Beyond the Target: Establishing Chemical Confidence Levels in Non-Target Analysis for Drug Development

Abstract

This article provides a comprehensive framework for establishing robust chemical confidence levels in non-target analysis (NTA), a critical methodology for identifying unknown or unexpected chemicals in drug development. Aimed at researchers and scientists, we explore the foundational principles of NTA confidence-level assignments and detail advanced workflows that integrate high-resolution mass spectrometry (HRMS) with machine learning (ML). The content covers practical strategies for optimizing NTA workflows, troubleshooting common challenges like data complexity and identification bottlenecks, and validating findings through multi-tiered approaches. By synthesizing current methodologies and future-looking trends, this guide empowers professionals to enhance the reliability and regulatory readiness of their non-targeted screening data.

The Five-Tiered Foundation: Understanding Confidence Level Assignments in NTA

In the field of non-targeted analysis, the confidence level framework provides a standardized system for evaluating the certainty of compound identification. This hierarchical classification spans from Level 1 (confirmed structure) to Level 5 (unknown), creating a systematic approach for researchers to communicate the reliability of their identifications. The framework was formally established following the 2017 Metabolomics Society meeting in Brisbane, which redefined metabolite identification credibility standards and added "Level 0" to the classification system [1]. This standardization is particularly crucial in applications such as clinical diagnostics, environmental monitoring, and drug development, where erroneous identifications can significantly impact research conclusions and subsequent decisions.

The fundamental challenge in non-target analysis lies in distinguishing true compound identifications from false positives, especially when dealing with complex biological samples containing thousands of metabolites at varying concentrations. Without a standardized confidence framework, cross-study comparisons become problematic, and the accumulation of misidentified compounds in databases can perpetuate errors in the scientific literature. The confidence level system addresses this by establishing clear, evidence-based criteria for each classification level, enabling researchers to properly assess and report their identification confidence.

The Five-Tiered Confidence Framework

Detailed Classification Criteria

The confidence level framework creates a transparent system for evaluating compound identifications based on the type and quality of analytical evidence available.

The five-tiered confidence framework illustrated above represents a hierarchical system where evidence quality decreases from Level 1 to Level 5. At the highest confidence level (Level 1), identifications require confirmation against a reference standard analyzed under identical experimental conditions, with matching of two orthogonal properties such as retention time and MS/MS spectrum [1] [2]. This level provides near-certain structural confirmation and is essential for definitive biomarker identification.

Level 2 identifications provide probable structure through library spectrum matching, typically using high-resolution MS/MS data compared to reference spectra in databases such as HMDB or MassBank, but without retention time confirmation [1]. Level 3 represents tentative candidates based on diagnostic evidence such as characteristic fragmentation patterns or spectral similarity to compounds within the same chemical class. At Level 4, only the molecular formula can be confidently assigned through accurate mass measurement, while Level 5 includes unidentified exact mass signals with no structural information available [2].

Analytical Evidence Supporting Each Level

Different types of analytical evidence contribute to the confidence level assignment, with each level requiring progressively more rigorous data.

Table: Analytical Evidence Requirements for Confidence Levels

Confidence Level	MS/MS Spectrum	Retention Time	Accurate Mass	Reference Standard	Ion Mobility (CCS)
Level 1	Required (matched)	Required (matched)	Required	Required (authentic)	Optional (increasingly used)
Level 2	Required (matched)	Not required	Required	Not required	Optional (supports level)
Level 3	Characteristic fragments	Not required	Required	Not required	Not typically available
Level 4	Not required	Not required	Required	Not required	Not typically available
Level 5	Not available	Not available	Detected	Not required	Not typically available

The table above demonstrates how each confidence level builds upon specific analytical evidence. For the highest confidence identifications, the incorporation of collision cross-section (CCS) values from ion mobility spectrometry provides an additional orthogonal parameter that can significantly increase confidence [3]. Modern instruments like the timsTOF Pro 2 enable measurement of CCS values, which serve as an additional molecular descriptor that is independent of mass and retention time. When CCS values match those of reference standards, they can provide supporting evidence that may elevate confidence levels.

Experimental Protocols for Confidence Level Assignment

Sample Preparation and LC-MS Analysis

Robust sample preparation is fundamental to achieving reliable identifications across all confidence levels. For serum non-target analysis, a typical protocol involves protein precipitation using cold methanol [4]. Specifically, 100 μL of serum is combined with 370 μL of cold methanol, followed by vortexing and incubation at -80°C for 30 minutes to precipitate proteins. The sample is then centrifuged at 3,200 × g for 30 minutes at 4°C, with the supernatant transferred to a new vial for analysis [4]. For large-scale studies, this process can be automated using 96-well plate formats with phospholipid removal plates to enhance throughput and reduce matrix effects [2].

Liquid chromatography separation typically utilizes reversed-phase columns with either water-acetonitrile or water-methanol mobile phase systems. Two alternative gradient approaches are commonly employed: Gradient A (biased toward non-polar molecules) and Gradient B (providing better coverage of medium-polarity molecules) [4]. For comprehensive coverage, many laboratories employ both reversed-phase and hydrophilic interaction liquid chromatography (HILIC) methods to capture the broad chemical diversity present in biological samples.

Mass spectrometric analysis for confidence level assignment requires high-resolution instruments capable of accurate mass measurement and MS/MS fragmentation. Data-dependent acquisition (DDA) methods typically involve full-scan MS1 spectra (e.g., 50-1,200 m/z range) followed by isolation and fragmentation of the most intense ions. Key instrument parameters include collision energy (typically 6-35 eV, sometimes ramped), capillary voltage (2,200 V for positive mode), and source temperature (150°C) [4]. The inclusion of quality control samples—including pooled quality control (QC) samples and dilution QC (dQC) samples—throughout the analytical sequence is essential for monitoring instrument stability and assessing quantitative accuracy [1].

Data Processing and Compound Identification

Following data acquisition, raw LC-MS files undergo preprocessing including peak detection, retention time alignment, and intensity normalization. Open-source tools like XCMS are commonly used with parameters such as full width at half maximum (FWHM) = 8 seconds for peak detection [4]. For metabolite annotation, the PerformDetailMatch() function in MetaboAnalystR enables database matching with user-defined mass tolerance (e.g., 5 ppm) and supports HMDB, KEGG, and other major metabolite databases [5].

For Level 1 identifications, the workflow requires comparison to authentic reference standards analyzed under identical experimental conditions. The RegisterData() function in MetaboAnalystR can facilitate seamless integration of multiple batches of LC-MS data, which is particularly important for longitudinal studies where reference standards may be analyzed across different sequences [5]. Level 2 identifications rely on MS/MS spectral matching to reference libraries, with tools like MetaboAnalystR incorporating scoring algorithms that consider both spectral similarity and retention time prediction [5].

Advanced computational approaches are increasingly being employed to enhance confidence level assignments. The Denoising Search algorithm, for example, removes both electronic and chemical noise from MS/MS spectra, significantly improving spectral matching quality [6]. When tested on 240 metabolites, this approach reduced the required injection amount by 35-fold while maintaining identification confidence, demonstrating particular value for low-abundance compounds where spectral quality is often compromised [6].

Technology Comparisons for Confidence Level Attainment

Instrumentation Platforms and Their Capabilities

Different mass spectrometry platforms offer varying capabilities for achieving specific confidence levels in non-target analysis. The comparison below highlights how instrument selection impacts the potential confidence levels attainable.

Table: Mass Spectrometry Platform Comparison for Confidence Level Attainment

Instrument Platform	Best Suited Confidence Levels	Key Strengths	Typical Annotation Rates	Limitations
Orbitrap Astral	Levels 1-3	Ultra-high sensitivity; Rapid MS/MS acquisition	2.5× increase vs. Exploris 240 [6]	Higher cost; Complex data handling
timsTOF Pro 2	Levels 1-3	CCS measurement; dia-PASEF technology	>70% MS2 coverage [3]	Requires specialized databases with CCS values
Triple Quadrupole	Level 1 (targeted)	High sensitivity; Excellent quantitative performance	N/A (targeted only)	Limited to pre-defined compounds
Q-TOF	Levels 2-4	Good balance of resolution and sensitivity	Variable (database dependent)	Lower fragmentation efficiency vs. Astral

The table demonstrates how platform selection directly impacts the depth and confidence of compound identifications. The Orbitrap Astral platform achieves significantly improved annotation rates through a combination of exceptional sensitivity (low attomole range) and rapid MS/MS acquisition, enabling more high-quality spectra for confident identifications [6]. The timsTOF Pro 2 contributes to confidence level assignment through the addition of collision cross-section (CCS) values as a fourth dimension of separation, which helps distinguish isobaric compounds and reduces false identifications [3].

Data Analysis Tools and Their Impact

Software tools play a critical role in confidence level assignment, with different platforms offering varying approaches to data processing and metabolite annotation.

MetaboAnalystR provides a comprehensive open-source solution for non-target analysis, incorporating automated peak profiling, metabolite annotation, and functional interpretation [5]. The tool utilizes a hybrid architecture with Rcpp-based C++ acceleration that provides 10-50× faster processing compared to pure R implementations [5]. For statistical analysis, it integrates established packages like pcaMethods and limma, ensuring analytical reliability while maintaining an accessible interface for non-bioinformatics experts.

Enhanced Structure-Guided Molecular Networking (E-SGMN) represents an advanced approach that leverages the high-speed MS/MS capabilities of instruments like the Orbitrap Astral [6]. This method organizes metabolites into molecular families based on spectral similarity and fragmentation patterns, enabling the transfer of annotation confidence within chemical classes. When combined with denoising algorithms, this approach can significantly increase annotation coverage while maintaining confidence levels.

TASQ Software combined with timsTOF instrumentation enables the incorporation of ion mobility filtration, which dramatically improves selectivity in complex matrices [3]. The mobility filtering window removes chromatographic and spectral interferences, as demonstrated in the analysis of thiacloprid in onion matrix at 1 ng/mL concentration, where background interferences were effectively eliminated, resulting in a perfect database match [3].

The Scientist's Toolkit: Essential Research Reagents and Materials

Successful confidence level assignment requires not only sophisticated instrumentation but also carefully selected reagents and reference materials. The following toolkit outlines essential components for establishing confident identifications in non-target analysis.

Table: Essential Research Reagent Solutions for Confidence Level Assignment

Reagent/Material	Function	Application Example	Impact on Confidence Level
Authentic Reference Standards	Retention time and spectrum matching	Level 1 confirmation	Enables highest confidence (Level 1)
Stable Isotope-Labeled Internal Standards	Retention time alignment; Quantitative correction	dQC sample preparation	Improves quantitative accuracy across all levels
Quality Control Materials	Monitor instrument performance; Assess technical variation	Pooled QC samples; dilution QC (dQC)	Essential for validating all confidence levels
Database Subscriptions	MS/MS spectrum matching	HMDB, MassBank, NIST	Critical for Level 2-3 assignments
Specialized Solid-Phase Extraction Plates	Matrix removal; Sample cleanup	96-well phospholipid removal plates	Reduces interferences, improves spectral quality
Chromatography Optimization Kits	Column and mobile phase selection	Retention time predictor development	Supports Level 2 with retention time prediction

The reagents and materials highlighted above each contribute specific functions that collectively enable confident compound identification. Authentic reference standards are particularly crucial as they represent the only pathway to Level 1 confirmations [1] [2]. The emerging practice of incorporating dilution QC (dQC) samples addresses a critical gap in non-target analysis by distinguishing technical variation from true biological differences, thereby increasing confidence in quantitative differences observed for putatively identified compounds [1].

For laboratories focusing on specific chemical domains, customized spectral libraries provide significant advantages over general databases. As emphasized in contemporary practice, "self-built libraries prioritize accuracy over size," recognizing that excessively large databases can complicate data interpretation and lead to resource waste during subsequent validation phases [1]. This approach is particularly valuable for targeted applications such as pharmaceutical impurity profiling or environmental contaminant screening.

The framework for confidence levels in non-target analysis provides an essential foundation for generating reliable, reproducible data across diverse application domains. As analytical technologies continue to evolve, with platforms like the Orbitrap Astral and timsTOF Pro 2 offering unprecedented sensitivity and spectral acquisition rates, the potential for achieving higher confidence levels for more compounds continues to expand [3] [6]. However, technology alone cannot ensure confidence—rigorous experimental design, appropriate quality control procedures, and transparent reporting of confidence levels remain fundamental to generating scientifically valid results.

The future of confidence level assignment will likely see increased integration of computational approaches, such as the Denoising Search algorithm and Enhanced Structure-Guided Molecular Networking, which enhance the quality and quantity of conf identifications without requiring additional instrumental analysis [6]. Additionally, the incorporation of orthogonal parameters such as collision cross-section values provides an independent measure for confirming identifications and represents a significant advancement in confidence level assignment [3]. As these technologies and methods mature, the field moves closer to comprehensive and confident characterization of complex mixtures, enabling deeper biological insights and more confident decision-making in applied settings.

The Critical Role of High-Resolution Mass Spectrometry (HRMS) in NTA

Non-Targeted Analysis (NTA) represents a paradigm shift in chemical analytical approaches, moving beyond the limitations of traditional targeted methods that search for small, pre-defined sets of chemicals. Instead, NTA employs sophisticated analytical techniques to simultaneously detect, identify, and potentially quantify thousands of unknown chemicals present in complex samples without prior knowledge of their identity [7]. This capability is particularly crucial given the thousands of chemicals in commerce and the environment for which little to no exposure data exists. At the heart of this transformative approach lies High-Resolution Mass Spectrometry (HRMS), which provides the foundational analytical power necessary to explore this vast, unknown chemical space [7].

The fundamental challenge that NTA addresses is the limitation of conventional environmental monitoring strategies, which predominantly rely on targeted chemical analysis and inherently overlook a wide range of known "unknowns" including transformation products and emerging contaminants [8]. As the rapid proliferation of synthetic chemicals continues to lead to widespread environmental pollution through diverse sources such as industrial effluents, household personal care products, and agricultural runoff, the need for comprehensive analytical approaches becomes increasingly urgent [8]. HRMS-enabled NTA provides researchers and regulatory agencies with the capability to identify and monitor compounds of emerging concern that are difficult to measure with traditional methods, thereby supplying decision-makers with critical information to better assess and manage potential risks [7].

The integration of chromatography with HRMS platforms, including quadrupole time-of-flight (Q-TOF) and Orbitrap systems, generates the complex datasets essential for NTA [8]. These instruments resolve isotopic patterns, fragmentation signatures, and structural features necessary for compound annotation through post-acquisition processing involving centroiding, extracted ion chromatogram (EIC/XIC) analysis, peak detection, alignment, and componentization to group related spectral features into molecular entities [8]. This technological foundation enables researchers to find chemicals that have never been reported or noticed, opening new frontiers in environmental chemistry, exposomics, and public health protection.

HRMS as the Analytical Foundation for Confidence Level Assignment

The Confidence Scale for Compound Identification

The assignment of confidence levels for compound identification represents a critical framework in NTA, providing a standardized approach to communicate the certainty of chemical annotations. This hierarchical system enables researchers to distinguish between tentatively and conclusively identified compounds, with HRMS data serving as the fundamental source of evidence across all levels. The confidence scale typically ranges from Level 1 (confirmed structure) to Level 5 (exact mass of interest), with each successive level incorporating additional analytical evidence derived from HRMS measurements [9] [10].

At the most confident Level 1 identification, the analytical evidence must include matching retention time and fragmentation spectrum with an authentic standard measured under identical analytical conditions—a process fundamentally reliant on HRMS platforms [10]. Level 2 identification (probable structure) demonstrates the critical importance of high-resolution tandem mass spectrometry, requiring confirmation through library spectrum match or diagnostic evidence from fragmentation spectra [10]. Level 3 (tentative candidate) relies primarily on HRMS data through spectral library matches or in silico fragmentation predictions, while Level 4 (unequivocal molecular formula) depends exclusively on the accurate mass measurement capabilities unique to HRMS instruments [10]. This structured approach to confidence assignment ensures transparency in reporting and helps stakeholders understand the limitations and certainty associated with chemical identifications in NTA studies.

HRMS Capabilities for Confidence Level Assignment

The superior capabilities of HRMS platforms directly enable the ascending levels of confidence in chemical identification. Orbitrap and Q-TOF systems provide the high mass accuracy (<5 ppm) and resolution (>20,000) necessary for confident molecular formula assignment (Level 4), with modern instruments routinely achieving 1-2 ppm mass accuracy and resolutions exceeding 50,000-100,000 [8] [10]. This exact mass measurement capability allows researchers to exclude many potential candidate structures during the identification process, significantly narrowing the chemical search space.

For higher confidence levels (2 and 1), HRMS systems provide the tandem mass spectrometry (MS/MS) data essential for structural elucidation. The combination of accurate precursor mass selection and high-resolution fragmentation spectra enables researchers to discern between structurally similar compounds and establish diagnostic fragmentation patterns [10]. When coupled with chromatographic separation, HRMS systems further support confidence through retention time matching, with advanced approaches incorporating predicted retention time indices (RTIs) based on quantitative structure-retention relationship (QSRR) models to provide additional orthogonal evidence for compound identification [11]. The integration of ion mobility spectrometry (IMS) with HRMS adds yet another dimension of separation through collision cross-section (CCS) values, which can be predicted from chemical structure and used as additional confirmatory evidence [10].

Table 1: HRMS Instrument Capabilities Supporting Confidence Level Assignment

Confidence Level	Identification Type	Key HRMS Data Requirements	Typical HRMS Performance Metrics
Level 1	Confirmed structure	Retention time match + MS/MS spectrum match to reference standard	Mass accuracy < 2 ppm, MS/MS spectral library match score > 0.8
Level 2	Probable structure	Characteristic MS/MS fragments or library spectrum match	Mass accuracy < 5 ppm, MS/MS spectral library match score > 0.7
Level 3	Tentative candidate	Spectral similarity to known compounds or class-specific fragments	Mass accuracy < 5 ppm, diagnostic fragment ions present
Level 4	Unequivocal molecular formula	Accurate mass measurement for molecular formula assignment	Mass accuracy < 1-5 ppm, isotope pattern match (RMSD < 20)
Level 5	Exact mass of interest	Accurate mass measurement only	Mass accuracy < 5 ppm, detected in sample but not blanks

Experimental Design and Workflow for HRMS in NTA

Comprehensive NTA Workflow with HRMS

The implementation of HRMS within NTA follows a systematic, multi-stage workflow that transforms raw samples into chemically actionable information. This comprehensive process integrates sample preparation, instrumental analysis, and sophisticated data processing to address the unique challenges of non-targeted chemical discovery [8].

