orthogonal, data-driven method by identifying redundant features through mass distance frequency analysis without requiring a predefined adduct list.
Molecular networking: GNPS-based molecular networking(Wang et al. 2016) connects structurally related compounds through MS/MS spectral similarity, enabling propagation of annotations from known to unknown compounds within the same spectral family.
In silico prediction: Tools like SIRIUS/CSI:FingerID(Dührkop et al. 2019), CFM-ID(Allen et al. 2014), and MS2DeepScore(Huber et al. 2020) predict molecular properties or spectra from structure, expanding the searchable chemical space beyond experimentally characterized compounds.
Machine learning approaches: Deep learning models are increasingly used to embed spectral data into chemical space, enabling analogue searching and compound class prediction even without exact database matches (see Statistical Analysis chapter for details).
For a comprehensive review of bioinformatics tools tackling the dark metabolome, see Koelmel et al.(Koelmel et al. 2025).
7.3 Step 1: constrain candidates by molecular formula
Once the annotation limits are clear, the next step is to reduce the candidate space using accurate mass, isotope patterns, and chemical plausibility rules before structural searching.
Cheminformatics will help for MS annotation. The first task is molecular formula assignment. For a given accurate mass, the formula should be constrained by predefined element type and atom number, mass error window and rules of chemical bonding, such as double bond equivalent (DBE) and the nitrogen rule. The nitrogen rule is that an odd nominal molecular mass implies also an odd number of nitrogen. This rule should only be used with nominal (integer) masses. Degree of unsaturation or DBE use rings-plus-double-bonds equivalent (RDBE) values, which should be interger. The elements oxygen and sulphur were not taken into account. Otherwise the molecular formula will not be true.
\[RDBE = C+Si - 1/2(H+F+Cl+Br+I) + 1/2(N+P)+1 \]
To assign molecular formula to a mass to charge ratio, Seven Golden Rules (Kind and Fiehn 2007) for heuristic filtering of molecular formulas should be considered:
- Apply heuristic restrictions for number of elements during formula generation. This is the table for known compounds:
## Mass.Range.[Da] Library C.max H.max N.max O.max P.max S.max F.max Cl.max
## 1 < 500 DNP 29 72 10 18 4 7 15 8
## 2 <NA> Wiley 39 72 20 20 9 10 16 10
## 3 < 1000 DNP 66 126 25 27 6 8 16 11
## 4 <NA> Wiley 78 126 20 27 9 14 34 12
## 5 < 2000 DNP 115 236 32 63 6 8 16 11
## 6 <NA> Wiley 156 180 20 40 9 14 48 12
## 7 < 3000 DNP 162 208 48 78 6 9 16 11
## Br.max Si.max
## 1 5 NA
## 2 4 8
## 3 8 NA
## 4 8 14
## 5 8 NA
## 6 10 15
## 7 8 NAPerform LEWIS and SENIOR check. The LEWIS rule demands that molecules consisting of main group elements, especially carbon, nitrogen and oxygen, share electrons in a way that all atoms have completely filled s, p-valence shells (‘octet rule’). Senior’s theorem requires three essential conditions for the existence of molecular graphs
The sum of valences or the total number of atoms having odd valences is even;
The sum of valences is greater than or equal to twice the maximum valence;
The sum of valences is greater than or equal to twice the number of atoms minus 1.
Perform isotopic pattern filter. Isotope ratio abundance was included in the algorithm as an additional orthogonal constraint, assuming high quality data acquisitions, specifically sufficient ion statistics and high signal/noise ratio for the detection of the M+1 and M+2 abundances. For monoisotopic elements (F, Na, P, I) this rule has no impact. isotope pattern will be useful for brominated, chlorinated small molecules and sulphur-containing peptides.
Perform H/C ratio check (hydrogen/carbon ratio). In most cases the hydrogen/carbon ratio does not exceed H/C > 3 with rare exception such as in methylhydrazine (CH6N2). Conversely, the H/C ratio is usually smaller than 2, and should not be less than 0.125 like in the case of tetracyanopyrrole (C8HN5).
Perform NOPS ratio check (N, O, P, S/C ratios).
