From Target–Nontarget to Nontarget Screening: A Review on Screening Methods for Organic Pollutants Based on High-Resolution Mass Spectrometry

Abstract

Although certain substances exist in the environment at trace concentrations, their ecological risks remain significant, making nontarget screening (NTS) one of the most important topics in the fields of environmental and analytical chemistry today. In recent years, with the development of high-resolution mass spectrometry (HRMS), the higher precision and accuracy have provided mass spectrometry (MS) characteristics with deeper application significance, offering new perspectives for NTS. This review briefly introduces the principles and methodologies of different MS characteristics and summarizes emerging advancements in screening strategies based on these characteristics. We innovatively propose a target–nontarget to nontarget screening workflow, which links different MS characteristics to achieve comprehensive NTS. This workflow prioritizes the currently applied MS characteristics, aiming to integrate various analytical steps to expand the screening coverage while maintaining reliability. The prioritization includes 1) mass defect (MD) analysis, 2) isotope patterns, 3) characteristic fragments, 4) construction of MS characteristics through chemical reactions, and 5) intermolecular connectivity MS characteristics mining. Subsequently, we provided an overview of the mining and prediction of MS characteristics based on machine learning. Finally, we evaluated the challenges of transitioning from target–nontarget to nontarget workflows and proposed future perspectives for NTS.

1. Introduction

With the rapid development of industry, large quantities of chemicals, in particular, small molecular organic compounds, have been released into the environment. These substances subsequently undergo complicated and consistent migration and transformation in a wide variety of environmental matrices, resulting in complex compositions and trace contents of various known and unknown pollutants in the environment. Despite extremely low concentrations of certain substances, their ecological risks remain significant. For instance, approximately 1 μg/L of N-(1,3-dimethylbutyl)-N′-phenyl-p-phenylenediamine-quinone (6PPD-quinone) can induce acute toxicity in coho salmon. Consequently, comprehensive screening of pollutants and evaluating their ecological risks have become critical topics in environmental science and analytical chemistry. This has led to the emergence of the concept of nontarget screening (NTS). A typical screening workflow comprises sample pretreatment, data acquisition, data preprocessing, data analysis and mining, and structural identification. − Meanwhile, it is crucial to design methodologies from a holistic perspective to ensure that the processes are coordinated and highly integrated. For example, a well-designed pretreatment strategy that retains as many compounds as possible is essential, as it fundamentally defines the chemical space available for data acquisition and subsequent analysis.

In recent years, advances in mass spectrometry (MS) technology, especially high-resolution mass spectrometry (HRMS), have provided new opportunities for NTS. Early mass spectrometers had limited mass resolution, which hindered the accurate determination of molecular composition. With advances in analyzers, HRMS has achieved breakthroughs in both resolution and acquisition speed, making it suitable for coupling with chromatographic techniques. According to Hollender et al., a resolution above 20,000 is generally required for HRMS. Currently, the most commonly used HRMS analyzers include time-of-flight (TOF) and Orbitrap, with TOF offering high acquisition rates and Orbitrap providing superior resolution and dynamic range. Furthermore, Fourier transform ion cyclotron resonance (FT-ICR) analyzers provide ultrahigh resolution, capable of reaching above 1,000,000. Overall, the combination of high sensitivity, high mass accuracy, rapid data acquisition, and rich multidimensional structural information makes HRMS the core technology for modern molecular identification. ,, However, along with these advantages comes the challenge of processing and analyzing the vast amount of data generated by HRMS during screening, including MS¹ and MS² information. For general NTS, the ideal goal is a theoretically unbiased analysis that attempts to detect all organic compounds present in a sample without any prior assumptions. However, the vast diversity of environmental chemicals means that such fully unbiased NTS quickly leads to an overwhelming candidate space, increased data-processing burden, and reduced sensitivity for low-abundance features. Therefore, researchers often narrow the coverage of NTS by focusing on specific classes or groups of substances or on compounds already present in existing librariesthe latter corresponding to “suspect screening”this intermediate strategy being what we refer to as “target–nontarget” screening. The target–nontarget screening workflows of per- and polyfluoroalkyl substances (PFAS) and organophosphate esters (OPEs) have been widely used, with results similar to suspect screening. Additionally, top-down techniques like effect-directed analysis (EDA) and protein affinity purification with nontargeted analysis (APNA) provide more specific and effective approaches to screen compounds with potentially specific effects. However, the development of effect assay systems and the acquisition of specific target proteins have limited the broader application of these methods.

In general, target–nontarget screening methods usually emphasize identifying and characterizing the inherent characteristics of substances, such as homologues, isotope patterns, characteristic fragments, and derivatization properties. We recognize that within a screening workflow, the MS characteristics of compounds have varying degrees of priority and are not equally applicable. Inspired by this, we propose that by integrating or constructing MS characteristics of different substances, it is possible to combine multiple independent, explicit target steps (or screenings). This strategy enables the coverage of NTS to expand from narrow to broad, aiming for the most comprehensive coverage possible. In other words, it is a target–nontarget to nontarget screening strategy.

In this review, we recommend that the target–nontarget to nontarget screening should be prioritized in the following order, as shown in Figure : first, searching intrinsic MS characteristics, including 1) mass defect (MD) analysis, 2) isotope patterns, and 3) characteristic fragments, which include characteristic fragment ions (CFIs), characteristic neutral losses (CNLs), and fragment mass differences. Next, if these intrinsic characteristics are not significant, we focus on constructing MS characteristics through chemical reactions, such as 4) chemical derivatization and chemical isotope labeling (CIL) reactions/exposure coordinated with experimental predesigns. Finally, based on preliminary screening results, we search for new chemicals through intermolecular connectivity MS characteristics via 5) molecular networking and paired mass distance (PMD) analysis. Meanwhile, the rapid advancement of machine learning (ML) has enabled the mining and prediction of MS characteristics from an alternative perspective.

1.

Prioritization of MS characteristics in the screening process. The molecular networks section references the work by Meyer et al. on the spectral similarity of mefenamic acid.

2. Intramolecular Intrinsic MS Characteristics Searching

Certain compounds, due to their elemental compositions or molecular structures, inherently exhibit specific MS characteristics, such as MD, isotope patterns, and characteristic fragmentation behavior. Therefore, during data processing, extracting and analyzing these intrinsic features, which require no additional treatment, should be a priority for researchers. MD analysis, as a preliminary MS¹ level filter, primarily aids in identifying the homologous series of compounds. The differences in the natural stable isotope abundances can provide important information about the presence of bromine and chlorine in molecular formulas. Additionally, characteristic fragments generated during in-source fragmentation (ISF) or MS² processes act as structural fingerprints, playing a crucial role in the annotation of unknown compounds.