Diagram 1: Comprehensive HRMS-NTA Workflow. The workflow progresses through four critical stages from sample preparation to validation, with HRMS central to data generation and processing stages.

Sample Preparation and Data Acquisition Protocols

The initial stages of the NTA workflow focus on preparing samples for HRMS analysis and generating high-quality data. Sample preparation requires careful optimization to balance selectivity and sensitivity, with researchers employing techniques such as solid phase extraction (SPE), QuEChERS, and pressurized liquid extraction (PLE) to remove interfering components while preserving as many compounds as possible with adequate sensitivity [8]. For broader chemical coverage, multi-sorbent strategies combining materials like Oasis HLB with ISOLUTE ENV+, Strata WAX, and WCX have proven effective [8].

Following sample preparation, HRMS platforms coupled with liquid or gas chromatographic separation (LC/GC) generate the complex datasets essential for NTA. Specific experimental protocols vary by instrument platform but share common elements:

Liquid Chromatography Conditions:

• Column: C18 reversed-phase column (e.g., 100 × 2.1 mm, 1.7-1.8 μm particle size)
• Mobile Phase: Water/acetonitrile or water/methanol with 0.1% formic acid or ammonium acetate buffers
• Gradient: 5-95% organic modifier over 15-30 minutes
• Flow Rate: 0.2-0.4 mL/min
• Injection Volume: 1-10 μL

HRMS Acquisition Parameters:

• Ionization: Electrospray ionization (ESI) in positive and negative modes
• Mass Range: m/z 50-1000 or 150-1500
• Resolution: >25,000 (typically 50,000-240,000 for Orbitrap, >20,000 for Q-TOF)
• Mass Accuracy: <5 ppm with internal calibration
• Data Acquisition: Data-dependent acquisition (DDA) or data-independent acquisition (DIA)

Quality assurance measures include batch-specific quality control (QC) samples, internal standards, and system suitability tests to ensure data integrity throughout the acquisition process [8].

HRMS Data Processing and Compound Identification Workflow

The transformation of raw HRMS data into chemically meaningful information involves sophisticated computational workflows that leverage the high-quality data generated by modern HRMS platforms. The process begins with raw data conversion from vendor-specific formats to open formats (e.g., mzML), followed by peak detection, retention time alignment, and componentization to group related spectral features into molecular entities [8] [10].

Diagram 2: HRMS Data Processing for Confidence Assignment. The workflow transforms raw HRMS data through feature detection and candidate search to final confidence level assignment.

For compound identification, multiple computational approaches are employed:

Spectral Library Matching: Experimental MS/MS spectra are matched against reference libraries such as MassBank, NIST, METLIN, and GNPS using similarity metrics like cosine similarity, spectral entropy, or MS2DeepScore [10]. This approach typically provides Level 2 confidence annotations.

In Silico Fragmentation: Tools like MetFrag, CFM-ID, and GrAFF-MS predict fragmentation spectra from candidate structures and compare them to experimental data, extending identification capabilities beyond available reference libraries [10].

Retention Time Prediction: Machine learning models predict retention time indices (RTIs) from molecular structure or fragmentation data, providing orthogonal confirmation for compound identification [11].

Ion Mobility Integration: When available, collision cross-section (CCS) values provide an additional dimension for confirmation through comparison with experimental or predicted CCS databases [10].

Table 2: Key Research Reagent Solutions for HRMS-NTA

Reagent/Resource Category	Specific Examples	Function in NTA Workflow	Performance Considerations
Extraction Materials	Oasis HLB, ISOLUTE ENV+, Strata WAX/WCX, QuEChERS	Comprehensive extraction of diverse chemical classes from complex matrices	Chemical coverage, recovery efficiency, matrix effects
Chromatography Columns	C18, HILIC, phenyl-hexyl stationary phases	Separation of complex mixtures prior to HRMS analysis	Peak capacity, retention reproducibility, pH stability
HRMS Calibration Solutions	ESI-L Tuning Mix, sodium formate clusters	Mass accuracy calibration during HRMS acquisition	Mass accuracy stability, calibration range coverage
Spectral Libraries	MassBank, NIST, METLIN, GNPS, MoNA	Reference spectra for compound identification by spectral matching	Library size, spectral quality, chemical domain coverage
In Silico Prediction Tools	MetFrag, CFM-ID, SIRIUS+CSI:FingerID	In silico spectrum prediction and compound identification	Prediction accuracy, computational efficiency, chemical space coverage
Quantitative Structure-Retention Relationship Databases	C3-14 n-alkylamide RTI system, Unified CCS Compendium	Retention time and CCS prediction for identification confidence	Prediction accuracy, transferability between laboratories

Comparative Performance of HRMS Platforms in NTA Applications

Technical Specifications and Performance Metrics

The analytical performance of different HRMS platforms directly impacts their effectiveness in NTA applications. The two predominant HRMS technologies used in NTA—Orbitrap and quadrupole time-of-flight (Q-TOF) systems—each offer distinct advantages and limitations for non-targeted screening [8].

Orbitrap mass analyzers provide exceptionally high mass resolution (typically 140,000-240,000 at m/z 200) and high mass accuracy (<1-3 ppm with internal calibration), which significantly enhances molecular formula assignment confidence and facilitates the separation of isobaric compounds [8] [10]. The Fourier transform-based detection principle enables selective ion accumulation and multiplexed acquisition schemes, though this can sometimes create spectral dependencies between adjacent ions. Modern Orbitrap systems typically offer a dynamic range of 4-5 orders of magnitude and are frequently coupled to chromatographic systems with tightly controlled conditions that minimize retention time drift [8].

In comparison, Q-TOF instruments provide high resolution (typically 30,000-100,000), excellent mass accuracy (<2-5 ppm), and faster acquisition speeds, which is advantageous for comprehensive characterization of complex mixtures with ultra-fast chromatography [8]. The TOF technology offers inherently parallel detection without ion trapping limitations, providing more linear dynamic range (up to 5 orders of magnitude) and minimal spectral skewing. However, Q-TOF systems may require more frequent mass calibration and can exhibit greater susceptibility to retention time drift compared to Orbitrap systems coupled with high-performance liquid chromatography [8].

Table 3: Comparative Performance of HRMS Platforms in NTA Applications

Performance Parameter	Orbitrap Systems	Q-TOF Systems	Impact on NTA Performance
Mass Resolution	140,000-240,000 (at m/z 200)	30,000-100,000	Higher resolution improves separation of isobaric compounds
Mass Accuracy	<1-3 ppm (with internal calibration)	<2-5 ppm (with frequent calibration)	Better accuracy reduces molecular formula candidates
Acquisition Speed	6-20 Hz (depending on resolution)	10-100 Hz	Faster acquisition better captures narrow chromatographic peaks
Dynamic Range	10^4-10^5	10^4-10^5	Wider range enables detection of low-abundance compounds
Fragmentation Efficiency	HCD and CID capabilities	CID capabilities with collision energy ramping	Flexible fragmentation improves structural elucidation
Retention Time Stability	Typically lower drift due to coupled LC systems	Potentially higher drift in some configurations	Better stability improves alignment across multiple samples
Spectral Library Match Scores	Comparable performance when using appropriate libraries	Comparable performance with optimized conditions	Directly impacts confidence level assignment

Experimental Performance in Real-World Applications

The performance differences between HRMS platforms manifest in practical NTA applications, particularly for complex environmental and biological samples. Studies comparing annotation rates across different instrument platforms reveal that both Orbitrap and Q-TOF systems can successfully identify hundreds to thousands of compounds in complex matrices, though the specific annotations may vary due to differences in ionization efficiency, fragmentation patterns, and mass accuracies [10].

In a comparative study of wastewater samples analyzed by both platforms, Albergamo et al. tentatively identified 884 and 550 of the prioritized LC/HRMS features in positive and negative electrospray ionization modes, respectively, using in silico fragmentation tools [10]. However, only 106 and 139 of these annotations yielded high enough scores for further verification, highlighting the challenge of confident identification regardless of platform. Subsequent analytical standard confirmation validated 25 of 42 tested candidate structures, demonstrating that even high-confidence annotations require experimental verification [10].

For machine learning-enhanced NTA applications, recent research demonstrates significant improvements in identification probability (IP) when combining HRMS data with predictive models. One study incorporating reference spectral library searches and retention time index errors achieved a weighted F1 score of 0.65 and a Matthews correlation coefficient of 0.30 for pesticides at concentrations of 1 to 1000 ppb in blank samples [11]. Compared to spectral library matching alone, the average identification probabilities for pesticides increased by 54.5%, 52.1%, and 46.7% when spiked in blank, 10× diluted, and 100× diluted tea matrices, respectively [11]. These results highlight how computational approaches can leverage HRMS data to substantially enhance confidence in compound annotations.

Advanced Applications and Future Directions

Machine Learning Integration with HRMS Data

The integration of machine learning (ML) with HRMS data represents a transformative advancement in NTA, addressing the principal challenge of extracting meaningful environmental information from vast chemical datasets [8]. ML algorithms excel at identifying latent patterns within high-dimensional data, making them particularly well-suited for contamination source identification and compound prioritization [8]. For instance, ML classifiers including Support Vector Classifier (SVC), Logistic Regression (LR), and Random Forest (RF) have successfully screened 222 targeted and suspect per- and polyfluoroalkyl substances (PFASs) across 92 samples, achieving classification balanced accuracy ranging from 85.5% to 99.5% across different sources [8].

The systematic framework for ML-assisted NTA encompasses four critical stages: (1) sample treatment and extraction, (2) data generation and acquisition via HRMS, (3) ML-oriented data processing and analysis, and (4) result validation [8]. Within the data processing stage, ML techniques address key challenges through dimensionality reduction (PCA, t-SNE), clustering (HCA, k-means), and classification (RF, SVC) algorithms [8]. These approaches enable researchers to move beyond intensity-based prioritization, which risks overlooking low-concentration but high-risk contaminants, toward pattern recognition that captures source-specific chemical signatures [8].

Recent innovations include the development of ML models that enhance identification probability through integrated analysis of spectral and retention data. The implementation of a k-nearest neighbors (KNN) algorithm that incorporates retention time index errors derived from molecular fingerprint-based and cumulative neutral loss-based regression models has demonstrated significant improvements in distinguishing true positive spectral matches [11]. This approach exemplifies how ML can leverage the rich data generated by HRMS to overcome fundamental challenges in NTA, particularly the high rate of false positives and uncertain annotations.

Emerging Applications and Methodological Innovations

HRMS-based NTA continues to expand into new application domains while evolving methodologically to address existing limitations. In the field of exposomics, NTA plays an essential role in characterizing the broad spectrum of chemical exposures, with HRMS enabling comprehensive analysis of biological samples for both endogenous metabolites and exogenous contaminants [9] [11]. The development of large-scale collaborative initiatives such as the Network for EXposomics in the U.S. (NEXUS) highlights the growing importance of HRMS-NTA in public health research [9].

In regulatory contexts, HRMS-NTA is increasingly applied to challenging analytical scenarios such as extractables and leachables (E&L) testing for medical devices and materials biocompatibility assessment [9]. These applications demand particularly high confidence in identifications, driving innovations in reporting standards and confidence assessment protocols aligned with regulatory guidance from organizations like the FDA and ISO [9]. The BP4NTA (Best Practices for Non-Targeted Analysis) community has emerged as a key organization promoting harmonized methods, with technical subcommittees focused on specific application domains like E&L testing [9].

Methodologically, several cutting-edge approaches are extending the capabilities of HRMS-NTA:

Generative Models: New ML approaches like Mass2SMILES, JTVAE, Spec2Mol, and MSNovelist generate chemical structures directly from MS/MS spectra, potentially enabling annotation of compounds completely absent from existing databases [10].

Integrated Multi-platform Analysis: Combining data from complementary analytical platforms including LC-HRMS, GC-HRMS, and ion mobility spectrometry increases coverage of the chemical space and provides orthogonal confirmation for identifications [9] [10].

Large-Scale Spectral Databases: Resources like the Analytical Methods and Open Spectral Database (AMOS), which provides access to >6,500 analytical methods and >900,000 spectra, significantly expand the reference data available for compound identification [9].

Tiered Validation Strategies: Comprehensive validation approaches integrating reference material verification, external dataset testing, and environmental plausibility assessments ensure that ML-assisted NTA results are both analytically sound and environmentally relevant [8].

As these methodological innovations mature and HRMS technology continues to advance, the role of NTA in chemical discovery, environmental monitoring, and public health protection will undoubtedly expand, solidifying the critical position of HRMS as the foundational analytical technology for comprehensive chemical characterization.

Why Traditional Targeted Analysis Fails for Emerging Contaminants and Unknowns

Traditional targeted chemical analysis is a cornerstone of environmental monitoring and regulatory compliance, designed to detect and quantify a predefined set of analytes with high precision. However, this approach operates on a fundamental assumption: that the chemicals of concern are already known. In the face of emerging environmental contaminants (EECs)—such as novel pesticides, pharmaceuticals, industrial chemicals, and their transformation products—this assumption fails. EECs are characterized by their structural diversity and lack of analytical standards, making them invisible to targeted methods that rely on reference materials for identification and quantification [12]. With over 350,000 chemicals in use globally and thousands of intentional and unintentional chemical releases occurring annually—nearly 30% of which are of unknown composition—the limitations of targeted analysis are not just theoretical but pose a significant practical challenge for public health and ecological safety [13] [14].

This article objectively compares the performance of traditional targeted analysis against Non-Targeted Analysis (NTA) using High-Resolution Mass Spectrometry (HRMS). By framing the comparison within the context of chemical confidence levels, we demonstrate how NTA addresses the critical blind spots of targeted methods, transforming how researchers and drug development professionals identify and assess unknown chemical threats.

Fundamental Comparison: Targeted vs. Non-Targeted Analysis

The core difference between these methodologies lies in their scope and purpose. Targeted analysis is a closed system, while NTA is an open, discovery-oriented one.

Table 1: Core Methodological Comparison

Aspect	Traditional Targeted Analysis	Non-Targeted Analysis (NTA)
Analytical Scope	Limited to a predefined list of analytes [15]	Broad, unbiased screening for known and unknown chemicals [15]
Dependence on Standards	Requires analytical standards for every target [12]	Can identify compounds without a priori standards [14]
Identification of "Unknowns"	Fails when the chemical is not predefined [13]	Explicitly designed for identifying unknown or suspected contaminants [12]
Key Instrumentation	Typically low-resolution MS (e.g., GC-MS, QQQ-MS)	High-Resolution Mass Spectrometry (HRMS) like Q-TOF and Orbitrap [8]
Data Interpretation	Compares data against a library of known standards	Uses advanced informatics, computational tools, and ML to propose identities [12] [8]

The Critical Role of Confidence Levels in Chemical Identification

A standardized framework for reporting the confidence of identifications is crucial for interpreting NTA results and comparing them to the definitive identifications provided by targeted analysis. The community widely adopts the Schymanski confidence scale [13], which provides a transparent system for assigning a level of certainty to each identification.

• Level 1: Confirmed Structure is comparable to targeted analysis and requires confirmation with an authentic reference standard.
• Level 2: Probable Structure is often the realistic goal in rapid NTA, supported by high-resolution MS/MS spectrum match and other orthogonal data like retention time [13].
• Level 3: Tentative Candidate is assigned when no reference standard is available, but a candidate structure is proposed based on spectral data.
• Levels 4 & 5 represent an unequivocal molecular formula or a mass signal only, providing insufficient evidence for structural identification.

The distinction is clear: targeted analysis is designed to achieve Level 1 confidence for a limited set of compounds, while NTA systematically works to assign the highest possible confidence level (ideally Level 2 or 3) to a vast array of previously unknown features in a sample [13].

Experimental Evidence: NTA Performance in Mock Scenarios

The practical superiority of NTA for identifying unknowns has been demonstrated in designed mock rapid-response scenarios. A key study tested a focused NTA method on three real-world situations: a surrogate nerve agent in a beverage, illicit drugs in a home, and an industrial chemical spill into water [13].

The experimental protocol involved:

• Sample Preparation: Varying by scenario, but often using generic solid-phase extraction (SPE) with multi-sorbent strategies (e.g., Oasis HLB with ISOLUTE ENV+) to broaden the range of recoverable compounds [13] [8].
• Data Acquisition: Analysis using Liquid Chromatography-High Resolution Mass Spectrometry (LC-HRMS) with both full-scan and data-dependent MS/MS modes to collect precise molecular and fragment ion data [13].
• Data Processing: A multi-stream approach using five different data processing software tools and the NTA WebApp to achieve a consensus identification rapidly [13].
• Hazard Assessment: Using the Hazard Comparison Module (HCM) to rapidly assemble toxicity profiles for identified chemicals [13].

The results were telling. The NTA workflow correctly assigned structures to more than half of the 17 total features investigated across the scenarios, achieving Level 2 or 3 identifications within a 24-72 hour window critical for rapid response [13]. This demonstrates that NTA is not only viable but highly effective for identifying unknown stressors when targeted methods have failed.

Table 2: Key Metrics for Rapid Response NTA

Performance Metric	Targeted Analysis Performance	NTA Performance in Mock Scenarios [13]
Speed	Fast for predefined lists, fails completely for unknowns.	Results delivered in 24-72 hours after sample receipt.
Confidence	Level 1 for targeted compounds.	Achieved Level 2 or 3 identifications for most features.
Hazard Information	Available only for pre-selected chemicals.	Integrated hazard profiles for identified unknowns via the HCM.
Transferability	Highly standardized but inflexible.	Demonstrated as a viable supplemental tool for rapid response laboratories.

The Scientist's Toolkit: Essential Reagents and Materials for NTA

Transitioning to NTA requires a different set of tools and reagents focused on broad chemical coverage rather than specificity for a few analytes.

Table 3: Essential Research Reagent Solutions for NTA Workflows

Item	Function in NTA
Multi-Sorbent SPE Cartridges (e.g., Oasis HLB, ISOLUTE ENV+, Strata WAX/WCX)	Broad-range extraction and clean-up from various matrices; different sorbents recover compounds with diverse physicochemical properties [8].
LC-HRMS Grade Solvents (e.g., Methanol, Acetonitrile)	Used in generic chromatographic gradients (0-100%) to separate a wide hydrophobicity range of compounds with minimal analyte loss [14].
Generic Reversed-Phase LC Columns (e.g., C18)	Provides the primary separation mechanism for a vast space of semi-polar organic compounds in liquid samples [14].
Retention Time Index (RTI) Calibration Mix	A set of known chemicals used to calibrate and project retention times across different chromatographic systems, aiding in candidate prioritization [16].
Quality Control (QC) Samples	Pooled sample aliquots analyzed throughout the batch to monitor instrument stability and data quality throughout the NTA workflow [8].
HRMS Instrumentation (Q-TOF, Orbitrap)	Provides the high mass resolution and accuracy needed to determine elemental compositions and generate MS/MS spectra for structural elucidation [8] [14].