## Element.ratios Common.range.(covering.99.7%) Extended.range.(covering.99.99%)
## 1 H/C 0.2–3.1 0.1–6
## 2 F/C 0–1.5 0–6
## 3 Cl/C 0–0.8 0–2
## 4 Br/C 0–0.8 0–2
## 5 N/C 0–1.3 0–4
## 6 O/C 0–1.2 0–3
## 7 P/C 0–0.3 0–2
## 8 S/C 0–0.8 0–3
## 9 Si/C 0–0.5 0–1
## Extreme.range.(beyond.99.99%)
## 1 < 0.1 and 6–9
## 2 > 1.5
## 3 > 0.8
## 4 > 0.8
## 5 > 1.3
## 6 > 1.2
## 7 > 0.3
## 8 > 0.8
## 9 > 0.5- Perform heuristic HNOPS probability check (H, N, O, P, S/C high probability ratios)
df <- data.frame(
stringsAsFactors = FALSE,
Element.counts = c("NOPS all > 1","NOP all > 3","OPS all > 1",
"PSN all > 1","NOS all > 6"),
Heuristic.Rule = c("N< 10, O < 20, P < 4, S < 3",
"N < 11, O < 22, P < 6","O < 14, P < 3, S < 3",
"P < 3, S < 3, N < 4","N < 19 O < 14 S < 8"),
DB.examples.for.maximum.values = c("C15H34N9O8PS, C22H44N4O14P2S2, C24H38N7O19P3S","C20H28N10O21P4, C10H18N5O20P5",
"C22H44N4O14P2S2, C16H36N4O4P2S2",
"C22H44N4O14P2S2, C16H36N4O4P2S2","C59H64N18O14S7")
)
df## Element.counts Heuristic.Rule
## 1 NOPS all > 1 N< 10, O < 20, P < 4, S < 3
## 2 NOP all > 3 N < 11, O < 22, P < 6
## 3 OPS all > 1 O < 14, P < 3, S < 3
## 4 PSN all > 1 P < 3, S < 3, N < 4
## 5 NOS all > 6 N < 19 O < 14 S < 8
## DB.examples.for.maximum.values
## 1 C15H34N9O8PS, C22H44N4O14P2S2, C24H38N7O19P3S
## 2 C20H28N10O21P4, C10H18N5O20P5
## 3 C22H44N4O14P2S2, C16H36N4O4P2S2
## 4 C22H44N4O14P2S2, C16H36N4O4P2S2
## 5 C59H64N18O14S7- Perform TMS check (for GC-MS if a silylation step is involved). For TMS derivatized molecules detected in GC/MS analyses, the rules on element ratio checks and valence tests are hence best applied after TMS groups are subtracted, in a similar manner as adducts need to be first recognized and subtracted in LC/MS analyses.
Seven Golden Rules were built for GC-MS and Hydrogen Rearrangement Rules were major designed for LC-CID-MS/MS(Tsugawa et al. 2016). Based on extensively curated database records and enthalpy calculations, “hydrogen rearrangement (HR) rules” could be extending the even-electron rule for carbon (C) and heteroatoms, oxygen (O), nitrogen (N), phosphorus (P), and sulfur (S). They used high abundance MS/MS peaks that exceeded 10% of their base peaks to identify common features in terms of 4 HR rules for positive mode and 5 HR rules for negative mode.
Seven Golden Rules and Hydrogen Rearrangement Rules might also be captured by statistical models. However, such heuristic rules could reduce the searching space of possible formula.
molgen generating all structures (connectivity isomers, constitutions) that correspond to a given molecular formula, with optional further restrictions, e.g. presence or absence of particular substructures (Gugisch et al. 2015).
mfFinder can predict formula based on accurate mass (Patiny and Borel 2013).
RAMSI is the robust automated mass spectra interpretation and chemical formula calculation method using mixed integer linear programming optimization (Baran and Northen 2013).
Here is some other Cheminformatics tools, which could be used to assign meaningful formula or structures for mass spectra.
- RDKit Open-Source Cheminformatics Software
- cdk The Chemistry Development Kit (CDK) is a scientific, LGPL-ed library for bio- and cheminformatics and computational chemistry written in Java (Guha 2007).
- Open Babel Open Babel is a chemical toolbox designed to speak the many languages of chemical data (O’Boyle et al. 2011).
- ClassyFire is a tool for automated chemical classification with a comprehensive, computable taxonomy (Djoumbou Feunang et al. 2016).
- BUDDY can perform molecular formula discovery via bottom-up MS/MS interrogation(Xing et al. 2023).
7.4 Step 2: collapse redundant MS1 features
Before MS/MS ranking or database matching, it is usually necessary to reduce the feature table into a smaller set of representative precursor candidates.