2.1. Mass Defect

MD is defined as the difference between the exact mass of a compound and its nominal mass. More specifically, we conventionally define 1 Da (dalton) or 1 amu (unified atomic mass unit) as 1/12 of the mass of a ¹²C atom; therefore, any other atom exhibits a MD except ¹²C. Besides these two units, MD can also be expressed using mDa or parts per thousand (ppt).

Polyhalogenated molecules have a unique negative MD, which allows them to be distinguished from other compounds and visualized using MD plots. , For example, PFAS, with more than 8,000 unique known structures, have a wide variety of chemical compositions and homologues, making their identification undoubtedly challenging. Therefore, filtration at the MS¹ level prior to further structural confirmation at the MS² level is crucial, as it helps reduce data complexity and derive chemical formulas and identify possible homologues. Compared to other organic pollutants, PFAS tend to exhibit more significant MDs due to their high halogen content, aiding in preliminary filtering. Typically, when the nominal mass is calculated by rounding, researchers select a MD range of approximately −0.1 to 0.15 Da or −0.15 to 0.15 Da to filter for potential PFAS. This range is not fixed but depends on the composition of the elements. MD is also a commonly used primary filter in the field of drug metabolism. Relative to their parent compounds, phase I and phase II metabolites consistently have a MD within a 50 mDa window. Therefore, applying a MD template can make metabolite signals more distinct. ,

In 1963, Kendrick set CH₂ (14/14.01565) as a scale factor to compute masses for hydrocarbons, which led to the development of the Kendrick MD (KMD) analysis. KMD can be calculated using the following equations:

$$\mathrm{scaled}\mathrm{mass}=\mathrm{exact}\mathrm{mass}(\mathrm{IUPAC})\times \mathrm{scale}\mathrm{factor}$$

$$\mathrm{scaled}\mathrm{MD}=\mathrm{nominal}\mathrm{scaled}\mathrm{mass}(\mathrm{rounded}\mathrm{up})-\mathrm{scaled}\mathrm{mass}$$

Recently, various nontraditional Kendrick mass scale factors have been widely used in environmental analysis field, such as H/Cl (34/33.96102), H/Br (78/77.91051), and CF₂ (50/49.99681). ,,

KMD analysis is a powerful feature recognition method for visualizing homologous series with repeated structural units. For example, researchers used the CF₂ normalized adjusted KMD in the range of 0.85–1.0 or 0–0.15 as the primary filter for PFAS in full scan data, resulting in the discovery of 35 novel PFAS compounds in commercial fluorinated products. KMD is also widely applied to polycyclic aromatic hydrocarbons (PAHs) and their analogues due to the presence of numerous homologues. For instance, the MD scale factors of (H₂/C₂H₂), (H₂/C₂H₄), and C₄H₂ have been used to characterize PAHs. ,

Moreover, the concept of higher-order MD analysis is an extension of the KMD. Simply put, first-order mass transformation, commonly known as KMD analysis, is a renormalization of mass by using a scale factor. Based on first-order mass transformation, the second- or higher-order mass transformations further normalize MDs by introducing an additional scale factor, allowing their subsequent grouping and identification. For example, using Orbitrap MS equipped with an electrospray ionization (ESI) source, higher-order MD plots with three repeating units −C₂H₄O–, −C₂F₄O–, and −CF₂O– reduce the complexity of handling MS data for perfluoropolyether (PFPE) formulations, demonstrating promising application prospects.

2.2. Isotope Patterns

Isotopes are variants of a chemical element with the same number of protons but different numbers of neutrons in their atomic nuclei. Among them, bromine (⁷⁹Br: ⁸¹Br = 50.69:49.31, close to 1:1) and chlorine (³⁵Cl: ³⁷Cl = 75.76:24.24, close to 3:1) often generate significantly characteristic peak patterns in MS, which facilitates researchers in screening compounds containing these elements. Some post-treatment processes, such as disinfection, membrane treatment, or thermal treatment of waste materials, especially in halogen-rich e-waste sites, can lead to complex dehalogenation and halogenation reactions. Therefore, in addition to the parent substances, mixed chlorinated/brominated transformation products (TPs), such as mixed Br/Cl TPs of TBBPA (X-BBPA), polyhalogenated carbazoles, and 2-chloro-4-bromophenol, may also be detected in actual environments. This further enhances the practicality and the applicability of the strategy. For example, in the identification of halogenated disinfection byproducts (DBPs), the isotope patterns of Br and Cl serve as the primary filters in the screening process. As shown in Figure , the characteristic isotope distributions at m/z 256.8414 and m/z 300.7909 demonstrate the presence of BrCl₂ and Br₂Cl, respectively. This was further confirmed by MS² while providing fragmentation information. The two substances were eventually identified as 3-bromo-5,6-dichloro-4-hydroxy-pyran-2-one and 3-chloro-5,6-dibromo-4-hydroxy-pyran-2-one (level 3, according to the five confidence levels proposed by Schymanski et al.).

2.

Relative abundance of Cl and Br isotope distributions. a) Common isotopic distributions (number of Cl and Br atoms ranging from 0 to 2). b) Isotopic distribution of C₅Br₂ClO₃. c) Isotopic distribution of C₅BrCl₂O₃.

Since manually searching for Cl and Br isotope patterns in the total ion chromatogram (TIC) is both cumbersome and time-consuming, several scripts or ML tools have been developed to significantly enhance the efficiency of isotope pattern matching. , Typically, the core functionality of these tools focuses on identifying the isotope peaks at the same retention time (RT) and matching the isotopic intensity ratios. For example, according to the isotope pattern of bromine, if two ion peaks of an ion cluster show a specific mass difference (1.998 Da) and the abundance ratio between the highest peak I _m/z and the peak I _m/z+2 is lower than 100/40, it indicates the presence of a brominated compound. Based on this principle, in our previous study involving hydroponic exposure of tetrabromobisphenol A (TBBPA) to pumpkin seedlings, various glycosyl TBBPAs were identified by filtering the spectra of typical isotope patterns of Br through a R script. For trace compounds in more complex matrices, highly customized and modular algorithms are required. For instance, a computationally efficient isotopic profile deconvoluted chromatogram (IPDC) screening algorithm was developed by combining C/Br/Cl/S isotopic patterns, chromatographic behavior, and ML techniques. This approach successfully detected a wide range of legacy and unknown halogenated contaminants in a Lake Michigan lake trout extract. Recently, a user-friendly software named HaloSeeker 1.0 (updated to version 2.0) has been made freely available for nontarget screening of halogenated compounds in complex environmental media. This software integrates a variety of streamlined tools, including raw file processing, isotope fingerprinting, MD analysis, and formula assignment, making this isotope pattern method easier to grasp and operate. Furthermore, beyond the recognition of the two most characteristic isotope patterns mentioned above (Cl, Br), the isotope patterns of some uncommon elements can also be used to assist in compound annotation, such as ³²S/³⁴S and ¹⁰B/¹¹B. − Software tools and data processing platforms such as Compound Discoverer have integrated the functions for quantitative evaluation of the overall isotope pattern, thereby improving the efficiency of automated annotation.