Advanced Integration: Machine Learning and Retention Time Prediction

Overcoming the identification bottleneck in NTA requires advanced computational power. Machine Learning (ML) is now being integrated into NTA workflows to enhance pattern recognition, structure identification, and toxicity prediction [12] [8]. ML classifiers like Random Forest (RF) and Support Vector Classifier (SVC) have been successfully used to classify samples according to their contamination source with balanced accuracy ranging from 85.5% to 99.5% by recognizing complex, source-specific chemical fingerprints [8].

A critical advancement is the use of ML for retention time (RT) prediction. RT is an essential orthogonal parameter for increasing confidence in candidate selection. Two primary approaches exist:

• Projection Methods: Use a small set of chemicals analyzed on two systems to project RTs from a source database to the lab's system [16].
• Prediction Methods: Use ML models trained on public RT data to predict the RT of a candidate structure directly from its chemical structure [16].

A 2025 study found that the accuracy of both methods is directly linked to the similarity of the chromatographic systems, with the pH of the mobile phase and column chemistry being the most impactful factors. For cases where the training data is similar to the lab's system, ML prediction models can perform on par with experimental projection methods [16].

NTA and ML Identification Workflow

The evidence is clear: traditional targeted analysis is fundamentally inadequate for the modern challenge of emerging contaminants and unknown chemical releases. Its failure is inherent in its design—it can only find what it is programmed to look for. Non-Targeted Analysis, powered by HRMS and advanced computational tools like machine learning, provides a powerful, complementary paradigm. By following a structured workflow and adhering to confidence level frameworks, NTA moves beyond simple detection to provide probable and tentative identifications that enable rapid response and informed decision-making. For researchers and scientists committed to a comprehensive understanding of the chemical environment, integrating NTA into their analytical arsenal is not just an advantage—it is a necessity.

In non-targeted analysis (NTA) and suspect screening, the confident identification of unknown chemicals represents a significant scientific challenge. The Metabolomics Standards Initiative (MSI) has established a framework of confidence levels to address this, where the highest confidence annotations require orthogonal data from multiple independent techniques [17]. This guide objectively compares the performance and contribution of four core analytical components—m/z, isotope patterns, fragmentation spectra, and retention time—in achieving these confidence levels. These components form an integrated system where the strengths of one compensate for the limitations of another, enabling researchers to navigate the complex detectable chemical space, which NTA has been shown to enhance by 20-fold compared to targeted analysis alone [18]. For researchers in drug development and environmental science, understanding the optimal application and limitations of each component is critical for reliable structure elucidation and subsequent risk assessment.

Core Component Performance Comparison

The following table provides a systematic comparison of the four core identification components, detailing their specific roles, technical requirements, and contributions to confidence levels.

Table 1: Performance Comparison of Core Identification Components

Component	Primary Role in Identification	Technical Requirements & Common Data	Contribution to MSI Confidence Levels
Precursor m/z	Determines the molecular mass and enables formula generation for the precursor ion [17].	High-resolution mass spectrometry (HRMS; ~1-5 ppm mass accuracy); reported in Daltons (Da) or mass-to-charge ratio (m/z).	Level 4: Unknown feature of interest. Level 3/2: Provides the first piece of evidence to search databases for possible structures [17].
Isotope Patterns	Validates the proposed molecular formula; indicates the presence of specific elements (e.g., Cl, Br, S) [19].	HRMS with sufficient resolution; Relative abundance and exact mass of isotopic peaks (e.g., M+1, M+2).	Level 3/2: Agreement between observed and theoretical isotope ratios increases confidence in the molecular formula, supporting a probable structure [19] [17].
Fragmentation Spectra (MS/MS)	Reveals substructures and functional groups; the most informative component for distinguishing between isomers [17] [20].	Tandem MS (MS/MS) with collision-induced dissociation (CID); product ion spectra (m/z and intensity).	Level 1: A confirmed match to a reference standard's MS/MS spectrum is a key orthogonal parameter for confident 2D structure annotation [17].
Retention Time (RT)	Provides a hydrophobicity-based index that is orthogonal to mass spectral data [17].	Consistent chromatographic conditions (column, solvent gradient, temperature); reported in minutes or seconds.	Level 1: A confirmed match to a reference standard's RT under identical analytical conditions is a key orthogonal parameter [17].

Experimental Protocols for Data Acquisition and Interpretation

Protocol for Interpreting Isotope Patterns

Isotope patterns provide a powerful tool for validating molecular formulas and identifying specific elements.

• Sample Preparation: The compound of interest is dissolved in a suitable solvent and introduced via liquid chromatography (LC) or direct infusion.
• Instrumentation and Data Acquisition: Data is collected using a high-resolution mass spectrometer (e.g., Orbitrap, Q-TOF) in MS1 mode. The instrument must be well-calibrated to ensure accurate mass measurement of the isotopic peaks [19].
• Data Interpretation:
  • Formula Assignment: The accurate mass of the monoisotopic peak is used to generate a list of candidate molecular formulas.
  • Pattern Matching: The observed isotopic pattern (relative intensities of M, M+1, M+2 peaks) is compared to the theoretical pattern for each candidate formula. Software tools perform this comparison automatically.
  • Elemental "Fingerprinting": Visually identify patterns indicative of specific elements. A pair of peaks separated by 2 Da with a ~3:1 intensity ratio suggests a single chlorine atom. A similar pair with a ~1:1 ratio suggests a single bromine atom [19].
• Performance Consideration: In ion trap instruments, resolution can be compromised at high ion counts. For accurate isotope peak analysis, it is best to use the high-mass end of the isotope cluster or reduce the target ion count to improve resolution [19].

Protocol for Acquiring and Interpreting MS/MS Spectra

Fragmentation spectra are the most informative data for structural elucidation.

• Sample Preparation: Similar to the protocol for isotope patterns.
• Instrumentation and Data Acquisition:
  • Fragmentation Strategy: Common data-dependent acquisition (DDA) selects the top N most intense ions from an MS1 scan for fragmentation. Data-independent acquisition (DIA) and SWATH are alternatives that fragment all ions within predefined windows [21] [22].
  • Fragmentation Parameters: Isolate the precursor ion with a narrow window (e.g., 1-2 m/z), and fragment using collision-induced dissociation (CID) with optimized collision energies [22].
• Data Interpretation:
  • Database Search: The experimental MS/MS spectrum is searched against spectral libraries (e.g., MassBank, mzCloud, GNPS) using similarity scoring algorithms [17].
  • In Silico Fragmentation: For compounds not in libraries, tools like MS-FINDER are used. This software retrieves candidate structures from databases, predicts their fragmentation using hydrogen rearrangement (HR) rules, and ranks the candidates by matching theoretical fragments to the experimental spectrum [20]. MS-FINDER has been validated to correctly identify the structural isomer with 82.1% accuracy within the top-3 candidates [20].
• Performance Consideration: The choice of N in DDA represents a trade-off. A higher N increases coverage but reduces the number of MS1 scans, potentially affecting quantitative accuracy [22]. Machine learning-assisted in silico tools are crucial as experimental MS/MS libraries cover only a tiny fraction of known metabolites [17] [20].

Protocol for Leveraging Retention Time

Retention time provides an orthogonal physicochemical property.

• Chromatographic Conditions: Ultra-high-performance liquid chromatography (UHPLC) with a defined stationary phase (e.g., C18), solvent gradient, flow rate, and column temperature is used [22] [23].
• Calibration: The retention time of the unknown is compared to that of an authentic analytical standard analyzed under identical chromatographic conditions.
• Prediction Models: As an alternative, in silico retention time prediction models can be used to provide supporting evidence when a standard is unavailable [17].

Integrated Workflows and Signaling Pathways

The following diagram illustrates the logical workflow for integrating the four core components to achieve a confident identification in non-targeted analysis.

Diagram 1: Identification confidence level workflow.

The Scientist's Toolkit: Essential Research Reagents and Materials

The following table lists key software tools and resources that are essential for implementing the workflows described in this guide.

Table 2: Key Research Reagent Solutions for Non-Targeted Analysis

Tool / Resource Name	Type	Primary Function in Identification
Compound Discoverer (Thermo Scientific)	Commercial Software	A comprehensive platform for processing NTA data, performing database searches for SSA, and predicting molecular formulas [15] [18].
MS-FINDER	Open-Source Software	Performs in-silico fragmentation for structure elucidation using hydrogen rearrangement rules, crucial for identifying compounds absent from libraries [20].
xcms (Bioconductor)	Open-Source R Package	Used for peak detection, retention time alignment, and statistical analysis in LC-MS-based metabolomics [21].
MassBank	Public Spectral Library	A community-wide repository of experimental MS/MS spectra used for spectral matching against known reference compounds [17] [20].
ViMMS	Simulation Framework	Allows for in-silico simulation of LC-MS/MS methods to optimize fragmentation strategies before resource-intensive instrument time [22].

From Raw Data to High-Confidence IDs: Advanced NTA Workflows and Machine Learning Integration

In the field of chemical exposome characterization, non-targeted analysis (NTA) has emerged as a powerful, discovery-based approach for identifying unknown or unsuspected chemicals in complex samples. [15] Unlike targeted methods that search for predefined analytes, NTA employs high-resolution mass spectrometry (HRMS) to detect a broad spectrum of substances without prior knowledge of their presence. [12] [15] This capability is crucial for advancing environmental health research, identifying emerging contaminants, and ensuring the safety of consumer products. [24] [15]

However, the transition from raw sample to confident chemical identification is fraught with analytical challenges. The structural diversity of potential contaminants, their typically low concentrations, and the lack of available analytical standards necessitate a rigorous, multi-stage workflow. [12] [24] This guide dissects and compares the methodologies within a foundational four-stage workflow for NTA: Sample Treatment, Data Acquisition, ML-Oriented Processing, and Validation. By objectively examining the performance of different approaches at each stage, we provide a framework for researchers to optimize their protocols and enhance the confidence levels of chemical assignments in their non-targeted analysis research.

Stage 1: Sample Treatment and Preparation

The initial stage of sample treatment is critical for determining which chemicals will be detectable in subsequent analysis. The objective is to extract a broad range of analytes from the sample matrix while minimizing co-extraction of interfering substances. The chosen methods profoundly influence the "detectable chemical space." [15]

Comparative Methodologies and Performance

Sample preparation strategies vary significantly depending on the sample matrix. The table below summarizes common approaches and their performance implications across different sample types.

Table 1: Comparison of Sample Treatment Methods for Various Matrices

Sample Matrix	Common Extraction & Migration Methods	Key Performance Considerations	Commonly Detected Chemical Classes
Plastic Food Contact Materials (FCMs)	Extraction with 95% ethanol; Migration to food simulants (e.g., 95% ethanol) at 60°C for 10 days. [24]	Mimics worst-case real-world scenarios; identifies migratable compounds. [24]	Oligomers, degradation products, additives, and contaminants. [24]
Water	Solid-phase extraction (SPE). [15]	Effectiveness depends on sorbent chemistry; can target a wide or specific polarity range.	Pharmaceuticals, per- and polyfluoroalkyl substances (PFAS). [15]
Soil & Sediment	Pressurized liquid extraction, ultrasonic extraction. [15]	Efficient for complex, solid matrices; can be tailored for specific contaminant classes.	Pesticides, polyaromatic hydrocarbons (PAHs). [15]
Dust	Solvent shaking, Soxhlet extraction. [15]	Addresses complex mixture of organic chemicals in a solid indoor environment matrix.	Flame retardants, plasticizers. [15]
Human Biospecimens	Protein precipitation, liquid-liquid extraction. [15]	Requires high sensitivity due to low analyte concentrations; must remove proteins and lipids.	Plasticizers, pesticides, halogenated compounds. [15]

Experimental Protocol: Migration Testing for Plastic FCMs

A typical protocol for assessing non-intentionally added substances (NIAS) in plastic Food Contact Materials involves: [24]

• Sample Preparation: The FCM (e.g., polypropylene film) is cut into pieces to increase surface area.
• Migration Exposure: The pieces are immersed in a food simulant (e.g., 95% ethanol, D2) under controlled time and temperature conditions (e.g., 10 days at 60°C) to simulate long-term storage. [24]
• Extract Collection: The simulant is collected after the exposure period and is often concentrated prior to analysis to enhance the detection of trace-level migrants.

Stage 2: Data Acquisition via High-Resolution Mass Spectrometry

Data acquisition transforms the chemical extract into instrumental data, serving as the foundation for all subsequent discoveries. The choice of chromatographic and mass spectrometric platforms directly defines the "detectable space" of the NTA study. [15]

Instrumental Platforms and Their Detectable Chemical Space

The two primary chromatographic techniques coupled to HRMS are Liquid Chromatography (LC) and Gas Chromatography (GC), each with distinct advantages.

Table 2: Comparison of Data Acquisition Platforms in NTA

Platform	Ionization Methods	Ideal Chemical Space	Relative Usage in NTA Studies [15]	Key Strengths
LC-HRMS	Electrospray Ionization (ESI+, ESI-), Atmospheric Pressure Chemical Ionization (APCI). [15]	Polar, non-volatile, and thermally labile compounds. [15]	51% (LC-only); 43% use both ESI+ and ESI-. [15]	Broad coverage of pharmaceuticals, pesticides, and many industrial chemicals.
GC-HRMS	Electron Ionization (EI), sometimes complemented by Chemical Ionization (CI). [15]	Volatile and semi-volatile, thermally stable compounds. [15]	32% (GC-only). [15]	Highly reproducible, library-matchable EI spectra; excellent for hydrocarbons, PAHs, many pesticides.
Dual Platform	Combination of the above.	Maximizes the breadth of detectable chemical space.	16% (Both LC & GC). [15]	Most comprehensive approach for capturing a wide array of chemical properties.

Experimental Protocol: LC-HRMS Analysis with QTOF

A typical LC-HRMS method for NTA involves: [24]

• Chromatography: An aliquot of the sample extract is injected onto a UPLC system equipped with a C18 column (e.g., 2.1 x 100 mm, 1.7 µm). Separation is achieved using a gradient of water and methanol (both modified with 0.1% formic acid) at a flow rate of 0.3 mL/min over a 13-minute run time. [24]
• Mass Spectrometry: The eluent is analyzed using a quadrupole time-of-flight (QTOF) mass spectrometer with electrospray ionization (ESI) in both positive and negative modes.
• Data Acquisition: Data-independent acquisition (DIA or MSE) is often used, which fragments all ions in a predefined range without pre-selection, ensuring a comprehensive record of all detectable features and their fragmentation patterns for later interrogation. [24]

Stage 3: ML-Oriented Data Processing and Compound Identification

The vast datasets generated by HRMS require sophisticated computational tools for peak picking, compound identification, and prioritization. This is where machine learning (ML) and automated workflows demonstrate their significant potential. [12]

Machine Learning Workflow and Data Processing Tools

ML-assisted NTA leverages computational models to optimize workflows, improve structure identification, and predict toxicity. [12] The process follows a systematic framework from data to deployment.

Diagram 1: Iterative machine learning lifecycle for NTA.

The core of ML-oriented processing involves software tools for peak picking and compound identification. The choice between vendor and open-source software presents a key decision point.

Table 3: Comparison of Data Processing and Identification Tools

Tool Category	Examples	Common Identification Methods	Usage in NTA Studies [15]	Considerations
Vendor Software	Thermo Compound Discoverer, Agilent MassHunter. [15]	Spectral library matching (mzCloud, MassBank), suspect screening against custom databases. [15]	Majority of studies (e.g., 57 out of 76 reviewed). [15]	Integrated, user-friendly, but often proprietary and costly.
Open-Source Software	MZmine, MS-DIAL. [15]	Library matching, in-silico fragmentation, formula prediction.	Fewer studies (7 out of 76 reviewed). [15]	High flexibility and transparency; requires computational expertise.
In-silico Prediction	QSAR models, OPERA. [25]	Prediction of physicochemical

Leveraging Machine Learning for Pattern Recognition and Source Identification

Non-target analysis (NTA) represents a paradigm shift in analytical chemistry, moving from hypothesis-driven investigations toward discovery-based science. In fields ranging from exposomics to drug development, researchers face the monumental challenge of identifying unknown or unsuspected chemicals without a priori knowledge of what exists within complex samples [15]. High-resolution mass spectrometry (HRMS) generates immense, data-rich landscapes containing thousands of chemical features from a single sample—a volume and complexity that vastly exceeds human analytical capacity [26]. This data deluge has created a critical bottleneck in converting raw instrumental data into confident chemical identifications, particularly within the structured confidence levels framework that governs non-target analysis assignment research.

Machine learning (ML) has emerged as a transformative technology for pattern recognition and source identification in this context. By automating the detection of subtle patterns within high-dimensional chemical data, ML algorithms can dramatically accelerate the processes of feature prioritization, compound classification, and structural elucidation [12]. This comparative guide objectively evaluates the performance of different machine learning approaches applied to non-target analysis, with a specific focus on their capabilities for advancing chemical confidence levels assignment. We present experimental data and standardized protocols to help researchers select appropriate ML strategies for their specific NTA challenges, whether in environmental exposomics, pharmaceutical development, or other domains requiring comprehensive chemical characterization.

Machine Learning Approaches for Chemical Pattern Recognition

Algorithm Comparison and Performance Metrics

Table 1: Comparison of Machine Learning Algorithms for NTA Pattern Recognition

Algorithm Category	Best Applications in NTA	Reported Accuracy/Performance	Key Advantages	Major Limitations
Convolutional Neural Networks (CNNs) [27]	Image-like spectral data pattern recognition; MS1/MS2 feature detection	>85% accuracy in spectral similarity tasks [27]	Excellent at identifying local patterns; Minimal need for feature engineering	Requires large training datasets; Computationally intensive; "Black box" nature
Transformer Architectures [27] [28]	Spectral sequence prediction; Retention time modeling; Large-scale spectrum-structure relationships	15-30% improvement over RNNs in sequence modeling tasks [27]	Processes entire sequences simultaneously; Superior context awareness	Extreme computational demands; Complex implementation
Ensemble Methods (Bagging/Boosting) [27]	Compound classification; Source attribution; Confidence level prediction	75-90% accuracy in compound category classification [27] [12]	Reduces overfitting; Handles mixed data types well; More interpretable	Limited deep pattern discovery; Requires careful parameter tuning
Self-Supervised Learning [27]	Leveraging unlabeled HRMS data; Pretraining for limited labeled data scenarios	Effective with as little as 10% labeled data [27]	Overcomes labeled data scarcity; Creates transferable representations	Emerging methodology; Validation frameworks immature

The selection of appropriate machine learning algorithms depends heavily on the specific NTA challenge, available computational resources, and the nature of the chemical data. Convolutional Neural Networks (CNNs) excel at recognizing spatial patterns in spectral data, functioning similarly to their image recognition capabilities by detecting local relationships in mass spectrometry heatmaps or fragmentation spectra [27]. Transformer architectures, while computationally demanding, have demonstrated remarkable performance in sequence-based chemical data processing, such as predicting retention times or mapping fragmentation pathways by treating spectral data as linguistic sequences [27] [28]. For more traditional classification tasks, ensemble methods like random forests (bagging) and gradient boosting machines provide robust performance with greater interpretability—a valuable characteristic when working within the confidence framework for NTA where justification of identifications is essential [27].