Full scan mass spectra always contain lots of redundant peaks such as adducts, isotope, fragments, multiple charged ions and other oligomers. Such peaks dominated the features table(Xu et al. 2015; Sindelar and Patti 2020; Mahieu and Patti 2017). Annotation tools could label those peaks either by known list or frequency analysis of the paired mass distances(Ju et al. 2020; Kouřil et al. 2020).
7.4.1 Adducts list
You could find adducts list here from commonMZ project.
7.4.2 Isotope
Here is Isotope pattern prediction.
7.4.3 CAMERA
Common annotation for xcms workflow(Kuhl et al. 2012).
7.4.4 RAMClustR
The software could be found here (Broeckling et al. 2014; Broeckling et al. 2016). The package included a vignette to follow.
7.4.5 BioCAn
BioCAn combines the results from database searches and in silico fragmentation analyses and places these results into a relevant biological context for the sample as captured by a metabolic model (Alden et al. 2017).
7.4.6 mzMatch
mzMatch is a modular, open source and platform independent data processing pipeline for metabolomics LC/MS data written in the Java language. (Chokkathukalam et al. 2013; Scheltema et al. 2011) and MetAssign is a probabilistic annotation method using a Bayesian clustering approach, which is part of mzMatch(Daly et al. 2014).
7.4.7 xMSannotator
The software could be found here(Uppal et al. 2017).
7.4.8 mWise
mWise is an Algorithm for Context-Based Annotation of Liquid Chromatography–Mass Spectrometry Features through Diffusion in Graphs(Barranco-Altirriba et al. 2021).
7.4.9 MAIT
You could find source code here(Fernández-Albert et al. 2014).
7.4.10 pmd
Paired Mass Distance(PMD) analysis for GC/LC-MS based nontarget analysis to remove redundant peaks(M. Yu et al. 2019).
7.4.11 nontarget
nontarget could find Isotope & adduct peak grouping, and perform homologue series detection (Loos and Singer 2017).
7.4.12 Binner
Binner Deep annotation of untargeted LC-MS metabolomics data (Kachman et al. 2020)
7.4.13 mz.unity
You could find source code here (Mahieu, Spalding, Gelman, et al. 2016) and it’s for detecting and exploring complex relationships in accurate-mass mass spectrometry data.
7.4.14 MS-FLO
ms-flo A Tool To Minimize False Positive Peak Reports in Untargeted Liquid Chromatography–Mass Spectroscopy (LC-MS) Data Processing (DeFelice et al. 2017).
7.4.15 CliqueMS
CliqueMS is a computational tool for annotating in-source metabolite ions from LC-MS untargeted metabolomics data based on a coelution similarity network (Senan et al. 2019).
7.4.16 InterpretMSSpectrum
This package is for annotate and interpret deconvoluted mass spectra (mass*intensity pairs) from high resolution mass spectrometry devices. You could use this package to find molecular ions for GC-MS (Jaeger et al. 2016).
7.4.17 NetID
NetID is a global network optimization approach to annotate untargeted LC-MS metabolomics data(Chen et al. 2021).
7.4.18 ISfrag
De Novo Recognition of In-Source Fragments for Liquid Chromatography–Mass Spectrometry Data(Guo et al. 2021)
7.4.19 FastEI
Ultra-fast and accurate electron ionization mass spectrum matching for compound identification with million-scale in-silico library(Yang et al. 2023)
7.5 Step 3: connect MS1 features to MS2 acquisition
After redundant features are reduced, the next task is to connect the representative MS1 features to informative MS2 data for structural ranking.
7.5.1 PMDDA
Three step workflow: MS1 full scan peak-picking, GlobalStd algorithm to select precursor ions for MS2 from MS1 data and collect the MS2 data and annotation with GNPS(Yu et al. 2022).
7.5.2 HERMES
A molecular-formula-oriented method to target the metabolome(Giné et al. 2021).
7.5.3 dpDDA
Similar work can be found here with inclusion list of differential and preidentified ions (dpDDA)(Y. Zhang et al. 2023).
7.5.4 Extending from MS2 to MSn
A computational approach to generate adatabase of high-resolution-MS n spectra by converting existing low-resolution MSn spectra using complementary high-resolution-MS2 spectra generated by beam-type CAD(Lieng et al. 2023).
7.6 Step 4: score and prioritize candidates with MS/MS
MS/MS evidence is usually the most informative step for moving from formula-level annotation to candidate structure ranking.