2.3. Characteristic Fragments

Although MD analysis and isotope pattern recognition serve as preliminary screening tools at the MS¹ level, providing researchers with some insight into the sample, they are not sufficient for definitive identification. Characteristic fragments can offer more robust evidence for the NTS of specific compounds at the molecular structure level. The generation and acquisition efficiency of these fragments directly depend on the MS data acquisition mode and the resulting molecular fragmentation behavior. Therefore, it is necessary to take full advantage of different data acquisition strategies. Besides full scan, data-dependent acquisition (DDA) and data-independent acquisition (DIA) are also commonly used for NTS in HRMS. In DDA, precursor ions are selectively filtered for MS² analysis based on preset conditions such as ion intensity, neutral loss (NL), or isotope patterns. Instead, DIA does not filter precursor ions but applies collision energies (CEs) to all precursor ions within one or more m/z windows, collecting their MS² information indiscriminately. All-ion fragmentation (AIF, MS^E, or MS^ALL) and sequential window acquisition of all theoretical fragment-ion spectra (SWATH) are two commonly used DIA methods. , In some cases, ISF can be useful, as it provides additional MS² information. , However, it is necessary to note that relevant studies have also indicated that ISF is a main cause of the large number of unknown peaks in metabolomics data, thereby leading to an overestimation of the estimated values of biochemical diversity, that is, the so-called “dark metabolome”, and its impact must be taken into consideration.

The common characteristic fragment-based screening workflow involves three steps. First, diagnostic product ion peaks are extracted in MS² mode (or ISF), and their RTs are marked. Next, possible precursor ion candidates are identified based on their exact mass obtained from the full scan mode at the corresponding RTs. Finally, structural identification is performed on these candidates. ,, Here, we summarize the CFIs and CNLs commonly used in screening, as shown in Table . For substances with unknown characteristic fragments, retrospective analysis of their authentic standards can be used to characterize their fragmentation behavior and determine their characteristic fragments.

1. Common Characteristic Fragments (Including CFIs and CNLs) Used in the Screening .

^aThe results of PFAS refer to the summary by Bugsel et al; please note that the structural formulas may not be unique, and isomers may exist.

2.3.1. Halogenated Compounds

2.3.1.1. Organochlorine and Organobromine Compounds

Organic compounds containing Cl and Br have unique in-source fragments in their mass spectrometric response, namely, ³⁵Cl^– and ³⁷Cl^–, ⁷⁹Br^– and ⁸¹Br^–. This method has been successfully applied to screen novel pollutants in polar bear serum using HPLC-ESI-Orbitrap-MS, revealing hundreds of unrecognized halogenated contaminants. However, the abundances of Br^– and Cl^– in the ESI source (the most frequently used source in LC-HRMS) may be much lower compared to the quasi-molecular ion peaks, leading to lower sensitivity when using the unique in-source fragment screening method for these compounds. In contrast, when using a negative ion chemical ionization (NCI) source in GC-HRMS, the unique in-source fragment screening strategy is more sensitive for detecting nonpolar or weakly polar brominated or chlorinated compounds. Applying this strategy, two low abundant single benzene-ring TPs of TBBPA were identified in pumpkin seedlings.

2.3.1.2. Organoiodine Compounds

Compared with C–Br and C–Cl bonds, the C–I bond has a lower bond energy, making it more susceptible to cleavage and resulting in abundant I^– under collision-induced dehalogenation. Thus, I^– (m/z 126.9039) can be considered as a CFI for the identification of organoiodine compounds. An I^–-based screening workflow has been developed for the identification of unknown organoiodine compounds in seaweed, leading to the discovery of 28 potential organoiodine compound peaks at the RTs marked by the detection of specific fragment I^–. Another similar NTS method, named data-independent precursor isolation and characteristic fragment (DIPIC-Frag), was also used to identify organoiodine compounds by screening I^–. However, relying solely on I^– in the negative-ion mode limits the coverage of NTS. To address this, IodoFinder analyzed the fragmentation characteristics of iodinated compounds obtained from ESI+ and ESI– and developed recognition strategies for both modes based on ML. Unlike ESI–, ESI+ is more likely to exhibit characteristic iodine-containing neutral losses (I-NLs) between pairs of fragment ions, including ΔI, ΔHI, ΔCHIN, ΔCHIO, and Δ2I.

2.3.1.3. Organofluorine Compounds

Due to the strong electronegativity of fluorine, the bond energy of the C–F bond is quite high (485 kJ/mol, compared to 240 kJ/mol for the C–I bond), which makes it difficult to cleave under most circumstances. Therefore, CFIs or CNLs typically involve groups such as C _n F2 _n +1^–, C _n F2 _n SO₃ ^–, HF, CF₂, and so on. , This fluorinated fragment-based NTS is commonly used in various matrices, including water, soil, beer, polar bear serum, and cord blood, and can be performed using both ISF and DIA modes. Recently, the ISF characteristics and patterns of PFAS in the ESI source have been systematically characterized, which can help to avoid the misannotation of PFAS during identification. Besides the CFIs and NLs, the mass differences between fragments can also be used to screen unknown PFAS. It is important to distinguish between NLs and fragment mass differences, as the latter may also occur between unrelated fragments that differ by a certain formula but originate from different fragmentation pathways. The algorithm “FindPFΔS” based on Python can help discover and utilize the fragment mass differences of PFAS, thereby further mining the data. After validation and retrospective analysis with precharacterized samples and standards, FindPFΔS successfully detected 94% of a PFAS standard mixture (36 of 38 compounds from 10 compound classes) and identified unknown PFAS homologues in paper extracts. The algorithm shows that many different unknown PFAS fragments can be detected by a single mass difference without prior knowledge of their chemical formulas. The integration of ¹⁹F NMR spectroscopycapable of unbiased detection of all fluorine-containing compounds and providing information on fluorinated moietiesfurther enhances the comprehensiveness and accuracy of HRMS-based PFAS screening. ,

2.3.2. Ester Compounds

Ester groups have relatively high proton affinities, which means that they can accept a proton (H⁺) easily. Once protonated, the charge distribution and electron density within the molecule change, often leading to fragmentation and rearrangement, such as α-cleavage and McLafferty rearrangement. The ions produced by the cleavage of esters are excellent CFIs.