Self-supervised learning represents an emerging paradigm that is particularly valuable for NTA applications where labeled chemical data is scarce. By learning inherent data structures from unlabeled HRMS data, these systems can create powerful foundational models that subsequently require minimal fine-tuning with labeled examples to perform specific identification tasks [27]. This approach mirrors the success of large language models in natural language processing, adapted to the chemical "language" of mass spectrometry.

Experimental Protocol for Benchmarking ML Algorithms in NTA

To objectively compare machine learning algorithms for pattern recognition in non-target analysis, we propose the following standardized experimental protocol:

1. Dataset Curation and Preprocessing

• Collect HRMS data from diverse matrices (water, soil, biological samples) with known spiked compounds across chemical classes (pharmaceuticals, pesticides, industrial chemicals, metabolites)
• Apply uniform data preprocessing: retention time alignment, peak picking, and intensity normalization using open-source tools (MS-DIAL, MZmine) [15]
• Annotate data with confidence levels (1-5) following the Schymanski framework [26]

2. Feature Engineering and Representation

• Extract three data representations: (1) Raw spectral vectors (m/z, intensity), (2) Chemical descriptor features (molecular weight, polarity, complexity indices), and (3) Image-like representations of spectral patterns
• Apply dimensionality reduction (PCA, t-SNE) to high-dimensional features for visualization and analysis

3. Model Training and Validation

• Implement k-fold cross-validation (k=5) with strict separation of training and test sets
• Train each algorithm category (CNNs, Transformers, Ensemble Methods, Self-Supervised) on identical hardware and data partitions
• Employ early stopping with a patience of 10 epochs to prevent overfitting

4. Performance Evaluation Metrics

• Calculate standard classification metrics: accuracy, precision, recall, F1-score
• Assess confidence calibration: expected calibration error (ECE) and reliability diagrams
• Evaluate computational efficiency: training time, inference time, memory footprint

This protocol enables direct comparison of algorithmic performance while controlling for data quality and computational resource variables. Implementation requires approximately 2-4 weeks depending on dataset scale, with the feature engineering phase typically consuming 40-50% of the total project timeline.

Machine Learning-Enhanced Workflows for Confidence Level Assignment

Confidence Framework Integration

Table 2: ML Applications Across Chemical Identification Confidence Levels

Confidence Level	Traditional Identification Requirements	ML Enhancement Capabilities	Reported Performance Gains
Level 1 (Confirmed Structure)	Reference standard match; Retention time; MS/MS spectrum	Retention time prediction; Spectral similarity ranking; Automated database mining	45% reduction in standard acquisition needs; 3x faster verification [12]
Level 2 (Probable Structure)	Library spectrum match; Diagnostic evidence	In silico MS/MS prediction; Consensus scoring across multiple libraries	80% agreement with experimental spectra for known compounds [12]
Level 3 (Tentative Candidate)	Class-specific fragmentation; Literature data	Chemical class prediction from fragmentation patterns; Structure-function relationship modeling	92% accuracy in compound class assignment [15]
Level 4 (Unequivocal Molecular Formula)	Elemental composition from mass accuracy	Molecular formula assignment from isotopic patterns; Database prioritization	95% accurate formula assignment from high-resolution mass data [26]
Level 5 (Exact Mass)	m/z value only	Mass trend analysis; Homologue series detection; Blank subtraction automation	99% accuracy in detecting reproducible features across samples [26]

Machine learning technologies offer distinctive value propositions across the confidence level hierarchy for non-target analysis. At Confidence Level 1, where definitive structural confirmation requires authentic standards, ML models can dramatically reduce the need for physical standards by accurately predicting retention times and mass spectral patterns for candidate structures [12]. For Level 2 identifications, in silico fragmentation tools enhanced by machine learning can generate theoretical MS/MS spectra for tentative candidates, with recent advances achieving approximately 80% agreement with experimental spectra for known compound classes [12].

At Confidence Level 3, where specific stereochemistry may be unknown but compound class assignment is possible, machine learning classifiers excel at recognizing subtle patterns in fragmentation spectra that distinguish between chemical categories (e.g., phospholipids versus triglycerides, or polyfluoroalkyl substances versus hydrocarbon surfactants) [15]. For Level 4 assignments, ML algorithms improve molecular formula determination by integrating multiple lines of evidence beyond simple mass accuracy, including isotopic pattern recognition, heuristic rules regarding element probability, and database-derived likelihoods [26]. Even at Level 5, where only accurate mass information is available, machine learning can prioritize features for further investigation by recognizing patterns in detection frequency, intensity relationships across sample types, and mass defect trends characteristic of particular compound families [26].

Workflow Visualization

ML-Enhanced Confidence Level Assignment Workflow

The diagram above illustrates the integrated machine learning workflow for non-target analysis. The process begins with raw HRMS data acquisition and feature detection, followed by machine learning-powered pattern recognition that simultaneously supports multiple confidence levels of identification. This parallel processing capability represents a significant advancement over traditional sequential approaches, enabling more efficient utilization of analytical data and computational resources. The workflow culminates in comprehensive source identification and apportionment, leveraging the multi-level confidence assignments to provide nuanced insights into chemical origins and transformations.

Source Identification and Apportionment Capabilities

Experimental Data on Source Tracking Performance

Table 3: ML Performance in Chemical Source Identification Applications

Application Domain	ML Technique	Source Identification Accuracy	Key Experimental Findings	Limitations & Challenges
Environmental Source Tracking [15]	Random Forest Classification	89% accuracy in pollution source attribution	Successfully discriminated agricultural, urban, and industrial sources; Key features: pesticide profiles, PAH ratios, halogenated compound patterns	Performance degraded with aging/transformed chemicals (15% accuracy drop)
Exposomics Personal Care Product Attribution [15]	CNN Spectral Pattern Recognition	78% accuracy in product category matching	Identified fragrance signatures across household products; Detected metabolite-parent relationships in biological samples	Co-formulant interference reduced discriminative power
Pharmaceutical Impurity Sourcing [29] [30]	Anomaly Detection + Clustering	94% accuracy in manufacturing process defect identification	Correlated impurity profiles with specific synthetic pathways; Predicted degradants from stability data	Limited by proprietary process knowledge gaps
Metabolite Biological Pathway Assignment [12]	Graph Neural Networks	82% accuracy in pathway attribution	Mapped unknown metabolites to biotransformation pathways using mass similarity networks	Performance varied significantly by pathway (35-92% range)

Machine learning dramatically enhances source identification in non-target analysis by recognizing complex multivariate patterns that elude univariate statistical approaches. In environmental applications, random forest classifiers have demonstrated approximately 89% accuracy in attributing chemical profiles to specific pollution sources (agricultural, urban wastewater, industrial) by considering the complete contaminant fingerprint rather than individual marker compounds [15]. For exposomics applications, convolutional neural networks can match personal care product signatures across environmental and biological samples with 78% accuracy, enabling connections between product usage and human exposure through metabolite detection [15].

In pharmaceutical contexts, anomaly detection algorithms combined with clustering techniques have achieved 94% accuracy in identifying manufacturing process deviations and predicting impurity formation pathways, providing crucial quality control insights during drug development [29] [30]. For metabolic pathway assignment, graph neural networks represent an emerging approach that structures mass spectral relationships as networks, achieving 82% accuracy in mapping unknown metabolites to their biotransformation pathways by leveraging both chemical similarity and co-occurrence patterns across samples [12].

Source Identification Methodology

ML-Driven Source Identification Process

The source identification process begins with comprehensive chemical feature space characterization, followed by machine learning-powered pattern recognition to reduce dimensionality and extract meaningful signatures. These patterns are compared against source signature libraries using classification algorithms that both assign samples to known sources and flag novel or unknown source profiles for further investigation. The workflow produces quantitative source apportionment estimates, critically indicating the proportional contribution of each identified source to the overall chemical profile. This approach enables researchers to move beyond simple detection to meaningful source attribution—a crucial capability for solving complex environmental and biological exposure challenges.

Table 4: Research Reagent Solutions for ML-Enhanced NTA

Tool Category	Specific Solutions	Function in ML-NTA Workflow	Implementation Considerations
Data Generation Platforms	LC-HRMS; GC-HRMS; Ion Mobility-MS	Creates foundational data for ML pattern recognition	51% of studies use LC-HRMS only; 16% use both LC/GC-HRMS [15]
Open-Source Software Tools	MS-DIAL; MZmine; OpenMS	Feature detection, alignment, and preprocessing for ML	Only 7 of 57 studies used open-source tools [15]
Commercial Analysis Suites	Compound Discoverer; MassHunter	Integrated workflows from feature detection to identification	Dominant in current practice but create reproducibility challenges [15]
Spectral Libraries	NIST; mzCloud; GNPS	Training data for ML models; Verification of identifications	NIST most common for GC-HRMS; Limited for true unknown identification [15]
In Silico Prediction Tools	CFM-ID; MetFrag; SIRIUS	Generate theoretical spectra for confidence levels 2-3	80% agreement with experimental spectra for known compounds [12]
Computational Infrastructure	Cloud AI platforms; High-performance computing	Enable resource-intensive ML training and inference	58% of deployments use cloud-based platforms [31]

Successful implementation of machine learning for pattern recognition in non-target analysis requires both analytical chemistry tools and computational resources. High-resolution mass spectrometry platforms form the foundation, with liquid chromatography-HRMS (LC-HRMS) employed in 51% of studies, gas chromatography-HRMS (GC-HRMS) in 32%, and both platforms combined in only 16% of investigations—highlighting a significant opportunity for expanded chemical space coverage through complementary separations [15]. For data processing, open-source tools like MS-DIAL and MZmine provide transparent algorithms crucial for reproducible research, though currently only about 12% of studies leverage these open-source options, with the majority relying on commercial vendor software [15].

Spectral libraries serve as essential training data for supervised machine learning approaches, with the NIST library dominating GC-HRMS applications and various MS/MS libraries supporting LC-HRMS identifications. In silico prediction tools have evolved from rudimentary rule-based systems to sophisticated machine learning models that can predict mass spectral fragmentation patterns with approximately 80% accuracy for known compound classes, dramatically enhancing Confidence Level 2 and 3 assignments [12]. Computational infrastructure represents perhaps the most significant practical consideration, with cloud-based AI platforms dominating deployment (58% of implementations) due to their scalability and accessibility, particularly for research groups without dedicated high-performance computing resources [31].

Future Perspectives and Development Trajectories

The integration of machine learning with non-target analysis is rapidly evolving, with several emerging trends poised to further transform pattern recognition and source identification capabilities. Self-supervised learning approaches promise to address the fundamental challenge of labeled data scarcity in NTA by creating models that learn general chemical principles from unlabeled HRMS data before fine-tuning on specific identification tasks [27]. Transformer architectures, while computationally demanding, are demonstrating remarkable capabilities in predicting retention times and fragmentation patterns when trained on sufficiently large spectral datasets [28]. These advances parallel developments in natural language processing, treating mass spectral data as a chemical "language" with predictable patterns and relationships.

Interpretability remains a critical challenge for machine learning in regulatory and scientific contexts, spurring development of explainable AI (XAI) techniques that illuminate the reasoning behind ML-derived identifications [27]. Methods such as SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) are being adapted to mass spectral interpretation, helping chemists understand which fragment ions or chemical features most strongly influenced a particular classification decision [27]. This transparency is essential for advancing beyond "black box" models toward trustworthy systems that provide both identifications and chemically plausible justification for those assignments.

Looking forward, the field is moving toward increasingly automated and integrated workflows that combine robust experimental design, comprehensive data generation, and sophisticated machine learning into cohesive analytical systems. These systems will likely incorporate active learning approaches that strategically guide subsequent analyses based on initial findings, optimizing resource allocation for maximum information return. As these technologies mature, they hold the potential to transform non-target analysis from a specialized research activity into a routine component of chemical safety assessment, exposure science, and diagnostic applications—ultimately enabling more comprehensive understanding of the chemical environments that shape human and ecological health.

Seven Key Prioritization Strategies to Focus on High-Risk Compounds

In the analysis of complex chemical mixtures, non-target screening (NTS) using high-resolution mass spectrometry (HRMS) has become an essential discovery tool. However, a single sample can yield thousands of detected features, creating a significant bottleneck during the identification stage [32]. Without a structured approach to prioritize these features, valuable resources can be wasted on irrelevant signals, potentially causing truly high-risk compounds to be overlooked. This guide compares seven key prioritization strategies that enable researchers to focus confidently on the most relevant and hazardous chemicals, directly supporting the broader thesis of establishing chemical confidence levels in non-target analysis.

Comparative Analysis of Prioritization Strategies

The table below summarizes the seven core prioritization strategies, their primary functions, and key comparative aspects to guide method selection.

Table 1: Overview of the Seven Key Prioritization Strategies for Non-Target Screening

Strategy Number & Name	Primary Function	Key Tools & Databases	Relative Workflow Speed	Best for Identifying
P1: Target & Suspect Screening	Identifies known or suspected contaminants from lists [32].	PubChemLite, CompTox Dashboard, NORMAN Suspect List Exchange [32]	Fast	Compounds with existing regulatory or research interest
P2: Data Quality Filtering	Removes analytical artifacts and unreliable signals [32].	Peak shape analysis, blank subtraction, replicate consistency checks [32]	Fast	A clean, reproducible dataset for downstream analysis
P3: Chemistry-Driven Prioritization	Finds compounds based on chemical properties or class [32].	Mass defect filtering, homologue series analysis, diagnostic fragments [32]	Medium	PFAS, halogenated compounds, transformation products
P4: Process-Driven Prioritization	Highlights compounds changing due to a process [32].	Correlation analysis (e.g., upstream vs. downstream, before vs. after treatment) [32]	Medium	Persistent, formed, or removed compounds in dynamic systems
P5: Effect-Driven Prioritization	Isolates compounds responsible for biological effects [32].	Effect-Directed Analysis (EDA), Virtual EDA (vEDA) with statistical models [32]	Slow	Bioactive contaminants with direct risk potential
P6: Prediction-Based Prioritization	Ranks features by predicted risk using models [32].	MS2Quant (concentration), MS2Tox (toxicity), Risk Quotient (PEC/PNEC) [32]	Medium	High-risk compounds without full identification
P7: Pixel/Tile-Based Approaches	Analyzes regions of interest in complex chromatograms before peak detection [32].	Pixel-based (GC×GC) or tile-based (LC×LC) variance analysis [32]	Medium	Key chemical features in highly complex samples

A second critical table compares the quantitative risk-based outputs of these strategies, which is essential for confident risk assessment.

Table 2: Comparison of Risk Assessment and Quantitative Outputs Across Strategies

Strategy	Primary Risk Metric	Quantification Support	Key Data Inputs	Confidence Level for Identification
P1: Target/Suspect	Known hazard data from databases	Targeted methods possible post-identification	m/z, RT, isotope pattern, MS/MS spectra [32]	High (for targets) to Medium (for suspects)
P5: Effect-Driven	Direct biological activity (e.g., toxicity, receptor binding)	Requires post-identification quantification	Bioassay data, statistical correlation to chemical features [32]	Direct link to biological effect
P6: Prediction-Based	Risk Quotient (PEC/PNEC) [32]	Yes (e.g., via MS2Quant) [32]	MS/MS spectra, predictive model outputs [32]	Model-dependent
P3: Chemistry-Driven	Class-based known hazards (e.g., PFAS, PAHs)	Limited, class-based	Mass defect, isotope patterns, fragment ions [32]	Medium (for compound class)

Detailed Experimental Protocols

Protocol 1: Integrated Workflow for Prioritizing High-Risk Compounds

This protocol combines multiple strategies for a comprehensive assessment [32].

• Sample Preparation & HRMS Analysis: Prepare samples (e.g., water, sediment, biological extracts) using appropriate extraction methods. Analyze using LC-HRMS or GC-HRMS, acquiring data in both full-scan and data-dependent MS/MS modes [15].
• Data Preprocessing & Suspect Screening (P1 & P2): Process raw data (peak picking, alignment, componentization). Annotate features against suspect databases (P1) and apply data quality filters (P2) to remove features present in blanks and with poor replicate consistency [32].
• Chemistry- and Process-Driven Filtering (P3 & P4): Apply mass defect filters to highlight halogenated compounds (P3). Compare feature intensities across different sample groups (e.g., influent vs. effluent) to find compounds with poor removal (P4) [32].
• Risk Prediction & Prioritization (P6): For the remaining features, use in silico tools like MS2Tox to predict toxicity and MS2Quant to estimate concentration. Calculate a risk quotient (PEC/PNEC) to generate a risk-ranked shortlist [32].
• Structural Elucidation & Confirmation: For the top-ranked compounds, interpret MS/MS spectra, search spectral libraries, and, where possible, confirm identity with analytical standards.

Protocol 2: Effect-Directed Analysis (EDA) for Bioactive Compounds

This protocol directly links chemical features to biological activity [32].

• Fractionation: Separate the complex sample extract into multiple fractions using liquid chromatography (e.g., HPLC).
• Bioassay Testing: Test each fraction for a specific adverse biological effect (e.g., estrogenicity, cytotoxicity, genotoxicity).
• Chemical Analysis of Active Fractions: Analyze the bioactive fractions using HRMS (NTS workflow).
• Identification of Bioactives: Correlate the biological activity with the chemical features detected in the active fractions, leading to the identification of the causative agents.

Workflow Visualization

The following diagram illustrates the logical relationship and workflow for integrating the seven prioritization strategies.

The Scientist's Toolkit: Essential Research Reagents & Solutions

The following table details key materials and tools required for implementing the prioritization strategies discussed.