MS/MS annotation is performed to generate a matching score with library spectra. The most popular matching algorithm is dot product similarity. A recent study found spectral entropy algorithm outperformed dot product similarity (Li et al. 2021; Li and Fiehn 2023). Comparison of Cosine, Modified Cosine, and Neutral Loss Based Spectrum Alignment showed modified cosine similarity outperformed neutral loss matching and the cosine similarity in all cases. The performance of MS/MS spectrum alignment depends on the location and type of the modification, as well as the chemical compound class of fragmented molecules(Bittremieux et al. 2022). This work proposed a method weighting low-intensity MS/MS ions and m/z frequency for spectral library annotation, which will help annotate unknown spectra(Engler Hart et al. 2024). BLINK enables ultrafast tandem mass spectrometry cosine similarity scoring(Harwood et al. 2023). MS2Query enables the reliable and scalable MS2 mass spectra-based analogue search by machine learning(de Jonge et al. 2023). However, A spectroscopic test suggests that fragment ion structure annotations in MS/MS libraries are frequently incorrect(van Tetering et al. 2024).
Machine learning can also be applied for MS2 annotation(Codrean et al. 2023; Guo et al. 2023; Bilbao et al. 2023). Recent advancements in 2024 and 2025 have seen a surge in deep learning and large language model (LLM) applications. For instance, MSBERT uses mask learning and contrastive learning to embed tandem mass spectra into a chemically rational space(H. Zhang et al. 2024). Graph neural networks and transformers are also being adapted; Graph Transformers have been used for tandem mass spectrum prediction(Young et al. 2024), and graph embedding techniques are enhancing precursor-product ion pair analysis(Zheng et al. 2024). Furthermore, contrastive learning frameworks like CSU-MS2 are improving cross-modal compound identification(Xie et al. 2025), and LLMs are now being empowered to derive spectral embeddings(Xu et al. 2025) and predict collision cross-sections(Zhu et al. 2025).
You could check \[Workflow\] section for popular platform. Here are some stand-alone annotation software:
7.6.1 Matchms
Matchms is an open-source Python package to import, process, clean, and compare mass spectrometry data (MS/MS). It allows to implement and run an easy-to-follow, easy-to-reproduce workflow from raw mass spectra to pre- and post-processed spectral data. Spectral data can be imported from common formats such mzML, mzXML, msp, metabolomics-USI, MGF, or json (e.g. GNPS-syle json files). Matchms then provides filters for metadata cleaning and checking, as well as for basic peak filtering. Finally, matchms was build to import and apply different similarity measures to compare large amounts of spectra. This includes common Cosine scores, but can also easily be extended by custom measures. Example for spectrum similarity measures that were designed to work in matchms are Spec2Vec and MS2DeepScore(Huber et al. 2020).
7.6.2 MetDNA
MetDNA is the Metabolic reaction network-based recursive metabolite annotation for untargeted metabolomics (Shen et al. 2019).
7.6.3 MetFusion
Java based integration of compound identification strategies. You could access the application here (Gerlich and Neumann 2013).
7.6.4 MS2Analyzer
MS2Analyzer could annotate small molecule substructure from accurate tandem mass spectra. (Ma et al. 2014)
7.6.5 MetFrag
MetFrag could be used to make in silico prediction/match of MS/MS data(Ruttkies et al. 2016; Wolf et al. 2010).
7.6.6 CFM-ID
CFM-ID use Metlin’s data to make prediction (Allen et al. 2014) and 4.0 (Allen et al. 2014).
7.6.7 LC-MS2Struct
A machine learning framework for structural annotation of small-molecule data arising from liquid chromatography–tandem mass spectrometry (LC-MS2) measurements.(Bach et al. 2022)
7.6.8 LipidFrag
LipidFrag could be used to make in silico prediction/match of lipid related MS/MS data (Witting et al. 2017).
7.6.9 Lipidmatch
in silico: in silico lipid mass spectrum search (Koelmel et al. 2017).
7.6.10 BarCoding
Bar coding select mass-to-charge regions containing the most informative metabolite fragments and designate them as bins. Then translate each metabolite fragmentation pattern into a binary code by assigning 1’s to bins containing fragments and 0’s to bins without fragments. Such coding annotation could be used for MRM data (Spalding et al. 2016).
7.6.11 iMet
This online application is a network-based computation method for annotation (Aguilar-Mogas et al. 2017).