2.3.2.1. Organophosphate Esters

OPEs are widely used as flame retardants and plasticizers and have many potential adverse health effects in humans. , Over the past decade, the number of OPEs discovered in the environment has significantly increased, , largely due to advances in HRMS technology and its associated analytical methods. Research indicates that, in an electron ionization (EI) source, alkyl OPEs undergo three successive McLafferty rearrangements to produce the characteristic fragment H₄O₄P⁺(m/z 98.9842), while phenyl (or benzene-containing) substituted OPEs generate the fragment C₆H₈O₄P⁺ (m/z 175.0155). , Screening methods based on these CFIs have been widely applied to various matrices, such as house dust, sediments, wastewater, foodstuffs, and food packaging. Specifically, for alkyl-, aryl-, thio-, dehydrated, and various other OPE derivatives, studies have also summarized their CFIs and applied them in screening. ,− Additionally, user-friendly software for NTS of OPEs, such as OPEs-ID, can help researchers achieve efficient screening. Recent advances demonstrate that ³¹P NMR has emerged as a powerful complementary technique to MS for semiquantitative nontarget analysis of organophosphorus compounds.

2.3.2.2. Phthalate Esters (PAEs)

EI source, and even the “soft ionization” ESI, is too energetic to preserve the (quasi-)­molecular ions of PAEs. As shown in Figure , the reason is that during ionization, the charge sites on PAE molecules typically reside on the carbonyl oxygen, and the two ortho ester groups on the benzene ring contribute a high electron cloud density. As a result, the charge readily induces the carbonyl oxygen to attack the corresponding ester group (R) or ortho-carbonyl carbon, leading to in-source cleavage and rearrangement.

3.

Fragmentation mechanism of PAEs in positive ionization mode. This figure summarizes the results of refs , , , and .

In positive ionization mode, PAE molecules undergo McLafferty rearrangement, homolytic cleavage, and α-cleavage, resulting in the formation of five-membered cyclic intermediates, such as C₈H₅O₃ ⁺ (m/z 149.0233, common for most PAEs), C₉H₇O₃ ⁺ (m/z 163.0390, specific to dimethyl phthalate (DMP)), and C₈H₇O₄ ⁺ (m/z 167.0339, a potential CFI at relatively low CEs). These intermediates are relatively stable due to conjugation effects and thus exhibit higher abundance signals in the mass spectrum. In negative ionization mode, PAEs with linear, branched, and oxidized (oxo, hydroxyl, and carbonyl) alkyl chains and their TPs at different CEs produce C₇H₅O₂ ^– (m/z 121.0295), along with other ions such as C₈H₃O₃ ^– (m/z 147.0088) and C₈H₅O₄ ^– (m/z 165.0193), which can also be considered as CFIs for PAEs and their TPs. , Due to the good reproducibility of these CFIs, they have been widely used for PAE screening. , For example, using these CFIs, 48 suspect features of plasticizers were found in house dust, and 12 of them were tentatively identified.

2.3.2.3. Parabens

Parabens are widely used as preservatives in food, pharmaceuticals, and cosmetics. They are frequently detected in the environment and have been proven to have adverse effects, such as endocrine disruption. In a study analyzing urine samples from general populations across different age groups in China, a total of 34 paraben-related compounds were detected, including hydroxylated, sulfonated, and sulfated metabolites. In this research, C₇H₄O₃ ^– (m/z = 136.0166), C₆H₄O₂ ^– (m/z = 108.0217), and C₅H₃O₂ ^– (m/z = 95.0139) were designated as CFIs for peak extraction of parabens during NTS.

2.3.2.4. Organo-Sulfate and Sulfonates

Sulfate-like groups (HSO₄ ^–, SO₃ ^–, and HSO₃ ^–) may originate not only from exogenous pollutants that inherently contain these moieties, such as perfluorooctane sulfonate and sodium dodecyl sulfate, but also from their endogenous TPs generated via phase II sulfation. This dual origin enables screening to capture both parent pollutants bearing sulfate groups and their sulfated TPs. For example, one study identified 13 hydrocarbon surfactant ligands of L-FABP and PPARγ using protein APNA, with HSO₄ ^–, SO₃ ^–, and HSO₃ ^– serving as CFIs of sulfate, sulfonate, and sulfosuccinate, respectively. In another study, the use of SO₃ ^– and SO₃Cl^– as CFIs, and SO₂CH₂ and SO₂ as CNLs, enabled the identification of a variety of novel polychlorinated biphenyl (PCB) TPs in polar bear serum samples, such as PCB-sulfates and PCB-sulfonates. Similarly, the detection of CFIs (SO₄H^–, SO₃ ^–, SO₂Cl^–, and HSO₃ ^–) revealed the structures of sulfur-containing chlorinated paraffin analogues in human serum.

2.3.3. Carbocyclic Compounds

2.3.3.1. Polycyclic Aromatic Hydrocarbons

PAHs, as typical carbocyclic pollutants, are highly persistent in the environment and pose significant risks to human health. Their structural diversity, coupled with the fact that PAHs containing fused aromatic rings usually retain an almost intact molecular framework during ionization, makes it challenging to define common CFIs and CNLs for reliable identification. However, this inherent stability can instead be leveraged to distinguish PAHs from other compounds. Accordingly, PAH-Finder has recently been developed as a workflow for identifying PAHs and their derivatives. Instead of relying on specific CFIs, the model exploits PAH-specific fragmentation patterns and integrates ML models for recognition. After normalizing all fragment m/z values to 0–100% of the molecular ion, PAHs typically exhibit dense clusters of peaks in the high-m/z region, with base peak ratios close to 1. In addition, their planar conjugated structures readily form doubly charged ions under the EI, yielding a distinctive 0.5 Da isotope difference. These diagnostic characteristics enable PAH-Finder to achieve substantially lower false-negative and false-positive rates compared with conventional library searches or MD-based filtering.