Table 3: Essential Reagents and Tools for Non-Target Screening and Prioritization

Item Name	Function / Application	Example Use Case
LC-HRMS & GC-HRMS Systems	High-resolution separation and accurate mass measurement for broad chemical detection [15].	Fundamental platform for all NTS data acquisition.
Suspect List Databases	Digital libraries of known or suspected contaminants for initial screening [32].	P1: Rapid annotation of features from the NORMAN Suspect List Exchange.
Stable Isotope-Labeled Internal Standards	Controls for assessing extraction efficiency, matrix effects, and instrument performance.	P2: Differentiating true signals from artifacts; quality control.
Diagnostic Fragment Ion Libraries	Curated lists of mass fragments indicative of specific chemical classes.	P3: Confirming the presence of PFAS or plasticizers via characteristic fragments.
In Vitro Bioassay Kits	Testing sample toxicity for specific endpoints (e.g., estrogenicity, cytotoxicity).	P5: EDA to isolate fractions causing biological effects.
Software for Predictive Modeling	Tools for predicting concentration and toxicity directly from MS data.	P6: Using MS2Quant and MS2Tox to calculate a risk quotient.
Certified Reference Standards	Analytically pure chemicals for confirming compound identity and quantifying results.	Final confirmation of high-priority compounds identified via any strategy.

No single prioritization strategy is sufficient to navigate the complex data from non-target screening. A sequential, integrated workflow that combines chemical knowledge, biological effect data, and predictive modeling is the most effective path to identifying high-risk compounds. By applying these seven strategies, researchers can transform an overwhelming dataset into a manageable list of high-priority candidates, thereby building a more confident and comprehensive understanding of the chemical exposome.

In both modern material science and pharmaceutical development, non-targeted analysis (NTA) has become an indispensable tool for identifying unknown chemical constituents. For food contact materials (FCMs), this primarily focuses on uncovering non-intentionally added substances (NIAS)—impurities, breakdown products, or contaminants that may migrate into food [33]. In novel drug modality development, NTA characterizes complex therapeutic agents like cell and gene therapies, where comprehensive molecular understanding is critical for safety and efficacy. While the analytical techniques share common technological foundations, their application, regulatory frameworks, and the consequences of identification uncertainty differ substantially. This guide compares the performance of NTA approaches across these two critical fields, framed within the essential research on chemical confidence levels.

Analytical Frameworks and Regulatory Context

The drive for rigorous chemical characterization in both fields is underpinned by distinct regulatory imperatives and analytical challenges.

NIAS in Food Contact Materials

The European Union's updated Regulation (EU) 2025/351, known as the 19th Amendment, explicitly introduces a "high degree of purity" requirement for plastics. It mandates that NIAS must be assessed and controlled, creating a pressing need for robust NTA methods [34] [35]. The regulation defines specific migration thresholds: ≤ 0.05 mg/kg for individually assessed non-genotoxic substances, and a stringent ≤ 0.00015 mg/kg for substances assessed via other risk assessment pathways [35]. The challenge is amplified by the complexity of supply chains, where NIAS can originate from impurities in raw materials, breakdown products, or contaminants during production [33].

Novel Drug Modalities

The pharmaceutical landscape is increasingly dominated by complex new modalities. In 2025, they account for $197 billion, or 60%, of the total pharma projected pipeline value [36]. This category includes advanced therapies like cell therapies (CAR-T), gene therapies, and nucleic acids (RNAi, DNA/RNA therapies). Characterization of these products requires NTA to identify process-related impurities, product variants, and degradation products that are not part of the intended molecular structure. Unlike FCMs, where the concern is patient exposure via migration, the focus here is directly on patient safety and product efficacy.

Table 1: Comparison of Regulatory and Analytical Drivers

Aspect	NIAS in Food Contact Materials	Novel Drug Modalities
Primary Regulation	EU 19th Amendment (2025/351) [34] [35]	FDA Guidance, ICH Guidelines
Key Objective	Ensure a "high degree of purity," prevent food contamination [33]	Ensure patient safety, product efficacy, and consistency
Defined Limits	Specific migration limits (e.g., 0.05 mg/kg, 0.00015 mg/kg) [35]	Product-specific impurities and variants (often ppm relative to API)
Typical Sample	Polymer extracts, food simulants	Drug substance/product, in-process samples
Major Challenge	Long, complex supply chains; diverse NIAS sources [33]	Extreme structural complexity; large biomolecules

Experimental Protocols for NTA and Confidence Assignment

The workflow for NTA is foundational to generating reliable data. The following protocol, incorporating prioritization strategies, is adaptable to both FCM and pharmaceutical applications.

Sample Treatment and Extraction

• For FCMs: Migration testing is performed using food simulants (e.g., ethanol/water mixtures, vegetable oil) under standardized time/temperature conditions representing intended use [35].
• For Biologics: Samples are prepared using techniques that preserve molecular integrity, such as gentle denaturation or digestion for proteins/nucleic acids.
• Broad-Spectrum Extraction: Use multi-sorbent solid-phase extraction (SPE) strategies (e.g., Oasis HLB with ISOLUTE ENV+) to maximize the range of recoverable analytes with varying polarities [8].

Data Generation and Acquisition

• Platform: Utilize Liquid Chromatography coupled to High-Resolution Mass Spectrometry (LC-HRMS) with Quadrupole Time-of-Flight (Q-TOF) or Orbitrap mass analyzers [8].
• Chromatography: Employ reversed-phase or HILIC chromatography to separate a wide range of compounds. For highly complex samples, two-dimensional liquid chromatography (LC×LC) increases peak capacity [32].
• Data Acquisition: Use data-dependent acquisition (DDA) or data-independent acquisition (DIA) modes to collect full-scan MS and MS/MS data for unknown identification.

ML-Oriented Data Processing and Analysis

This stage transforms raw data into interpretable patterns and is critical for managing high-dimensional datasets [8].

• Data Preprocessing: Perform peak picking, alignment, and componentization to group related features (adducts, isotopes). Handle missing values using imputation methods like k-nearest neighbors.
• Prioritization: Apply a suite of strategies to focus on the most relevant features [32]:
  • P1: Target/Suspect Screening: Match against predefined databases (e.g., NORMAN Suspect List Exchange).
  • P2: Data Quality Filtering: Remove artifacts and unreliable signals based on blanks and replicate consistency.
  • P5: Effect-Directed Prioritization: Link features to biological endpoints using statistical models.
  • P6: Prediction-Based Prioritization: Calculate risk quotients (PEC/PNEC) using tools like MS2Tox to estimate toxicity from MS/MS fragments [32].
• Compound Identification: Assign confidence levels using the Schymanski scale (Level 1-5), from confirmed structure (Level 1) to unequivocal molecular formula (Level 5).

Result Validation

• Tiered Validation: Implement a multi-faceted approach [8]:
  • Analytical Confidence: Use certified reference materials (CRMs) or spectral library matches for verification.
  • Model Generalizability: Validate machine learning classifiers on independent external datasets using cross-validation.
  • Environmental/Process Plausibility: Correlate findings with known source markers or process parameters.

The following workflow diagram integrates the core steps of sample processing, data analysis, and the critical decision point for confidence level assignment, which is central to the thesis of this guide.

Figure 1: Core Workflow for Non-Targeted Analysis with Confidence Level Assignment. A critical branching point occurs after identification, where tentative identifications (Levels 2-4) are often grouped into chemical classes for risk assessment, particularly in NIAS evaluation [37].

Comparative Performance and Experimental Data

The performance of NTA is measured by its ability to accurately identify chemicals and support risk-based decisions. The table below summarizes key comparative data and approaches.

Table 2: Comparison of NTA Performance and Data Outputs

Performance Metric	NIAS in Food Contact Materials	Novel Drug Modalities
Typical Confidence Level	Predominantly Tentative (Levels 2-3) [37]	Requires Confirmed (Level 1) for Critical Impurities
Key Risk Assessment Method	Toxicological Risk Assessment (TRA); grouping into chemical classes [37]	Qualification by toxicology studies; ICH Q3 guidelines
Quantification Approach	Semi-quantification using surrogate standards; comparison to migration limits	Quantification using authentic standards; ppm relative to Active Pharmaceutical Ingredient (API)
Handling Uncertainty	Grouping tentative IDs into classes with similar toxicological concern is acceptable [37]	Uncertainty must be resolved for product-related impurities; often requires isolation and definitive ID
Typical Workflow Output	Identification of NIAS sources for supply chain management [33]	Understanding of product heterogeneity, degradation pathways, process-related impurities

A pivotal finding in NTA research for medical devices, which directly applies to FCMs, is that tentative or partial identification is often sufficient for risk assessment. Chemicals are frequently grouped into classes based on structural similarity and presumed toxicological action, and the class is treated as a single entity for assessment. This obviates the need for analytically demanding, confirmed identification of every single compound, significantly reducing the burden without compromising the safety conclusion [37].

The Scientist's Toolkit: Essential Research Reagent Solutions

Successful implementation of the NTA workflow relies on a suite of specialized reagents, materials, and software tools.

Table 3: Key Research Reagent Solutions for NTA Workflows

Tool / Reagent	Function	Application Example
Multi-Sorbent SPE Cartridges (e.g., Oasis HLB, ISOLUTE ENV+, WAX/WCX)	Broad-spectrum extraction and cleanup of diverse analytes from complex matrices [8].	Enriching NIAS from food simulants or impurities from drug product formulations.
HRMS Quality Control Standards	Monitoring instrument performance, mass accuracy, and retention time stability during long sequences.	Ensuring data integrity across large batch analyses in both fields.
Suspect List Databases (e.g., NORMAN Suspect List Exchange, EPA CompTox)	Predefined lists of m/z values for known or suspected contaminants for suspect screening (P1) [32].	Screening for common plastic additives/oligomers or known process impurities from biomanufacturing.
Quantitative & Predictive Software (e.g., MS2Quant, MS2Tox)	Predicting concentration (MS2Quant) and toxicity (MS2Tox) directly from MS/MS spectra for prioritization (P6) [32].	Prioritizing features with high risk quotients (PEC/PNEC) before full identification.
Chemical Class-Based Assessment Templates	Frameworks for grouping tentatively identified compounds with similar structures for collective risk assessment [37].	Efficiently managing the risk of numerous, poorly characterized NIAS in FCMs.

The application of NTA in identifying NIAS and characterizing novel drug modalities reveals a shared technological foundation but distinct approaches to managing uncertainty. The FCM field, guided by the new EU amendment, can leverage strategic grouping of tentative identifications to conduct robust risk assessments efficiently [37]. In contrast, the novel drug modality field often demands confirmed identification for critical quality attributes but deals with molecules of unparalleled complexity. The ongoing development of machine learning-based prioritization and predictive toxicology tools is bridging the gap between detection and decision-making for both fields [32] [8]. Ultimately, the choice of workflow and the required confidence level must be fit-for-purpose, driven by a combination of regulatory requirements and a fundamental commitment to product safety.

Overcoming NTA Bottlenecks: Strategies for Data Quality, Prioritization, and Library Gaps

Addressing Spectral Library Gaps and Limited Reference Standards

In non-targeted analysis (NTA), the confidence level of compound identification is fundamentally constrained by the availability of comprehensive spectral libraries and authentic reference standards. This limitation represents a critical bottleneck across multiple scientific disciplines, from clinical metabolomics to environmental safety assessment. Spectral library searching serves as the most common approach for compound annotation in untargeted metabolomics, where experimental MS/MS spectra are matched against reference spectra of known molecules to generate structural hypotheses [38]. However, the field remains severely constrained by spectral library gaps and limited reference standards, resulting in heavy reliance on tentative identifications [39]. The consequences of these limitations extend throughout the analytical workflow, impeding confident compound identification, quantitative accuracy, and ultimately, the translation of research findings into actionable knowledge. This guide objectively compares current strategies for addressing these challenges, providing experimental data and methodological frameworks to inform researcher decision-making.

Current Landscape: Spectral Library Limitations and Coverage Gaps

The Scale of the Problem

The fundamental challenge in NTA lies in the disparity between the vast chemical space of potential analytes and the limited coverage of existing spectral libraries. While publicly accessible MS/MS small molecule spectral libraries have grown significantly over the past decade, this expansion has not kept pace with the diversity of compounds encountered in real-world samples [38]. This coverage gap is particularly pronounced for specific compound classes, including:

• Non-Intentionally Added Substances (NIAS) in plastic food contact materials, including oligomers, degradation products, and reaction byproducts [39]
• Emerging contaminants and their transformation products
• Specialized metabolites from understudied biological sources
• Isomeric compounds that cannot be distinguished by MS/MS alone [38]

Table 1: Comparison of Major Spectral Library Resources

Library Name	Scope/Coverage	Key Strengths	Limitations
GNPS Community Libraries [38]	Natural products, lipids, drugs, pesticides, microbial metabolites	Broad community contribution; integration with analysis ecosystem	Variable quality control; gaps in specific compound classes
NIST Tandem Mass Spectral Library [38]	Human and plant metabolites	Comprehensive coverage for included domains; commercial quality control	Limited coverage of emerging contaminants; commercial access
METLIN Gen2 [38]	Lipids, dipeptides, metabolites	Large scale; MS/MS data	Limited public accessibility; composition details not fully released
MassBank [38]	Diverse small molecules	Open access; international collaboration	Inconsistent coverage across compound classes
USGS Spectral Library Version 7 [40]	Minerals, plants, chemical compounds, man-made materials	Extensive wavelength coverage (UV to far infrared); well-characterized samples	Limited for molecular identification by MS

Impact on Identification Confidence

The Metabolomics Standards Initiative defines different confidence levels for compound identification, with level 1 representing confirmed identification using reference standards, and level 2 or 3 annotations resulting from spectral library matching [38]. The limitations of spectral libraries directly impact these confidence levels:

• Level 1 Identifications are rare for novel or poorly characterized compounds due to the lack of reference standards [39]
• Level 2 Annotations (probable structures based on spectral similarity) are constrained by library coverage
• Level 3 Annotations (characteristic chemical class) often represent the highest achievable confidence for unknown compounds [38]

Without all possible isomers tested under identical mass spectrometry conditions and chromatographic co-migration validation, even high-spectral similarity matches may represent incorrect structural assignments [38].

Comparative Analysis of Computational Solutions

In Silico Spectral Prediction and Machine Learning Approaches

Computational methods have emerged as promising approaches to address spectral library gaps. These can be broadly categorized into database-driven methods and machine learning-based prediction tools.

Table 2: Performance Comparison of Computational Approaches for Spectral Prediction

Method/Approach	Underlying Technology	Reported Performance	Limitations
GLMR Framework [41]	Generative language model; two-stage retrieval	>40% improvement in top-1 accuracy vs. baselines; MassSpecGym benchmark	Requires candidate molecules for generation; computational intensity
JESTR [41]	Cross-modal representation learning; contrastive learning	<20% top-1 accuracy in MassSpecGym	Modality misalignment between spectra and structures
MIST [41]	Molecular fingerprint inference from chemical formula	Limited by formula assignment accuracy	Dependent on accurate formula determination
Carafe [42]	Deep learning trained directly on DIA data	Improved fragment ion prediction vs. DDA-trained models	Initially developed for proteomics; small molecule adaptation needed
Traditional Library Matching [38]	Spectral similarity scoring	Performance bound by library coverage	Limited to known compounds; cannot identify novel structures

The GLMR (Generative Language Model-based Retrieval) framework represents a significant advancement, addressing the fundamental challenge of modality misalignment between mass spectra (physical fragmentation patterns) and molecular structures (chemical information) [41]. By employing a two-stage process—pre-retrieval of candidate molecules followed by generative refinement—GLMR transforms cross-modal retrieval into a more tractable unimodal similarity task.

Experimental Validation of Computational Methods

Recent benchmarking studies provide quantitative performance data for these computational approaches. On the MassSpecGym dataset (approximately 230k mass spectra with structurally diverse splits), the current state-of-the-art model JESTR demonstrated less than 20% top-1 accuracy, highlighting the persistent challenge of cross-modal alignment [41]. In contrast, the GLMR framework achieved over 40% improvement in top-1 accuracy compared to existing methods, demonstrating the effectiveness of its generative approach [41].

The performance advantage of GLMR was further validated on the MassRET-20k dataset, which includes richer spectral variations and more challenging real-world cases. This improved performance stems from the framework's ability to leverage contextual priors from candidate molecules while generating refined molecular structures that better align with the input mass spectrum [41].

Experimental Protocols for Library Expansion

Strategic Reference Standard Acquisition and Synthesis

While computational methods show promise, experimental approaches using reference standards remain the gold standard for confident identification. Strategic approaches include:

• Purified compound measurement: Laboratory samples of specific minerals, plants, chemical compounds, and man-made materials measured under controlled conditions to establish spectro-chemical links [40]
• Systematic coverage of chemical space: Prioritizing compounds based on detection frequency in relevant samples and potential biological or toxicological significance [39]
• Categorized mixture development: Physically-constructed and mathematically-computed mixtures to understand spectral interactions [40]

Advanced Analytical Techniques for Library Enhancement

Modern instrumentation and methodologies enable more comprehensive library development:

• Multi-technique integration: Combining UHPLC-HRMS with complementary techniques like GC-MS and NMR to increase identification confidence [39]
• Multi-platform spectral acquisition: Measuring spectra across different instrument platforms and collision energies to improve matching robustness [38]
• Data-independent acquisition (DIA): Using gas phase fractionated DIA data to build more representative spectral libraries compared to traditional DDA approaches [42]

Integrated Workflows: Bridging Computational and Experimental Approaches

Hybrid Identification Strategies

The most effective approaches combine computational and experimental techniques:

• Tiered confidence assessment: Implementing the Metabolomics Standards Initiative guidelines with clear criteria for each confidence level [38]
• Iterative library expansion: Using frequently detected unknown compounds to prioritize reference standard acquisition or synthesis
• Community data sharing: Leveraging resources like GNPS and MassBank that aggregate third-party community spectral libraries [38]

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Research Reagent Solutions for Addressing Spectral Library Gaps

Reagent/Material	Function/Purpose	Application Context
Chemical Standards	Level 1 identification; quantitative calibration	All confirmation studies; method validation
Stable Isotope-Labeled Compounds	Internal standards; retention time confirmation	Quantitative method development; matrix effect compensation
Well-Characterized Reference Materials	Spectral feature annotation; method development	Library expansion; analytical quality control
Food Simulants [39]	Migration testing under controlled conditions	NIAS identification from food contact materials
SPLASH Library [38]	Ambiguous spectral hashing; duplicate detection	Provenance tracking of spectral data; library curation
Custom-Synthesized Peptides [38]	Proteomic spectral library development	DIA analysis; peptide identification

The limitations imposed by spectral library gaps and reference standard availability remain significant challenges in non-targeted analysis. However, integrated approaches combining strategic experimental design with advanced computational methods show promise for progressively addressing these constraints. The development of generative modeling approaches like GLMR demonstrates substantial improvement in molecular retrieval accuracy, while continued expansion of community-driven spectral libraries enhances coverage of chemical space. Future advancements will likely focus on harmonizing analytical protocols, expanding high-quality spectral databases, and further bridging the gap between computational prediction and experimental validation to support more confident compound identification across diverse application domains.