7.6.12 DNMS2Purifier
XGBoost based MS/MS spectral cleaning tool using intensity ratio fluctuation, appearance rate, and relative intensity(Zhao et al. 2023).
7.6.13 IDSL.CSA
Composite Spectra Analysis for Chemical Annotation of Untargeted Metabolomics Datasets(Baygi et al. 2023).
7.7 Step 5: refine annotation with retention, biology, and prior knowledge
After candidate structures are ranked by MS/MS evidence, orthogonal information can further improve confidence or reject implausible assignments.
7.7.1 Experimental design
Physicochemical Property can be used for annotation with a specific experimental design(Abrahamsson et al. 2023).
7.7.3 ProbMetab
Provides probability ranking to candidate compounds assigned to masses, with the prior assumption of connected sample and additional previous and spectral information modeled by the user. You could find source code here (Silva et al. 2014).
7.7.4 MI-Pack
You could find python software here (Weber and Viant 2010).
7.7.5 MetExpert
MetExpert is an expert system to assist users with limited expertise in informatics to interpret GCMS data for metabolite identification without querying spectral databases (Qiu et al. 2018).
7.7.6 MycompoundID
MycompoundID could be used to search known and unknown metabolites online (L. Li et al. 2013).
7.7.7 MetFamily
Shiny app for MS and MS/MS data annotation (Treutler et al. 2016).
7.7.8 CoA-Blast
For certain group of compounds such as Acyl-CoA, you might build a class level in silico database to annotated compounds with certain structure(Keshet et al. 2022).
7.7.9 KGMN
Knowledge-guided multi-layer network (KGMN) integrates three-layer networks, including knowledge-based metabolic reaction network, knowledge-guided MS/MS similarity network, and global peak correlation network for annotation (Zhou et al. 2022).
7.7.10 CCMN
CCMNs were then constructed using metabolic features shared classes, which facilitated the structure- or class annotation for completely unknown metabolic features(X. Zhang et al. 2024).
7.8 Step 6: search spectral databases
Spectral databases are most useful after candidate space has already been narrowed by formula, feature grouping, and MS/MS evidence.
7.8.1 MS
7.8.2 MS/MS
LibGen can generate high quality spectral libraries of Natural Products for EAD-, UVPD-, and HCD-High Resolution Mass Spectrometers(Kong et al. 2023).
MoNA Platform to collect all other open source database
GNPS use inner correlationship in the data and make network analysis at peaks’ level instead of annotated compounds to annotate the data.
ReSpect: phytochemicals
Metlin is another useful online application for annotation(Guijas et al. 2018).
LipidBlast: in silico prediction
NIST: Not free
GMDB a multistage tandem mass spectral database using a variety of structurally defined glycans.
HMDB is a freely available electronic database containing detailed information about small molecule metabolites found in the human body. HMDB 5.0(Wishart et al. 2022) greatly expanded the database to over 217,000 metabolite entries with enhanced spectral data coverage and improved chemical taxonomy.
KEGG is a collection of small molecules, biopolymers, and other chemical substances that are relevant to biological systems.
7.9 Step 7: search compound databases
Compound databases extend the searchable chemical space beyond spectral libraries, but they should be used with explicit uncertainty because many candidate structures will share the same elemental formula or similar exact mass.
PubChem is an open chemistry database at the National Institutes of Health (NIH).
Chemspider is a free chemical structure database providing fast text and structure search access to over 67 million structures from hundreds of data sources.
ChEBI is a freely available dictionary of molecular entities focused on ‘small’ chemical compounds.
RefMet A Reference list of Metabolite names.
CAS Largest substance database
CompTox compounds, exposure and toxicity database. Here is related data.
T3DB is a unique bioinformatics resource that combines detailed toxin data with comprehensive toxin target information.
FooDB is the world’s largest and most comprehensive resource on food constituents, chemistry and biology.
Phenol explorer is the first comprehensive database on polyphenol content in foods.
Drugbank is a unique bioinformatics and cheminformatics resource that combines detailed drug data with comprehensive drug target information.
LMDB is a freely available electronic database containing detailed information about small molecule metabolites found in different livestock species.
HPV High Production Volume Information System
There are also metabolites atlas for specific domain.
PMhub 1.0: a comprehensive plant metabolome database(Tian et al. 2023)
Atlas of Circadian Metabolism(Dyar et al. 2018)
Plantmat excel library based prediction for plant metabolites(Qiu et al. 2016).