2.3.3.2. Para-Phenylenediamine Quinones (PPD-Quinones)

PPD-quinones, which primarily originate from the oxidation of antioxidant PPDs, have garnered great attention as emerging pollutants. Although variations in side-chain substitutions lead to differences in properties, these substances have the same core skeleton, making CFI-based screening feasible. PPD-quinones with aryl-substituted side chains generate a CFI of C₁₁H₈NO⁺ (m/z 170.0600) and a CNL of C₁₂H₉NO₂ (199.0633 Da), while those with alkyl-substituted side chains have a CFI of C₆H₇N₂O₂ ⁺ (m/z 139.0502) and a CNL of C₆H₆N₂O₂ (138.0429 Da). These fragments have enabled the identification of six known and three novel PPD-quinones in air particulates, surface soil, and tire tissue.

2.3.4. Nitrogen Heterocyclic Compounds

2.3.4.1. Triazines

Triazines are six-membered heterocyclic compounds containing three nitrogen atoms and are widely applied in agriculture, materials, and pharmaceuticals. A recent study integrated suspect screening, molecular networking, and CFI searching to investigate triazine herbicide TPs in wastewater influent and effluent from treatment plants. A total of 41 TPs were detected, of which 12 were identified using 26 predefined CFIs. , Due to their shared core structures, several of these CFIsC₃H₄N₅ ⁺, C₂H₅N₄ ⁺, C₂H₂N₃ ⁺, and CH₆N₃ ⁺ (m/z 110.0461, 85.0509, 68.0243, and 60.0556, respectively)were also observed in melamine or alkylamine triazine (AAT) compounds. −

2.3.4.2. Triazoles

Triazoles, the largest group of fungicides, are widely detected in the environment. In studies characterizing the degradation byproducts of two triazole pesticides, myclobutanil and penconazole, a common fragment C₂H₄N₃ ⁺ (m/z 70.0400) was observed, which can be used as a CFI for the presence of the 1,2,4-triazole ring. This finding has also been supported by other studies on propiconazole and paclobutrazol. , For benzotriazoles, several studies have shown that most substances containing the benzotriazole substructure share C₆H₆N₃ ⁺ (m/z 120.0556) as a common CFI, as seen in analysis of their biotransformation products. In addition to C₆H₆N₃ ⁺, an “all-in-one” HRMS screening strategy, integrating target, suspect, and NTS, was developed for benzotriazole UV stabilizers (BZT-UVs), which summarized three CFIs: C₁₂H₁₀ON₃ ⁺, C₁₃H₁₀ON₃ ⁺, and C₁₅H₁₄ON₃ ⁺ (m/z 212.0818, 224.0824, and 252.1137, respectively).

A newly published article established a list of CFIs for nitrogen heterocyclic compounds and conducted specialized NTS based on this approach. The authors collected compounds with common structures from PubChem to construct a specific list of CFIs for nitrogen heterocyclic compounds under EI mode, such as pyrrole, pyridine, pyrazine, pyridazine, pyrimidine, indole, quinoline, and isoquinoline. When applied to the NTS in wastewater, this method significantly improved the identification of these compounds.

2.3.5. Antibacterial Agents

The extensive use of antibacterial agents has led to a sharp increase in their concentration in the environment and has also caused adverse effects such as antimicrobial resistance. , Antibacterial agents’ TPs may retain structures with antimicrobial activity and thus continue to contribute to antimicrobial resistance. Therefore, related research not only emphasizes the importance of the parent antibacterial agents in the environment but also includes their TPs. The complex transformation pathways make efficient screening particularly important, with CFI-based methods being among the most widely used.

2.3.5.1. Macrolides

Macrolides are a common type of antibiotic and have been widely detected in the environment. In the Chaobai River in Beijing, four macrolides, including clarithromycin, azithromycin, erythromycin, and roxithromycin, were each associated with more than 10 corresponding TPs in water samples. A total of 12 types of CFIs and CNLs were applied to the NTS of macrolides. By integrating this and other studies, we summarize five commonly observed CFIs in Table .

2.3.5.2. Sulfonamides

In a recently published study, researchers used C₆H₆NO₂S⁺ (m/z 156.0112) and C₆H₆N⁺ (m/z 92.0482) as CFIs, identifying 22 sulfonamides (SAs) and 29 TPs in wastewater contaminated with SAs. Both this and previous studies have demonstrated that although SA transformation pathways are diverse, acetylated products predominate in the TPs profile. Systematic analysis of acetylated SAs further revealed two diagnostic CFIs: C₈H₈NO⁺ (m/z 134.0601) and C₈H₈NO₃S⁺ (m/z 198.0230).

2.3.5.3. β-Lactams

β-lactams (β-Ls) are among the most commonly used antibiotics in medicine, including cephalosporins, penicillins, and β-lactamase inhibitors. C₅H₈NS⁺ (m/z 114.0372) and C₆H₁₀NO₂S⁺ (m/z 160.0427) are commonly used CFIs. ,,,

It is worth noting that while we have provided CFIs, representative of the core structures of major antibiotic classes, compound-specific CFIs are also effective for other antibiotics, such as lincomycin, , quinolone, and quaternary ammonium salts.

3. Construction of MS Characteristics

In addition to the above steps to screen substances with intrinsic characteristics, samples may also contain a large number of compounds without significant homologous features, isotope patterns, and characteristic fragments, such as bisphenols (BPs) and fatty ketones. For such substances, it is essential to construct MS characteristics based on their chemical reactivity or to design experimental strategies that facilitate their screening prior to analysis. Even for substances with the aforementioned MS characteristics, their identifiability during screening processes can be enhanced by further constructing isotope patterns, especially when one focuses on their reactions, transformations, and metabolism.

3.1. Chemical Derivatization

Chemical derivatization is an effective way to improve the analytical performance by changing the structure of the target compounds through chemical reactions. Functional groups such as carboxyl groups (−COOH), hydroxyl groups (−OH), amine groups (−NH₂), ketone groups (CO), aldehyde groups (−CHO), and thiol groups (−SH) are highly reactive and can be easily derivatized. Through reactions with appropriate derivatization reagents, MS characteristics, such as isotopic patterns, CFIs, and mass differences, can be artificially constructed to facilitate compound screening. In this section, we introduce some applications of derivatization in environmental screening, mainly focusing on the construction of MS characteristics.