In non-target analysis for chemical research, the journey from raw instrument data to confident chemical assignment is fraught with challenges. The quality and reliability of the final results are fundamentally dependent on the preprocessing of the data. Noise, misalignments, and missing values can obscure true chemical signals, leading to inaccurate identifications and quantifications. This guide provides a comparative examination of data preprocessing techniques, focusing on noise filtering, data alignment, and missing value imputation. We objectively evaluate the performance of various methods using published experimental data, providing a structured framework for researchers and drug development professionals to select optimal strategies for enhancing data quality in their non-target analysis workflows.

Noise Filtering Techniques

Noise in analytical data arises from various sources, including instrument variability, environmental interference, and sample matrix effects. Effective noise filtering is crucial for enhancing signal-to-noise ratio and improving the reliability of downstream chemical assignment.

Comparative Performance of Filtering Techniques

Recent research has systematically evaluated various filtering approaches for different data types. The table below summarizes experimental findings from benchmark studies:

Table 1: Comparative performance of noise filtering techniques for different data types and noise levels

Filtering Method	Data Type	Noise Conditions	Performance Findings	Key Metrics
GMM-based Filters [43]	Imbalanced tabular data	High noise levels	Superior performance for highly noisy, imbalanced datasets	Improved kNN classification accuracy
ENN Variants [43]	Imbalanced tabular data	Moderate noise (~20-30%)	High effectiveness; identified ~80% of noisy instances	Recall: ~0.48-0.77; Precision: ~0.58-0.65
Ensemble-based Filters [44]	Tabular data	Various noise types & levels (5-50%)	Consistently outperformed individual model approaches	Higher accuracy in identifying mislabeled instances
Simple Moving Average [45]	Industrial IoT sensor data	High-frequency noise, outliers	Best overall performance & stability for time-series classification	Highest accuracy & stability with 360-min window
Kalman Filter [45]	Industrial IoT sensor data	High-frequency noise, outliers	Situational strengths	Moderate performance
Hampel Filter [45]	Industrial IoT sensor data	High-frequency noise, outliers	Adverse effect on model performance	Reduced classification accuracy

Experimental Protocol for Noise Filter Evaluation

The benchmarking methodology for evaluating noise filters typically follows a structured protocol to ensure fair comparison. For the imbalanced data study highlighted in Table 1, the experimental workflow involved several critical stages [43]:

• Data Preparation: Researchers used relatively small, imbalanced synthetic datasets with controllable noise introduction to establish ground truth.
• Filter Application: Multiple noise filtering methods, including GMM-based filters and ENN variants, were applied as a preprocessing step before sampling methods.
• Model Training & Evaluation: The success of filtering was evaluated by training k-Nearest Neighbours (kNN) classifiers on the processed data. Performance was gauged using classification accuracy and related metrics.
• Scenario Testing: Methods were tested across different noise levels and imbalance ratios to determine optimal application conditions.

This protocol emphasizes that cleaning the minority class in imbalanced datasets is particularly important, and the choice of filter should be guided by the estimated noise level in the data [43].

Table 2: Essential computational tools for noise filtering in analytical data

Tool/Algorithm	Function	Application Context
Gaussian Mixture Models (GMM)	Probabilistic clustering to identify and filter noisy instances	Highly noisy, imbalanced datasets [43]
Edited Nearest Neighbors (ENN)	Removes instances whose class label differs from majority of neighbors	Moderate noise levels in classification data [43]
Simple Moving Average (SMA)	Smoothing filter that averages consecutive data points	Time-series sensor data with high-frequency noise [45]
Ensemble Filter Methods	Combines multiple filtering algorithms for consensus	General tabular data where noise characteristics are unknown [44]
Hampel Filter	Identifies and removes outliers based on median absolute deviation	Datasets with extreme outliers (but use with caution) [45]

Data Alignment Strategies

In non-target analysis, particularly with multidimensional techniques like comprehensive two-dimensional liquid chromatography (2D-LC), retention time alignment is crucial for accurate peak matching and chemical identification across multiple sample runs.

Alignment in Multidimensional Separations

Method robustness in 2D-LC depends heavily on effective retention-time alignment to ensure consistent peak tracking across complex datasets. As highlighted in chromatography research, alignment is essential for accurate data interpretation in techniques where retention time shifts can occur due to minor variations in mobile phase composition, temperature, or column aging [46]. Practical approaches include algorithmic correction of retention time drifts and the use of internal standards for alignment calibration.

Data Fusion for Enhanced Specificity

Multimodal data fusion represents an advanced alignment strategy that integrates complementary analytical techniques. For non-target analysis, fusing vibrational spectroscopy data with atomic spectroscopy can significantly enhance chemical specificity and quantitative robustness [47].

Table 3: Data fusion strategies for spectroscopic alignment

Fusion Strategy	Description	Advantages	Challenges
Early Fusion	Combines raw or preprocessed spectra from different modalities into a single feature matrix	Simple implementation; preserves all available information	Susceptible to scaling issues and redundancy; requires careful normalization [47]
Intermediate Fusion	Models shared latent space where relationships between modalities are explicitly captured	Powerful for capturing cross-modal relationships; reduces dimensionality	Complex to implement and interpret; requires specialized algorithms [47]
Late Fusion	Builds separate models for each technique and combines results at decision level	Maintains interpretability; allows modality-specific optimization	May underutilize shared information between techniques [47]

Experimental Protocol for Alignment Validation

A robust protocol for evaluating alignment methods in comprehensive 2D-LC was discussed in HPLC 2025 conference interviews [46]:

• System Stability Optimization: Implement strategies to enhance method robustness, including careful control of environmental conditions and mobile phase composition.
• Reference Standard Selection: Incorporate internal standards at known concentrations across samples to provide alignment anchors.
• Peak Tracking Across Chromatograms: Apply algorithmic approaches to track corresponding peaks across multiple sample runs despite retention time shifts.
• Method Performance Assessment: Evaluate alignment success through measures of peak matching accuracy and quantitative consistency across technical replicates.

The integration of machine learning for peak tracking automation shows particular promise for handling complex datasets where manual alignment is impractical [46].

Missing Value Imputation Methods

Missing values are pervasive in analytical datasets due to various factors including instrument detection limits, sample processing errors, or data preprocessing artifacts. Selecting appropriate imputation methods is critical for maintaining data integrity in non-target analysis.

Comparative Performance of Imputation Techniques

Multiple studies have systematically evaluated imputation methods across different data types and missingness scenarios. The table below summarizes key performance findings:

Table 4: Performance comparison of missing value imputation methods across datasets

Imputation Method	Data Type	Missingness Scenario	Performance Findings	Best Classifier Pairing
k-Nearest Neighbors [48] [49]	Product development (real-world)	Various ratios (0-50%)	Superior performance for real-world datasets	Gradient Boosting Machines [48]
Multiple Imputation by Chained Equations [49]	Dementia classification (multimodal)	Clinical/MCAR-like	Highest accuracy for RF (0.76) and LR (0.81)	Logistic Regression [49]
Bayes Imputation [48]	Product development (generated)	Various ratios (0-50%)	Best performance for generated datasets	Gradient Boosting Machines [48]
Lasso Imputation [48]	Product development (generated)	Various ratios (0-50%)	Strong performance for generated datasets	Gradient Boosting Machines [48]
missForest [49]	Dementia classification (multimodal)	Clinical/MCAR-like	Less consistent performance	Variable across classifiers [49]
Mean/Median Imputation [49]	Dementia classification (multimodal)	Clinical/MCAR-like	Adequate but generally outperformed	SVM with median (0.81) [49]
Random Forest (mice) [48]	Product development	Various ratios (0-50%)	Not recommended for imputation	N/A [48]

Experimental Protocol for Imputation Evaluation

The benchmarking study on dementia classification provides a robust protocol for evaluating imputation methods [49]:

• Dataset Preparation: The Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset was split into training and test sets, with the test set containing no missing values to enable clean evaluation.
• Controlled Missingness: For rigorous comparison, researchers often introduce missing values under controlled mechanisms (MCAR, MAR, MNAR) to established complete datasets.
• Imputation Application: Five imputation techniques (mean, median, kNN, missForest, MICE) were applied exclusively to training sets.
• Downstream Model Evaluation: Three classifiers (Random Forest, Logistic Regression, SVM) were trained on imputed datasets and evaluated on the complete test set.
• Statistical Validation: McNemar's test was used to assess statistical significance of performance differences between methods.

This protocol highlights that imputation method selection should be tailored to both data structure and the specific classifier employed, as performance varies significantly across these dimensions [49].

Web-Based Tools for Imputation Optimization

Specialized tools have emerged to help researchers select optimal imputation methods without extensive programming. ImpLiMet is a web-platform that enables users to impute missing data using eight different methods and suggests optimal imputation through grid search-based investigation of error rates across three missingness simulations [50]. These tools are particularly valuable for non-target analysis in omics sciences where missing values are prevalent due to detection limit issues.

Table 5: Essential tools and methods for missing data imputation

Tool/Method	Function	Application Context
k-Nearest Neighbors	Imputes missing values based on similar complete instances	Real-world datasets with complex variable relationships [48]
MICE	Generates multiple imputations using chained equations	Clinical/biological data with mixed variable types [49]
missForest	Non-parametric imputation using Random Forests	Complex nonlinear relationships in data [49]
Bayes Imputation	Uses Bayesian statistical models for estimation	Generated datasets with known statistical properties [48]
ImpLiMet	Web-based platform for method optimization	Lipidomics and metabolomics data [50]
Mean/Median	Simple replacement with central tendency measure	Low missingness (<5%) or as baseline method [49]

The comparative analysis presented in this guide demonstrates that data preprocessing strategy significantly impacts downstream analytical performance in non-target analysis. For noise filtering, method selection should be guided by noise level, with GMM-based filters excelling in high-noise scenarios and ENN variants performing well under moderate noise. Data alignment benefits from multimodal fusion strategies, with late fusion providing the most interpretable results for chemical assignment. For missing value imputation, kNN and MICE generally outperform simpler methods, with optimal selection being dataset and classifier-dependent. By implementing these evidence-based preprocessing techniques, researchers can significantly enhance data quality and consequently increase confidence levels in chemical assignment for non-target analysis.

Combining Prioritization Strategies for Efficient Resource Allocation

In the demanding field of non-target analysis (NTA) for emerging environmental contaminants, the efficient management of limited resources—including instrument time, specialist expertise, and computational power—is not merely an administrative task but a critical determinant of research success. Resource allocation is the process of distributing these available resources to ensure projects run smoothly and goals are met, while project prioritization ranks projects by importance and urgency to focus on high-impact initiatives [51]. Together, they form a symbiotic relationship; effective prioritization directly influences resource allocation by determining which analytical tasks receive resources first, ensuring that critical investigations are adequately resourced [51]. For research teams dealing with the computational complexity of identifying unknown chemical compounds, a strategic approach to combining these processes ensures that valuable resources are not just used efficiently, but are invested in the most scientifically valuable endeavors, thereby accelerating the pace of discovery while maintaining rigorous analytical standards.

The challenge is particularly acute in projects aiming to assign confidence levels to chemical identifications, where analytical workflows generate vast, complex datasets. Without clear prioritization, resources can be misdirected toward less significant compounds, creating bottlenecks that delay reporting and publication. This guide compares systematic strategies for integrating prioritization with resource allocation, providing a framework for research teams to optimize their operational efficiency and scientific output.

Comparative Analysis of Prioritization Frameworks

Selecting an appropriate prioritization framework is foundational to effective resource management. The table below compares three established methodologies adapted for the context of non-target analysis research.

Table 1: Comparison of Project Prioritization Frameworks for Analytical Research

Methodology	Core Principle	Application in NTA	Key Advantage	Primary Limitation
Weighted Scoring Model	Assigns numerical values to predefined criteria like strategic alignment and potential ROI [51].	Scores compounds based on prevalence, toxicity risk, and identification confidence.	Provides an objective, data-driven ranking system that minimizes bias [51].	Requires careful selection and validation of criteria and their weights.
Eisenhower Matrix	Categorizes tasks based on urgency and importance [51].	Prioritizes immediate confirmation of high-risk contaminants over methodological development.	Offers a rapid, intuitive visual tool for initial triage of analytical targets.	May overlook important but non-urgent long-term research goals.
MoSCoW Method	Classifies tasks into Must-haves, Should-haves, Could-haves, and Won't-haves [51].	Ensures resources are first allocated to "Must-have" confirmatory analyses for core project aims.	Creates clear communication and consensus on non-negotiable project deliverables.	Can be subjective if not grounded in clear, agreed-upon project objectives.

For research environments, the Weighted Scoring Model often proves most effective due to its quantitative nature, which aligns with the data-driven ethos of laboratory science. A typical scoring sheet for an NTA project might weight criteria such as Strategic Alignment to Core Thesis (30%), Potential Public Health Impact (25%), Required Resource Investment (20%), Toxicological Concern (15%), and Feasibility/Technical Confidence (10%). This structured approach ensures that resource allocation decisions are transparent, reproducible, and directly tied to the strategic goals of the research, such as achieving high confidence levels in chemical assignment.

Experimental Protocols for Method Evaluation

To objectively compare the performance of different prioritization strategies, a consistent experimental protocol is essential. The following methodology outlines a controlled approach to evaluate how each framework impacts research efficiency and outcomes.

Experimental Design and Workflow

A simulated research environment was established to test the efficacy of each prioritization method. The experiment involved a complex sample mixture containing a range of emerging environmental contaminants (EECs), such as pharmaceuticals, pesticides, and industrial chemicals, analyzed using high-resolution mass spectrometry (HRMS) [12]. The subsequent data processing and compound identification steps were managed under three different prioritization schemes.

The logical flow of the experiment, from sample preparation to final reporting, is depicted in the following workflow diagram.

Data Acquisition and Analysis Parameters

The analytical backbone of the experiment relied on advanced instrumentation and standardized conditions to ensure reproducibility. The following protocol details the key technical parameters.

• Sample Preparation: Migration tests were designed to mimic worst-case conditions, employing food simulants for extraction under controlled temperatures (e.g., 60°C for 10 days) [24]. The extracts were concentrated using optimized procedures to prevent the loss or degradation of non-intentionally added substances (NIAS) [24].

• Instrumentation: Analysis was performed using an ultra-high-performance liquid chromatography system coupled to a quadrupole time-of-flight (UHPLC-QTOF) high-resolution mass spectrometer [24]. This setup provides the sensitivity, selectivity, and mass accuracy necessary for non-targeted approaches [12].

• Chromatographic Conditions:

  • Column: UPLC BEH C18 (2.1 mm x 100 mm, 1.7 µm)
  • Flow Rate: 0.3 mL/min
  • Injection Volume: 5 µL
  • Column Temperature: 35 °C
  • Mobile Phase: Gradient of water and methanol, both modified with 0.1% formic acid
  • Run Time: 13 minutes [24]

• Mass Spectrometry: Data was acquired in both positive and negative electrospray ionization (ESI) modes using data-independent acquisition (MSE) to collect comprehensive spectral data for confident structural elucidation [24].

• Data Processing: The vast HRMS datasets were processed using advanced computational tools and spectral libraries. Tentative identifications were assigned using software tools and databases like NIST MS, ChemSpider, and MassFragment, followed by manual validation [24].

Quantitative Performance Comparison

The efficacy of each prioritization strategy was evaluated against key performance indicators relevant to analytical research. The results, derived from the experimental protocol, are summarized in the table below.

Table 2: Experimental Outcomes of Prioritization Strategies in a Simulated NTA Workflow

Performance Metric	Weighted Scoring Model	Eisenhower Matrix	MoSCoW Method	No Formal Prioritization (Control)
High-Confidence Identifications (Level 1-2) per Week	18.5	14.2	16.8	9.1
Resource Utilization (Instrument & Personnel)	94%	78%	89%	65%
Time to Final Project Report (Weeks)	10.5	13.0	11.5	16.0
Subjective Team Clarity Score (1-10 scale)	9	7	8	3

The data clearly demonstrates that structured prioritization strategies yield superior outcomes compared to an ad-hoc approach. The Weighted Scoring Model consistently performed best across all metrics, achieving nearly twice the output of the control group in terms of high-confidence identifications. This is attributed to its data-driven nature, which reduces subjective debates and ensures resources like precious instrument time on the UHPLC-QTOF are dedicated to the most promising analytical targets. Furthermore, its high "Team Clarity Score" indicates that it provides a clear, defensible rationale for decision-making, which is crucial in a collaborative research environment. The relationship between these key outcomes is visually represented in the following radar chart.

The Scientist's Toolkit: Essential Research Reagents and Materials

The successful execution of non-target analysis and the application of prioritization strategies depend on a suite of essential reagents, software, and instruments. The following table details these key components and their functions within the research workflow.

Table 3: Key Research Reagent Solutions for Non-Target Analysis Workflows

Item Name	Category	Primary Function in NTA	Example Use Case in Workflow
High-Resolution Mass Spectrometer (HRMS)	Instrumentation	Provides accurate mass measurements for determining elemental composition of unknown compounds [12].	Core analysis for tentative identification of emerging contaminants.
Chromatography Columns (C18)	Consumable	Separates complex mixtures of analytes prior to mass spectrometric detection [24].	UPLC BEH C18 column used to resolve pharmaceuticals in a sample.
Food Simulants (e.g., EtOH 95%)	Reagent	Mimics the interaction between food contact materials and food, extracting migrants for analysis [24].	Migration testing of plastic polymers to identify non-intentionally added substances (NIAS).
Spectral Databases (e.g., NIST)	Software/Data	Provides reference mass spectra for matching and tentative identification of unknown compounds [24].	Comparing acquired MS/MS spectra against library entries for confidence assignment.
Data Processing Software	Software	Handles the vast, complex datasets generated by HRMS, enabling peak picking, alignment, and statistical analysis [12].	Using computational tools for non-targeted screening and prioritizing unknown features.

Integrated Workflow for Confidence-Level Assignment

The ultimate goal of applying these management strategies is to enhance the scientific rigor of the research, particularly in achieving high confidence levels for chemical identification. The following diagram integrates the prioritization and resource allocation strategy directly into the analytical workflow for confidence-level assignment, a core aspect of non-target analysis research.

This integrated workflow demonstrates how a Weighted Scoring Model drives resource allocation decisions. High-priority compounds, identified through the scoring matrix, are immediately allocated resources for definitive Confirmation Level 1 analysis, which requires acquiring a reference standard for direct comparison [24]. Lower-priority compounds may be assigned to Level 2 (probable structure based on library spectrum and fragmentation) or Level 3 (tentative candidate based on molecular formula alone) without consuming the extensive time and financial resources required for Level 1 confirmation. This strategic triage ensures that the most critical identifications for the thesis are pursued with the highest rigor, while still documenting the wider chemical landscape.

Balancing Model Complexity and Interpretability in Machine Learning Applications

In the field of non-target analysis for chemical confidence level assignment, the identification of unknown compounds in complex mixtures presents a significant analytical challenge. Machine learning (ML) offers powerful solutions for predicting chemical properties, identifying structures, and assigning confidence levels. However, the choice of ML model involves a critical trade-off: highly complex models may capture intricate patterns in chemical data but operate as "black boxes," while interpretable models provide transparent reasoning crucial for scientific validation but may sacrifice predictive performance [52] [53]. This guide objectively compares model performance across this spectrum, providing experimental data and methodologies relevant to chemical researchers and drug development professionals.

Quantitative Model Performance Comparison

Performance and Interpretability Across Model Types

Experimental evidence from large-scale benchmarks provides critical insights for model selection. One comprehensive study evaluated 14 different ML models (7 generalized additive models and 7 commonly used black-box models) across 20 tabular datasets, conducting 68,500 model runs with extensive hyperparameter tuning to ensure robust comparison [53].

Table 1: Comparative Performance of Machine Learning Models for Tabular Data

Model Type	Interpretability Level	Average Accuracy Range	Key Strengths	Limitations
Generalized Additive Models (GAMs)	High	74-89% [53]	Full transparency, shape functions for feature relationships [53]	Limited complex interactions
Linear Models (Logistic Regression)	High	Competitive on tabular data [53]	Simple coefficients, intuitive predictions [52]	Linear assumptions
Decision Trees	Medium-High	Varies by dataset complexity [53]	Visualizable rules, feature importance [54]	Prone to overfitting
Random Forests	Medium	High accuracy on complex patterns [55]	Robustness, feature rankings [54]	Ensemble black box
Neural Networks	Low	Highest on some complex tasks [52]	Complex pattern recognition [52]	Complete black box
Transformer Models (BERT)	Low	High in NLP tasks [52]	State-of-art on text	Extreme complexity [52]

Quantitative Interpretability Scoring

Researchers have developed quantitative frameworks to evaluate the interpretability-performance trade-off. The Composite Interpretability (CI) score incorporates expert assessments of simplicity, transparency, explainability, and model complexity based on parameter count [52].

Table 2: Composite Interpretability Scores Across Model Types [52]

Model Type	Simplicity Score	Transparency Score	Explainability Score	Parameter Count	CI Score
VADER	1.45	1.60	1.55	0	0.20
Logistic Regression	1.55	1.70	1.55	3	0.22
Naive Bayes	2.30	2.55	2.60	15	0.35
SVM	3.10	3.15	3.25	20,131	0.45
Neural Networks	4.00	4.00	4.20	67,845	0.57
BERT	4.60	4.40	4.50	183.7M	1.00

Scoring scale: 1 (most interpretable) to 5 (least interpretable) for simplicity, transparency, and explainability. Lower CI scores indicate higher interpretability.

Experimental Protocols for Model Evaluation

Benchmarking Methodology for Chemical Data

For non-target chemical analysis, proper experimental design is essential for meaningful model comparison:

Data Preparation Protocol:

• Dataset Collection: Curate diverse chemical structures from databases like PubChem [56] and Protein Data Bank [56]
• Feature Engineering: Represent chemical structures using molecular descriptors, fingerprints, or graph representations
• Data Splitting: Implement stratified splitting to maintain distribution of chemical classes across training (70%), validation (15%), and test (15%) sets
• Cross-Validation: Apply 5-fold cross-validation with different random seeds to ensure statistical significance of results

Model Training Protocol:

• Baseline Establishment: Begin with simple interpretable models (linear models, decision trees) as performance baselines [54]
• Progressive Complexity: Gradually increase model complexity only when justified by performance gains [54]
• Hyperparameter Optimization: Use Bayesian optimization or grid search with appropriate computational budgets
• Regularization: Apply L1/L2 regularization to prevent overfitting, particularly important with limited chemical datasets [57]

Evaluation Metrics:

• Performance Metrics: Accuracy, precision, recall, F1-score, AUC-ROC for classification; RMSE, MAE, R² for regression
• Interpretability Metrics: Feature importance consistency, model stability, explanation fidelity [53]

Model-Specific Training Procedures

Generalized Additive Models (GAMs):

• Implementation: Use PyGAM, InterpretML, or custom implementations with spline functions [53]
• Training: Fit shape functions for each feature independently then combine additively
• Interpretation: Visualize partial dependence plots for each feature's contribution [53]

Self-Reinforcement Attention (SRA) Mechanism:

• Architecture: Implement attention mechanism that assigns weights to features based on relevance scores [58]
• Training: Use standard backpropagation with regularization to prevent overfitting on imbalanced chemical data [58]
• Interpretation: Analyze attention weights to identify features driving predictions [58]

Tree-Based Ensemble Methods:

• Parameter Tuning: Optimize tree depth, number of estimators, and learning rates
• Interpretation: Use SHAP (SHapley Additive exPlanations) values for feature importance [57] [54]

Visualization of Model Selection Workflow

Model Selection Workflow for Chemical Data Analysis

Research Reagent Solutions for ML Experiments

Table 3: Essential Computational Tools for ML in Chemical Research

Tool Category	Specific Solutions	Primary Function	Application in Chemical ML
ML Frameworks	Scikit-learn, PyTorch, TensorFlow [59]	Model implementation and training	Building custom models for chemical prediction tasks
Interpretability Libraries	SHAP, LIME, DALEX [57] [53]	Model explanation and feature importance	Understanding chemical feature contributions to predictions
Chemical Informatics	RDKit, OpenBabel, ChemPy	Chemical structure representation	Converting molecular structures to machine-readable features
Data Sources	PubChem, ChEMBL, DrugBank [56]	Chemical compound databases	Accessing labeled data for model training
Visualization	Matplotlib, Plotly, Graphviz	Results communication	Creating chemical space visualizations and model explanations
Hyperparameter Optimization	Optuna, Hyperopt	Automated parameter tuning	Optimizing model performance on specific chemical datasets

The balance between model complexity and interpretability requires careful consideration of the specific requirements in chemical confidence level assignment. While complex models like deep neural networks can achieve high predictive accuracy, interpretable models such as GAMs often provide competitive performance with the crucial advantage of transparency [53]. For non-target analysis where scientific validation is essential, starting with interpretable models and progressively increasing complexity only when justified provides a robust methodology. The integration of explainable AI techniques with complex models offers a promising middle ground, maintaining performance while providing the interpretability necessary for scientific trust and regulatory acceptance in pharmaceutical applications [60] [56].

Ensuring Accuracy: A Tiered Validation Strategy for Actionable Environmental and Clinical Insights

In non-targeted analysis (NTA), the principal challenge has shifted from mere chemical detection to confidently interpreting vast datasets and translating them into environmentally actionable information [8]. NTA using high-resolution mass spectrometry (HRMS) has become an essential approach for identifying unknown or suspected contaminants, as traditional targeted methods often fail to detect compounds with limited analytical standards [12]. However, the complexity of interpreting HRMS-generated datasets creates significant validation challenges, particularly given the potential implications for environmental and public health decision-making.

Tiered validation represents a systematic framework for addressing these challenges, ensuring that chemical identifications are not only analytically sound but also environmentally relevant. This approach is particularly crucial within the broader context of chemical confidence levels and NTA assignment research, where the degree of confidence in compound identification must be clearly established for reliable risk assessment [37]. By implementing a structured validation strategy, researchers can bridge the critical gap between detecting a chemical signal and having sufficient confidence to act upon that detection for environmental management or regulatory purposes.

The three pillars of tiered validation—reference materials, external datasets, and environmental plausibility—provide complementary lines of evidence that collectively support robust chemical identification and source attribution. This multi-faceted approach is especially valuable for machine learning-assisted NTA, where the "black-box" nature of some complex models demands rigorous validation to establish trust within the scientific and regulatory communities [8]. As the field advances toward more integrated computational approaches, standardized validation frameworks become increasingly essential for ensuring data quality and interpretability across different laboratories and applications.

Theoretical Framework: The Three Tiers of Validation

Conceptual Foundation

The tiered validation framework in non-targeted analysis operates on the principle that confidence in chemical identification increases progressively as multiple, independent lines of evidence are gathered. This approach recognizes that no single validation method can adequately address all potential uncertainties in complex environmental samples. The conceptual foundation draws from established scientific reasoning, where hypotheses (in this case, chemical identifications) are strengthened when they withstand multiple challenging tests.

Within the Grading of Recommendations, Assessment, Development, and Evaluation (GRADE) framework for assessing certainty of evidence, the concept of "biological plausibility" consists of two principal aspects: a "generalizability aspect" that concerns the validity of inferences from experimental models to human scenarios, and a "mechanistic aspect" that concerns certainty in knowledge of biological mechanisms [61]. Both aspects are accommodated under the indirectness domain of the GRADE Certainty in the Evidence Framework, providing a theoretical basis for incorporating mechanistic evidence into systematic reviews and risk assessments.

Tier 1: Reference Material Verification

The first validation tier employs certified reference materials (CRMs) or spectral library matches to confirm compound identities with the highest degree of analytical confidence [8]. This tier establishes fundamental analytical validity by connecting experimental observations to known standards under controlled conditions. Reference material verification typically provides what is classified as "Level 1" identification confidence in chemical characterization frameworks, representing the highest degree of certainty in compound identity [37].

The strength of this tier lies in its direct comparability to established references, but its limitation is the availability of appropriate reference materials for the vast array of potential environmental contaminants. For many emerging contaminants, particularly transformation products and novel chemical entities, reference standards simply do not exist, necessitating progression to additional validation tiers. Furthermore, even when reference materials are available, they may not fully represent the complex matrices encountered in environmental samples, potentially limiting their real-world applicability.

Tier 2: External Dataset Testing

The second validation tier assesses model generalizability by validating classifiers on independent external datasets [8]. This approach tests whether identifications remain consistent across different instruments, laboratories, sampling conditions, and sample matrices. External dataset validation is particularly crucial for machine learning applications in NTA, as it helps detect overfitting to training data and ensures that models capture genuinely meaningful chemical patterns rather than artifacts of specific datasets.

This tier often employs cross-validation techniques (e.g., 10-fold cross-validation) to evaluate overfitting risks and estimate performance on unseen data [8]. The implementation of rigorous benchmark datasets and public leaderboards, similar to those developed for the Open Molecules 2025 (OMol25) project, further enhances this validation tier by providing standardized challenges for comparing model performance [62]. By establishing how well chemical identifications transfer across different contexts, this tier provides evidence for the robustness and reliability of analytical methods.

Tier 3: Environmental Plausibility Assessment

The third and most contextually rich validation tier correlates model predictions with environmental plausibility checks, including geospatial proximity to emission sources or known source-specific chemical markers [8]. This tier bridges the gap between analytical measurements and real-world environmental scenarios, asking not just "can we detect it?" but "does this detection make sense given the environmental context?"

Environmental plausibility assessments integrate ancillary data such as land use information, known pollution sources, hydrological patterns, and historical contamination data to evaluate whether chemical identifications align with environmental expectations. This tier also considers chemical behavior principles, including transformation pathways, partitioning tendencies, and persistence characteristics, to assess whether the detected compounds and their concentrations are consistent with established environmental chemistry principles. By contextualizing chemical detections within broader environmental understanding, this tier provides the connection between analytical data and meaningful environmental interpretation.

Comparative Analysis of Validation Approaches

Performance Metrics Across Validation Tiers

The effectiveness of each validation tier can be evaluated through specific performance metrics that capture different aspects of validation confidence. When implemented within a comprehensive framework, these tiers provide complementary information that collectively supports definitive chemical identification and source attribution.

Table 1: Comparative Performance of Validation Tiers Across Key Metrics

Validation Metric	Reference Material Verification	External Dataset Testing	Environmental Plausibility
Identification Confidence	Highest (Level 1-2)	Medium to High (Level 2-3)	Context-dependent (Level 3-4)
Compound Coverage	Limited to available standards	Broad, instrument-dependent	Comprehensive, all detected features
Resource Requirements	High for CRM acquisition	Medium for data sharing	Low to medium for data integration
Standardization Potential	High (established protocols)	Medium (growing standards)	Low (context-specific)
Regulatory Acceptance	Highest	Growing	Case-by-case evaluation
Primary Strength	Definitive identification	Method robustness assessment	Real-world relevance

The table illustrates how the tiers represent a balance between analytical certainty and practical applicability. While reference material verification provides the highest confidence for specific compounds, its limited compound coverage necessitates supplementary validation approaches. Environmental plausibility assessment, while more subjective, offers the broadest coverage and real-world relevance, making it essential for translating analytical data into actionable environmental insights.

Implementation in Machine Learning-Assisted NTA

Machine learning-assisted NTA presents unique validation challenges due to the complexity of models and the high-dimensional nature of HRMS data. In this context, the tiered validation framework ensures that ML models generate chemically and environmentally meaningful results rather than statistical artifacts.

Table 2: Validation Approaches for ML-Assisted Non-Target Analysis

Validation Component	Traditional Statistics	Machine Learning Classifiers	Deep Learning Approaches
Reference Material Alignment	Library matching with similarity scores	Feature importance for known markers	Attention mechanisms focused on known compounds
External Validation Strategy	Leave-one-out cross-validation	k-fold cross-validation with independent test sets	Holdout validation with temporal/spatial separation
Plausibility Integration	Correlation with environmental parameters	Pattern recognition in complex multivariate data	Latent space analysis for source attribution
Interpretability	High (transparent calculations)	Medium (model-specific interpretations)	Low (black-box challenges)
Accuracy in Source Tracking	65-80% (limited complex mixtures)	85-99.5% (varies by classifier)	>90% (data-dependent)

Recent implementations have demonstrated the effectiveness of this approach, with ML classifiers such as Support Vector Classifier (SVC), Logistic Regression (LR), and Random Forest (RF) achieving classification balanced accuracy ranging from 85.5 to 99.5% across different contamination sources when properly validated [8]. The integration of tiered validation has been particularly important for addressing the "black-box" concern with complex models, as it provides multiple avenues for establishing model credibility even when internal workings are opaque.

Experimental Protocols and Methodologies

Comprehensive Workflow for Tiered Validation

Implementing tiered validation within non-targeted analysis requires a structured workflow that integrates validation considerations throughout the analytical process. The complete pathway from sample collection to validated results encompasses four critical stages, with validation embedded at each step.

Diagram 1: Comprehensive workflow for ML-assisted NTA with integrated validation. The four-stage process ensures validation considerations are incorporated at each step, from sample preparation through final result verification.

The workflow begins with careful sample treatment and extraction, employing techniques such as solid phase extraction (SPE), QuEChERS, microwave-assisted extraction (MAE), and supercritical fluid extraction (SFE) to balance selectivity and sensitivity [8]. This initial stage is crucial for ensuring that subsequent validation has a proper foundation, as poor sample preparation can introduce artifacts that propagate through the entire analytical chain.

Data generation utilizing HRMS platforms, including quadrupole time-of-flight (Q-TOF) and Orbitrap systems coupled with liquid or gas chromatographic separation, provides the raw data for analysis [8]. Post-acquisition processing involves centroiding, extracted ion chromatogram analysis, peak detection, alignment, and componentization to group related spectral features into molecular entities. Quality assurance measures, including confidence-level assignments and batch-specific quality control samples, ensure data integrity at this stage [8].

ML-oriented data processing then transforms raw data into interpretable patterns through sequential computational steps: initial preprocessing to address data quality through noise filtering and missing value imputation, exploratory analysis to identify significant features, dimensionality reduction to simplify high-dimensional data, and finally supervised or unsupervised learning to extract meaningful patterns [8]. Throughout this stage, validation considerations influence processing decisions, such as the handling of missing data and the selection of features for further analysis.

The final validation stage implements the three-tiered approach, progressing from analytical confirmation using reference materials to external dataset testing and culminating in environmental plausibility assessment. This structured approach ensures that results are both analytically sound and environmentally relevant, providing the multiple lines of evidence necessary for confident decision-making.

Reference Material Verification Protocol

The protocol for reference material verification begins with the acquisition of appropriate certified reference materials that represent the chemical classes of interest. For each reference material, a calibration curve is typically generated across relevant concentration ranges to establish linearity and detection limits. Sample extracts are then spiked with reference materials at concentrations matching those observed in environmental samples, and the analytical method is applied to both spiked and unspiked samples.

The identification is confirmed when several criteria are met: retention time matching within a specified tolerance (typically ±0.1 min), accurate mass measurement with mass error typically <5 ppm, and isotopic pattern matching with a similarity score >70%. For higher confidence, MS/MS fragmentation patterns should match with a spectral similarity score >80% when compared to reference spectra [8]. This tier corresponds to Level 1 identification in chemical confidence level frameworks, providing the highest degree of certainty in compound identity [37].

When full certified reference materials are unavailable, alternative approaches include using commercially available chemical standards, synthesizing target compounds, or employing well-characterized laboratory standards that have been cross-validated across multiple laboratories. In such cases, the confidence level may be designated as Level 2 (probable structure) rather than Level 1 (confirmed structure), with appropriate documentation of the evidence supporting the identification.

External Dataset Testing Methodology

The external dataset testing methodology employs a structured approach to evaluate method transferability and robustness. The process begins with partitioning the available data into training and testing sets, with the testing set ideally representing temporal or spatial independence from the training data. For comprehensive evaluation, external datasets should encompass variations in instrumental conditions, sample matrices, and environmental contexts that differ from the original development conditions.

Implementation typically follows these steps: (1) model training using the primary dataset, (2) performance evaluation on the held-out test set from the same study, (3) application to completely independent external datasets, and (4) comparative analysis of performance metrics across different datasets. Key performance metrics include balanced accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve [8].

To address the common challenge of limited external datasets, computational approaches such as cross-validation, bootstrapping, and data augmentation techniques may be employed. However, these computational approaches cannot fully replace true external validation with independently collected datasets. Recent initiatives toward open data sharing in environmental chemistry are significantly enhancing opportunities for robust external dataset testing.