Chemical isotope labeling (CIL) derivatization reagents offer significant advantages in terms of sensitivity, accuracy, selectivity, stability, analytical throughput, and applicability, improving the chemical detection and quantification of compounds. Isotope-labeled and unlabeled derivatization products exhibit nearly identical properties, ensuring consistent chemical structures and fragmentation behaviors. The isotope forms of the derivatization reagents can be used as an economical internal standard for absolute quantification, helping to bridge target and NTS. , The core of this isotope pattern construction workflow involves separately derivatizing samples with isotope-labeled and nonlabeled derivatization reagents and then mixing them in a specific ratio to generate paired peaks with exact m/z differences. This method has been successfully applied, such as screening aldehydes and ketones using 2,4-dinitrophenylhydrazine (DNPH) and isotope-labeled DNPH (DNPH-D₃) (Δm/z 3.0186), screening amino-containing chemicals by using H/D isotopic formaldehyde (Δm/z 2.0126, 4.0251, 6.0377, and 8.0502), and screening possible toxic DBPs by using light and heavy (¹³C₂,¹⁵N) isotope-labeled GSH (Δm/z 3.0037) as a thiol probe derivatization. , It is noteworthy that the MS² fragments generated from isotope-labeled derivatives may also retain isotope characteristics. In a recent study, glucuronate-conjugated metabolites were identified using N,N-Dimethyl ethylenediamine (DMED-D₀) and its deuterated counterpart DMED-D₆ through an amidation reaction. In addition to the characteristic mass difference (6.0371 Da) filter, two pairs of isotope-tagged CFIs (m/z 247.1294/253.1665 and 229.1188/235.1559) were also employed.

In addition to the aforementioned synthetic isotope derivatization reagents, natural isotope-containing derivatization reagents can also present unique MS fingerprints at lower cost, potentially emerging as a promising derivatization technique. For example, boron (B) has two natural isotopes of ¹⁰B and ¹¹B, which produce a pair of peaks in MS with a mass difference of 0.996 Da and a peak intensity ratio of approximately 1:4 (19.9:80.1). Based on this, researchers developed a derivatization method using the boronate affinity reagent 2-methyl-4-phenylaminomethylphenylboronic acid (2-methyl-4PAMBA). Utilizing LC-HRMS in combination with the B isotope pattern and MS ⁿ fragment characterization of the derivatives, they discovered a total of 45 cis-diol-containing metabolites in rice. Additionally, a bromine-containing derivatization reagent with stereodynamic chiral recognition characteristics, 1-benzoyl-pyrrolidine-2-carboxylic acid 5-bromo-2-formyl-phenyl ester (D-BPBr), has demonstrated potential for NTS due to its advantages in chromatographic separation, qualitative analysis, and quantification of chiral amino acids.

Derivatives can produce CFIs in MS due to the introduction of readily ionizable components. Among them, dansyl chloride (DnsCl) is commonly used as a representative reagent for derivatizing amines and phenolic hydroxyls. A strategy based on DnsCl derivatization coupled with ISF has been successfully applied to the screening of bisphenols in dust and airborne particles. CNLs of 234.0589, 63.9619, and 298.0208 Da correspond to the loss of the dansyl moiety (−C₁₂H₁₂NSO₂•), the elimination of SO₂ by an intramolecular rearrangement, and the consecutive loss of both (−C₁₂H₁₂NSO₂•–SO₂), respectively. Similarly, the CFIs (m/z 139.0212, 155.0161, and 157.0318) after derivatization of p-toluenesulfonylhydrazine (TSH) were also used as a key criterion for screening carbonyl compounds.

After reacting selectively with specific chemical groups, the resulting derivatives exhibit a predictable mass difference compared with their precursors, determined by the derivatization reagent. Therefore, differential analysis before and after derivatization helps to screen target analytes. For instance, O-(2,3,4,5,6-pentafluorobenzyl) hydroxylamine (PFBHA, C₇H₄ONF₅) selectively reacts with carbonyl groups, resulting in a mass difference of 195.0107 Da (C₇H₂NF₅) for a single carbonyl molecule. Based on this principle, a high-throughput screening method was developed combining chemical derivatization with ESI FT-ICR-MS to detect carbonyl molecules in complex natural organic matter (NOM) samples.

However, the derivatization method has some drawbacks. It can increase the analysis time and raise costs, especially when using isotope-labeled derivatization reagents, which are often expensive and not easily accessible. Derivatization may also lead to potential false positives and introduce extra groups, which may result in byproducts. Additionally, some reaction conditions may not be mild enough or biocompatible, and suitable derivatization agents may be lacking for some compounds. Given that environmental or biological samples are often limited in quantity, careful experimental design is required before proceeding with derivatization.

3.2. Construction of Isotope Patterns with Experimental Predesign

Constructing isotope patterns through appropriate pre-experimental design is an important approach initially used in elemental composition analysis in metabolomics. This approach involves using a mixture of unlabeled standards and stable isotope-labeled standards (typically substituted with D, ¹³C, or ¹⁵N) for coexposure or coreaction. Since labeled and unlabeled TPs go through the same transformation pathways, they exhibit the same MS dissociation behaviors and structures, resulting in stable Δm/z and intensity ratios. Unlike derivatization, the isotope tags in this approach actively participate throughout the reaction with the target analytes. This approach has become a promising approach for in-lab experiments aimed at investigating the transformation processes of pollutants, such as PAHs and mono-(2-ethylhexyl) phthalate (MEHP). , Moreover, this method can also provide information about the formation mechanisms of pollutants. In order to credibly track the nitrogenous byproducts and verify the incorporation of nitrogen from NH₄ ⁺ into organic compounds, a stable isotope labeling strategy of ¹⁴NH₄ ⁺ and ¹⁵NH₄ ⁺ was applied and further confirmed that one nitrogen atom in these nitrogenous byproducts originated from NH₄ ⁺. However, it must be noted that stable isotope-labeled standards are not always available for all analytes, and the mass difference may change if the transformation occurs on the labeled atoms.

4. Intermolecular Connectivity MS Characteristic Mining

After completing the initial screening process, we gained a more advanced understanding of the samples. At this stage, it becomes possible to extract more information through molecular connectivity characteristics. These characteristics are invaluable for refining the screening, filling the gaps, and discovering unknown substances such as analogues and TPs. This signifies the transition of NTS from discovering the unknowns to uncovering additional unknowns based on the established knowledge. Furthermore, these characteristics aid in describing the samples and analyzing the intrinsic connections from a holistic perspective.

4.1. Molecular Networking

Chemicals with similar structures may exhibit similar MS behavior and, therefore, share some fragment ions. Molecular networking links precursor ions with similar MS² spectra and clusters them into subclusters. Since molecules within the same subcluster have structural correlations, the presence of a known structure can facilitate the rapid identification of unknown structures. Therefore, based on the preliminary screening results, molecular networking can be used to discover new molecules, especially analogues of known compounds. Several open-access knowledge bases, such as the Global Natural Product Social Molecular Networking (GNPS), provide community-driven platforms that facilitate molecular network analysis.

In recent years, NTS strategies based on molecular networks have been gradually developed, especially in the fields of new pollutant screening and drug transformation. , For instance, after peak extraction, researchers created a molecular network in GNPS and obtained seed nodes by matching MS information (MS¹ and MS² data) with public databases, in silico annotation, and automated chemical structure classification. Subsequent visualization and analysis of mass differences between known and unknown nodes allow further characterization of unknown TPs. Employing this approach, the study identified 52 antimicrobial compounds and 49 TPs, 30 of which had not been previously reported in the environment. This potential structural correlation between nodes is an important driving factor in substance identification. In a separate study, a small set of known MS² spectra served as seeds within GNPS for iterative annotation of structurally related compounds, enabling high-throughput discovery of emerging pollutants and their TPs without relying on an extensive reference library. Currently, the development of several new grouping and annotation tools has enabled molecular networking strategies to overcome bottlenecks in different ways. One example is Ion Identity Molecular Networking (IIMN), which integrates chromatographic peak shape correlation analysis into molecular networks, thereby overcoming the unnecessary separation of molecular subnetworks due to the diverse fragmentation behaviors of molecules. Based on the hypothesis that natural products group closely around specific scaffolds in chemical space and molecular formula distributions are diagnostic for compound families, an innovative annotation tool, Structural Similarity Network Annotation Platform for Mass Spectrometry (SNAP-MS), was developed. This tool matches chemical similarity groupings in microbial natural product libraries with molecular networking subnetworks.

4.2. Paired Mass Distance Analysis

Paired mass distance (PMD) refers to the exact mass difference between the two ions. Due to the high precision of HRMS data, PMD can be endowed with statistical and analytical significance. PMD has direct applications in understanding MS behavior and suspect screening. First, PMD can identify adducts formed via complex in-source reactions. For example, a PMD of 21.9820 Da may indicate a mass distance between [M + H]⁺ and [M + Na]⁺, while a PMD of 15.9740 Da may indicate a mass distance between [M + K]⁺ and [M + Na]⁺. Second, PMD can recognize isotopic peaks, mining isotope information and patterns from the mass differences. Existing studies have used a PMD of 1.998 Da to filter potential Cl/Br-halogenated organic compound features. Third, PMD can be representative of a specific reaction unit. By integrating the reaction units with changes in functional groups or atoms corresponding to relevant biochemical reactions, potential TPs can be discovered based on known compounds. For instance, a PMD of 162.0528 Da (+C₆H₁₂O₆/–H₂O) implies that a glycosylation reaction may have occurred. This concept has been successfully applied to the screening of short-chain and medium-chain chlorinated paraffin metabolites in exposed rice cell systems. PMD analysis identified a total of 65 TPs formed through (multi)­hydroxylation, dichlorination, sulfation, and glycosylation reactions.

In addition to the aforementioned direct applications, PMD analysis can help achieve several crucial functions. To begin with, PMD analysis can be extended to the concept of PMD-based reactomics, which focuses on “reactions” rather than “compounds” themselves. It can be used to distinguish between endogenous metabolic processes and other transformation processes. , High-frequency PMDs identified from the Human Metabolome Database (HMDB) and the Kyoto Encyclopedia of Genes and Genomes (KEGG) often suggest associations with endogenous metabolic pathways, indicating an endogenous origin of substances or the involvement of endogenous metabolic processes in exogenous substances.

Subsequently, PMD analysis can be extended to the quantification of reactions, providing a macroscopic perspective for analyzing the transformation processes of compounds. One study explored the reactivity and transformation of NOM during UV photolysis, which helps illustrate how this approach can be applied in practice. Possible photodriven NOM transformation processes were examined by quantifying the reactions based on predefined PMDs, and the results showed that hydroxylation was the major reaction for lignin/CRAMs and condensed aromatics, involving 28.7% and 23.6% of related formulas, respectively. The loss of a nitro group in tannins was associated with 37.1% formulas.

Additionally, PMD analysis provides an effective method for characterizing the associations between molecules. It focuses not only on mass differences but also on structural changes. PMDs can be considered as associations (i.e., edges) among reactants (i.e., nodes). A recent study proposed a Simplified Network Analysis (SNA) method to reveal the complex transformation pathways of emerging contaminants in wastewater treatment processes. SNA defines and classifies nodes through NTS, and then, it uses PMD analysis to establish connections between the nodes and further analyze structural changes and transformation characteristics. Another study developed a strategy named directed paired mass distance (dPMD). Unlike the general PMD network that relies on reaction frequency analysis, dPMD integrates the actual relationships between samples (such as the sequential relationship between influent and effluent samples in a wastewater treatment plant) to infer the reaction direction. This method allows for a clearer understanding of reaction pathways and microbial interactions, thereby optimizing the control and treatment of dissolved organic matter in wastewater processes.

5. Mining and Prediction of MS Characteristics Based on Machine Learning

Although we previously discussed the MS characteristics of various small molecule compounds and their crucial role in NTS annotation, it must be acknowledged that accurate annotation still largely relies on laborious expert evaluation, which heavily depends on sample information as well as the experience and even intuition of scientists. This situation naturally led to the explosive growth in the field of ML over the past decade, bringing us to a revolutionary crossroads.

ML algorithms mainly consist of three types: supervised learning (such as decision trees, support vector machines (SVMs), and logistic regression), unsupervised learning (such as K-means clustering and principal component analysis), and semisupervised learning. Deep neural networks (DNNs), as a subfield of ML, typically contain dozens or even hundreds of hidden layers to achieve superior performance. Common network structures include recurrent neural networks (RNNs), convolutional neural networks (CNNs), transformers, and generative adversarial networks (GANs), among others. Additionally, a variety of advanced strategies and learning paradigms have emerged, integrating different learning modes to optimize and enhance the performance of the overall process.

Based on key MS¹ characteristics, such as isotope patterns, ML tools can assist in molecular formula assignment and peak extraction. ChloroDBPFinder is trained using molecular formulas from PubChem and combines a random forest model to establish a hierarchical prediction framework. This tool enables the identification of chlorinated disinfection byproducts, achieving an accuracy of 93.3% in recognizing Cl-containing molecular formulas and 92.9% in annotating the number of Cl atoms. LipoCLEAN identifies erroneous identifications by analyzing the differences in isotope distribution and integrates chromatographic behaviors, such as RT, thereby significantly enhancing the confidence and accuracy of the lipid identification results.

ML approaches for the annotation and identification of small molecules based on MS² follow three principal paradigms: spectrum-to-structure prediction, structure-to-spectrum prediction, and de novo methods. ,,, As shown in Figure , the core idea of spectrum-to-structure prediction is not direct spectral matching, but rather first translating mass spectra into molecular fingerprint codes, which are then used to search compound databases. For example, CSI:FingerID and MetFID belong to this type of tool. There are also some highly integrated specialized platforms, such as the PFAS automatic identification platform APP-ID. It consists of two modules: the PFAS_link module, designed to extract the PFAS characteristic peaks from environmental samples and build a PFAS-enhanced molecular network, and the PFAS_ID module, which employs a SVM model to predict structure fingerprints and match them based on similarity. As mentioned earlier, both PAH-Finder and IodoFinder identify structure fingerprints by leveraging characteristic fragmentation patterns, offering greater interpretability in terms of chemical mechanisms.

4.

Principles and common tools of two machine learning strategies: from structure to spectrum and from spectrum to structure. The classification of tools is based on ref ; IOKR refers to ref .

Structure-to-spectrum prediction refers to the process of generating theoretical mass spectra for compounds with known structures. This is achieved using computational chemistry and theoretical prediction models to simulate the mass spectral fragmentation process. The theoretical mass spectra are then compared with a large number of experimental mass spectra of unknown substances to achieve annotation. , A classic tool is CFM-ID, developed for the annotation of ESI-MS² spectra for a given compound structure. ,

De novo method does not rely on existing databases but directly generates reasonable compound structures from mass spectra. This method helps to discover “unknown unknowns”new substances that are not included in the existing databases. The de novo method represents the most cutting-edge direction, as exemplified by MSNovelist, which predicts molecular fingerprints based on CSI:FingerID and then converts these fingerprints into SMILES sequences through RNN to generate candidate molecular structures. , Recently, CSU-MS² (Contrastively Spectral-Structural Unification framework for MS/MS spectra and molecular structures), as its name suggests, maps both the MS² spectra and structures into a unified embedding space for direct cross-modal retrieval, facilitating a novel shift away from the traditional one-way structure-spectrum paradigm.

6. Summary and Future Perspectives

The realization of “true” NTS is a huge project that requires the integration of various technologies, methods, and even disciplines. It involves several critical steps, including sample collection, pretreatment, data acquisition and analysis, all of which must be highly interconnected. In this review, we propose a tentative holistic strategy, namely, target–nontarget to nontarget screening, to provide a relatively clear prioritization of screening work under current conditions. We classify common low molecular weight organic compounds into two categories: those with intrinsic MS characteristics and those without. For compounds with intrinsic MS characteristics, the workflow emphasizes extracting the inherent MS characteristics first, focusing on MD analysis and other information at the MS¹ level, such as natural isotope patterns (namely Cl- or Br-containing substances). At the MS² level, we summarize common characteristic fragments used in environmental analyses and their analytical methods. For compounds without intrinsic MS characteristics, we suggest that such characteristics can be constructed through chemical derivatization/reaction. Integrating experimental design into the screening process is crucial, especially for in-lab-simulated reaction systems. Finally, based on the intermolecular connectivity MS characteristics, we aim to refine the screening results, further describe the samples, and analyze the intrinsic connections from a more advanced holistic perspective.

However, applying this workflow to real environmental samples presents several challenges. First, under current conditions, a highly integrated, cross-platform, multimethod universal screening protocol has not yet been fully established. Although our prioritization strategy aims to achieve comprehensive screening, the complexity of the samples may actually lead to overlapping processes and results.

Alternatively, RT and collisional cross section (CCS) are critical parameters in separation techniques that provide an additional dimension of information. When combined with MS data, they increase the confidence of the annotation. Together with MS¹ and MS², they enable four-dimensional identification of compounds. However, RT is highly dependent on specific experimental conditions (mobile phase, column type, etc.), which poses a significant challenge for cross-laboratory comparison. In contrast, CCS is typically more reproducible across instruments and laboratories. Nevertheless, the limited coverage of CCS databases still restricts its application in NTS. It is essential to further improve the coverage and universality of the database, as well as the reliability of ML-based prediction methods to address this limitation.

Considering the interdependencies between modular steps, balancing throughput and reliability to form streamlined, scalable, and harmonized workflows to expand the analysis coverage will be a key trend for future exploration. Additionally, today, NTS usually stops after spectral annotation, followed by target analysis. However, due to the difficulty and high cost of obtaining authentic standards, it is impractical to perform exhaustive analysis on all possible compounds. Semiquantitative analysis may be the long-term direction for bridging target analysis and NTS in the future.

Second, the implementation of the characteristic fragment strategy still heavily depends on the researchers’ expertise and experience. Although we have summarized some known MS characteristic fragments in this review, it is impossible to achieve a comprehensive coverage of all substances. The rapid advancement of ML has enabled the mining and prediction of MS characteristics from an alternative perspective, which may become a future trend in analysis. However, the predictive accuracy of ML is still insufficient at present. The high-quality data available for training are very limited compared to the vast number of chemical substances. We need higher-performance computational tools and techniques to help us discover and predict potential CFIs, identify structural fingerprints of analytes, and provide prioritized targets. Recently, several ML-assisted methods have been developed to predict bioactive substances, replacing physical fractionation and achieving “virtual fractionation”, thereby guiding NTS and reducing false positives caused by common nonspecific groups. , Nevertheless, it must be acknowledged that the black-box nature of ML gives rise to issues of interpretability, which, in turn, leads to corresponding risks and a lack of trust. Further development and integration are still needed in terms of interpretability and workflow design.

Third, despite the availability of numerous derivatization reagents, achieving an unbiased derivatization remains challenging. The reactivity of the same functional groups can vary in different structures, potentially leading to the oversight of less reactive but environmentally significant substances during the derivatization process. For example, aryl ketones, which are less reactive than other ketone substances, may be neglected. This highlights the need for advancements in derivatization procedures and the development of more refined derivatization reagents.

The authors declare no competing financial interest.

Figure 1 was replaced after this paper was published ASAP on January 9, 2026. The corrected version was reposted on February 6, 2026.