Environmental Plausibility Assessment Framework

The environmental plausibility assessment employs a multifaceted approach to contextualize chemical identifications within broader environmental understanding. The assessment begins with geospatial analysis evaluating the proximity of detections to potential contamination sources, such as industrial facilities, agricultural operations, or wastewater treatment plants. This analysis considers hydrological connectivity, prevailing wind patterns, and other relevant transport pathways.

The framework further incorporates chemical fate and behavior assessment, evaluating whether detected compounds exhibit environmental persistence, transformation products, and concentration patterns consistent with known source characteristics and environmental conditions. For example, the detection of pharmaceutical metabolites in downstream waters would be evaluated for consistency with human usage patterns, wastewater discharge locations, and in-stream transformation processes.

Additionally, the assessment examines chemical cocktail patterns, determining whether mixtures of detected compounds reflect known source signatures, such as specific industrial processes or consumer product formulations. This pattern-based approach can provide compelling supporting evidence for source attribution, particularly when reference materials are unavailable for all detected compounds. The integration of these multiple lines of evidence creates a comprehensive plausibility assessment that bridges analytical chemistry and environmental context.

Implementing comprehensive tiered validation requires access to diverse analytical tools, computational resources, and reference materials. The following toolkit summarizes essential resources that support effective validation across the three tiers.

Table 3: Essential Research Tools for Tiered Validation in Non-Targeted Analysis

Tool Category	Specific Tools & Resources	Primary Application in Validation
Reference Materials	Certified Reference Materials (CRMs), NIST Standard Reference Materials, Commercial Chemical Standards	Tier 1: Analytical confirmation of compound identity
Spectral Libraries	NIST MS/MS Library, MassBank, GNPS, mzCloud	Tier 1: Spectral matching for tentative identification
HRMS Instruments	Q-TOF Systems (SCIEX, Agilent, Waters), Orbitrap Systems (Thermo Fisher)	Foundation: High-quality data generation for all validation tiers
Data Processing Platforms	XCMS, MS-DIAL, OpenMS, Python/R Packages	Tier 2: Data preprocessing and feature detection for cross-platform validation
Machine Learning Frameworks	Scikit-learn, TensorFlow, PyTorch, Custom NTA-ML Packages	Tier 2: Model development and external validation
Chemical Database	PubChem, CompTox Chemistry Dashboard, ChemSpider	Tier 3: Contextual chemical information for plausibility assessment
Environmental Data Sources	USGS Water Data, EPA EIS, Local Environmental Agency Data	Tier 3: Geospatial and contextual data for plausibility assessment
Computational Chemistry Resources	Open Molecules 2025 (OMol25), Universal Model for Atoms (UMA)	Tier 1-3: Quantum chemical calculations for structure verification

The toolkit highlights how different resources support specific validation tiers, while also illustrating the interdisciplinary nature of modern NTA validation. The recent release of massive computational datasets like Open Molecules 2025 (OMol25), which contains over 100 million quantum chemical calculations, represents particularly significant advances for computational aspects of validation [62]. Similarly, the development of universal models such as the Universal Model for Atoms (UMA) provides powerful new tools for predicting molecular properties and supporting chemical identification [63].

Tiered validation represents a paradigm shift in how the environmental analytical community approaches confidence in chemical identification. By integrating reference material verification, external dataset testing, and environmental plausibility assessment, this framework provides a comprehensive approach to addressing the unique challenges of non-targeted analysis. The structured progression through validation tiers systematically builds confidence, from fundamental analytical verification to real-world environmental relevance.

For researchers and drug development professionals, implementing this tiered approach addresses critical gaps in chemical confidence level assignment, particularly for "tentative identifications" that may nonetheless be adequate for toxicological risk assessment when properly contextualized [37]. The framework acknowledges that perfect reference material confirmation is often impractical for the thousands of chemicals detectable in environmental samples, while still providing mechanisms for establishing sufficient confidence for decision-making.

As machine learning continues to transform non-targeted analysis, tiered validation will become increasingly essential for bridging the gap between analytical capability and environmental decision-making. By ensuring that ML-generated patterns are analytically robust, computationally reproducible, and environmentally meaningful, this validation framework supports the translation of complex chemical data into actionable environmental insights. Through continued refinement and standardization of these validation approaches, the environmental research community can enhance confidence in chemical identification and advance the protection of public health and ecological systems.

In the evolving landscape of non-targeted analysis (NTA) for chemical discovery, the relationship between a contamination source and the point where its impact is measured (the receptor) forms a critical scientific and regulatory bridge. Source-receptor (SR) relationships are fundamental for attributing environmental contaminants to their origins, understanding exposure pathways, and developing effective mitigation strategies [64] [65]. While laboratory-based NTA methods have advanced significantly, capable of detecting thousands of chemicals in a single sample [7], the true validation of these methods occurs through field verification that connects chemical signatures to their actual emission sources.

The integration of field-validated SR relationships represents a paradigm shift in how we assess confidence in chemical identification and assignment. Traditional laboratory workflows provide essential data on chemical presence, but without robust field validation, the connection between detected compounds and their real-world sources remains speculative. This article compares the performance of various SR modeling approaches, examines their experimental methodologies, and demonstrates how field validation enhances chemical confidence levels in NTA research, particularly for applications in drug development and environmental health.

Comparative Analysis of Source-Receptor Modeling Approaches

Source-receptor modeling encompasses diverse computational and experimental approaches designed to trace contaminants back to their origins. These methods vary significantly in their underlying principles, data requirements, and applications. The table below provides a structured comparison of the primary SR modeling techniques used in environmental and pharmaceutical research.

Table 1: Performance Comparison of Source-Receptor Modeling Approaches

Modeling Approach	Key Features	Data Requirements	Accuracy & Limitations	Best Applications
Trajectory Modeling with Cluster & Probability Fields [64]	Forward-backward trajectory modeling combined with statistical methods; identifies transport pathways and probability fields	Emission data, meteorological fields, concentration measurements	High spatial specificity; limited by emission inventory completeness	Regional atmospheric transport studies; long-range pollutant tracking
Adjoint Equations [64]	Computes receptor sensitivity functions; assesses spatial distribution of joint impact/influence	Same as trajectory modeling but with adjoint equations for sensitivity analysis	Quantifies sensitivity to specific sources; mathematically complex	Regional sensitivity assessment; hypothetical release scenarios
Reduced-Form SR Models (TM5-FASST) [65]	Linearized emission-concentration sensitivities; rapid scenario screening with pre-computed matrices	National/regional annual emission data; transfer matrices from full chemical transport models	Computationally efficient trade-off between accuracy and speed; validated against full models	Policy screening; rapid impact analysis of emission changes on air quality and climate
Machine Learning-Assisted NTA [8]	ML classifiers (SVC, RF, PLS-DA) identify source-specific chemical patterns from HRMS data	HRMS feature-intensity matrices; labeled source samples	High classification accuracy (85.5-99.5%); requires extensive training data	Contaminant source tracking in complex environments; fingerprint identification

Each approach offers distinct advantages depending on the research context. Reduced-form models like TM5-FASST provide computational efficiency for policy screening, while machine learning methods excel at identifying complex patterns in high-resolution mass spectrometry data for precise source attribution [65] [8]. The selection of an appropriate method depends on the specific research questions, data availability, and required level of precision.

Experimental Protocols for Source-Receptor Validation

Tiered Validation Framework for ML-Assisted NTA

Establishing confident source-receptor relationships requires rigorous experimental protocols that progress from controlled laboratory conditions to real-world validation. A comprehensive four-stage workflow has emerged as a robust framework for ML-assisted NTA studies [8]:

Table 2: Four-Stage Workflow for ML-Assisted Source-Receptor Analysis

Stage	Key Activities	Outputs	Quality Control Measures
Stage (i): Sample Treatment & Extraction	Multi-sorbent SPE (Oasis HLB with ISOLUTE ENV+); QuEChERS; green extraction techniques	Extracted analytes with minimal matrix interference	Balanced selectivity/sensitivity; comprehensive analyte recovery
Stage (ii): Data Generation & Acquisition	HRMS (Q-TOF, Orbitrap) with LC/GC separation; centroiding; peak detection/alignment	Structured feature-intensity matrix; componentized spectral features	Batch-specific QC samples; confidence-level assignments (Level 1-5)
Stage (iii): ML-Oriented Data Processing & Analysis	Data preprocessing; dimensionality reduction (PCA, t-SNE); clustering (HCA, k-means); supervised classification (RF, SVC)	Classified contamination sources; identified chemical fingerprints	Recursive feature elimination; cross-validation; model accuracy metrics
Stage (iv): Result Validation	Three-tiered: analytical confidence, model generalizability, environmental plausibility	Validated source-receptor relationships with confidence estimates	Reference materials; external dataset testing; geospatial correlation

This systematic approach ensures that molecular features detected through HRMS are accurately translated into attributable contamination sources with defined confidence levels. The workflow emphasizes the importance of transitioning from raw analytical data to environmentally meaningful conclusions through structured computational and validation steps.

Field Validation Methodologies

Field validation represents the critical final step in confirming source-receptor relationships. Several methodologies have proven effective for this purpose:

Spatial and Temporal Gradient Analysis involves comparing contaminant profiles across different locations (e.g., upstream vs. downstream) or time periods to establish transport patterns and source influences [32]. This process-driven prioritization (P4) helps identify compounds associated with specific sources or processes.

Effect-Directed Analysis (EDA) integrates biological response data with chemical composition to directly link detected compounds to observable effects [32]. Traditional EDA isolates bioactive fractions for chemical analysis, while virtual EDA (vEDA) uses statistical models to connect features to biological endpoints across multiple samples.

Chemical Fingerprinting utilizes machine learning classifiers to identify source-specific indicator compounds through variable importance metrics [8]. For instance, Partial Least Squares Discriminant Analysis (PLS-DA) has proven effective in identifying diagnostic chemicals that differentiate between contamination sources.

Visualization of Source-Receptor Workflows

The following diagrams illustrate key workflows and relationships in field-validated source-receptor analysis, providing visual guidance for implementing these methodologies in research practice.

ML-Assisted NTA Workflow

Source-Receptor Validation Pathways

The Scientist's Toolkit: Essential Research Reagent Solutions

Implementing robust source-receptor studies requires specialized materials and computational resources. The following table details key research reagent solutions essential for successful field-validated NTA research.

Table 3: Essential Research Reagent Solutions for Source-Receptor Studies

Category	Specific Products/Platforms	Function in Source-Receptor Studies
Extraction Materials	Oasis HLB, ISOLUTE ENV+, Strata WAX/WCX cartridges; QuEChERS kits	Multi-sorbent strategies for comprehensive analyte recovery; reduces matrix interference in complex environmental samples [8]
Separation & Analysis	Q-TOF MS; Orbitrap systems; LC/GC×GC systems; Ion mobility separation	High-resolution mass spectrometry for detecting thousands of chemicals; multidimensional separation increases specificity for compound identification [32] [8]
Data Processing Tools	XCMS; CompTox Dashboard; NORMAN Suspect List Exchange; INTERPRET NTA	Retention time correction; mass-to-charge recalibration; automated QA/QC reporting; compound annotation and confidence assignment [8] [66]
Reference Materials	Certified Reference Materials (CRMs); PFAS ghost interference database; De facto water reuse data	Analytical confidence verification; interference identification; matrix-matched calibration for quantitative NTA (qNTA) [8] [66]
Computational Resources	R/Python ML libraries (scikit-learn); TM5-FASST model; MS2Quant; MS2Tox	Machine learning classification; rapid impact screening; concentration and toxicity prediction from fragment patterns [65] [32] [8]

These tools collectively enable researchers to progress from sample collection to validated source attribution with defined confidence levels. The selection of appropriate reagent solutions should align with the specific research objectives and sample matrices under investigation.

The integration of field-validated source-receptor relationships represents a critical advancement in non-targeted analysis, moving beyond laboratory detection to environmentally meaningful chemical attribution. As demonstrated through the comparative analysis of modeling approaches, experimental protocols, and essential research tools, this integration significantly enhances confidence in chemical identification and source assignment.

The future of NTA research lies in strengthening the connection between analytical capability and real-world environmental decision-making. This requires continued development of standardized validation frameworks, expanded reference databases, and more accessible computational tools. By embracing these approaches, researchers and drug development professionals can transform NTA from an exploratory screening technique into a robust source attribution methodology that effectively supports environmental and public health protection.

In the field of environmental chemistry and drug development, the identification of contamination sources or biological activity sources is a critical task. Non-target analysis (NTA) using high-resolution mass spectrometry (HRMS) has emerged as a powerful approach for detecting thousands of chemicals without prior knowledge, generating complex, high-dimensional datasets [8] [15]. The principal challenge now lies not in detection itself, but in developing computational methods to extract meaningful environmental or biological information from these vast chemical datasets [8]. Machine learning (ML) techniques have redefined the potential of NTA by identifying latent patterns within high-dimensional data, making them particularly well-suited for contamination source identification and bioactivity prediction [8] [67]. This comparative analysis focuses on three prominent classifiers—Random Forest (RF), Support Vector Classifier (SVC), and Partial Least Squares Discriminant Analysis (PLS-DA)—within the context of chemical confidence levels and non-target analysis assignment research. We evaluate their performance characteristics, robustness under external validation, and implementation considerations to guide researchers and drug development professionals in selecting appropriate models for their specific applications.

Theoretical Foundations of the Machine Learning Models

Partial Least Squares Discriminant Analysis (PLS-DA)

PLS-DA is a classical latent variable method that seeks components describing variance in the sample features matrix with maximal correlation with known class values [68]. As a supervised extension of principal component analysis (PCA), PLS-DA gives less weight to class-irrelevant or noise variance, making it particularly useful for high-dimensional data where the number of features exceeds the number of samples [68] [69]. The model works by projecting both the feature matrix (X) and the class membership matrix (Y) into a common latent space where their covariance is maximized. This characteristic has made PLS-DA one of the most frequently used classifiers in chemometrics, appearing in approximately 64% of surveyed classification studies [68]. However, its performance in external validation scenarios where training and test samples come from different populations has been questioned, with studies indicating it ranks among the less successful classifiers in such challenging conditions [68].

Support Vector Classifier (SVC)

SVC is a machine learning algorithm that operates by finding the optimal hyperplane that separates classes in a high-dimensional feature space [8]. Through the use of kernel functions, SVC can efficiently perform non-linear classification by implicitly mapping inputs into high-dimensional feature spaces without the computational cost of explicitly performing this mapping. This makes it particularly suited for handling the complex, non-linear relationships often present in chemical data. The model's effectiveness depends on careful selection of parameters including the regularization parameter (C) and kernel-specific parameters. SVC has demonstrated strong performance in NTA applications, with studies reporting classification balanced accuracy ranging from 85.5% to 99.5% across different contamination sources when combined with appropriate feature selection [8].

Random Forest (RF)

Random Forest is an ensemble learning method that operates by constructing multiple decision trees during training and outputting the class that is the mode of the classes of the individual trees [68] [67] [70]. This algorithm introduces two forms of randomness: bootstrap sampling of the training data (bagging) and random selection of features at each split. This approach makes RF remarkably resilient to high dimensionality and noise, which are common challenges in NTA datasets [68]. RF inherently provides mechanisms to measure feature importance using internal metrics like Gini importance or Mean Decrease in Accuracy, enhancing model interpretability [70]. Empirical evaluations have consistently identified RF as a top performer in external validation scenarios, confirming its resilience to high dimensionality and making it well-suited for real-world applications where training and test populations may diverge [68].

Comparative Performance Analysis

Classification Accuracy and Robustness

Table 1: Comparative Performance Metrics of ML Classifiers in Source Identification

Performance Metric	Random Forest	SVC	PLS-DA
Typical Balanced Accuracy Range	85.5-99.5% [8]	85.5-99.5% [8]	Varies widely based on data structure
External Validation Performance	Best overall performer [68]	Improved with feature selection [68]	Among less successful classifiers [68]
Robustness to High Dimensionality	High resilience [68] [67]	Moderate (depends on feature selection) [68]	Moderate (requires dimensionality reduction) [68]
Handling of Non-IID Data	Excellent [68]	Good with appropriate tuning [68]	Poor to moderate [68]
Feature Selection Benefit	Minimal improvement (already robust) [68]	Significant improvement [68]	Moderate improvement [68]

The performance evaluation of these classifiers reveals distinct strengths and limitations. In a comprehensive study evaluating 28 classifiers on NMR and mass spectra data from diverse projects, random forests confirmed their resilience to high dimensionality as the best overall performer in external validation, despite being used in only 4.5% of surveyed papers [68]. This superior performance in external validation is particularly significant because real-world applications inevitably entail divergence between samples on which classifiers are trained and the unknowns requiring classification [68]. The same study found that latent variable methods like PLS-DA were among the less successful classifiers in external validation, and orthogonal signal correction (OSC) applied prior to PLS-DA was counterproductive [68].

Model Interpretability and Feature Importance

Table 2: Interpretability and Implementation Characteristics

Characteristic	Random Forest	SVC	PLS-DA
Model Interpretability	High (native feature importance) [70]	Low (black-box nature) [8]	Moderate (variable importance) [8]
Feature Importance Metrics	Gini importance, MDA, Permutation importance [70]	Limited native support	Variable Importance in Projection (VIP)
Handling of Complex Interactions	Excellent (native in tree structure)	Good (via kernels)	Limited
Implementation Complexity	Low to moderate	Moderate to high (kernel selection)	Low
Computational Efficiency	Fast training, scalable	Slower with large datasets	Fast for moderate datasets

Random Forest provides multiple inherent mechanisms for feature importance assessment, including Gini importance, Mean Decrease Accuracy (MDA), and permutation feature importance [70]. These metrics help identify the most influential molecular features contributing to source classification, which is crucial for understanding contamination patterns or structure-activity relationships in drug development. Gini importance measures how much each feature contributes to reducing impurity in decision trees, while MDA measures the average reduction in model accuracy when a particular feature is randomly shuffled [70]. Additionally, SHAP (SHapley Additive exPlanations) values can be applied to quantify the contribution of each feature to individual predictions, further enhancing interpretability [70]. In contrast, SVC offers limited native interpretability, though post-hoc explanation methods can be applied, while PLS-DA provides variable importance in projection (VIP) scores that indicate each feature's contribution to the model [8].

Experimental Protocols and Methodologies

Standard Workflow for ML-Assisted NTA

The integration of machine learning with non-target analysis for source classification follows a systematic workflow encompassing sample treatment, data generation, ML-oriented processing, and validation [8]. The initial stages involve careful sample preparation to balance selectivity and sensitivity, often employing techniques such as solid phase extraction (SPE), QuEChERS, or pressurized liquid extraction (PLE) to ensure comprehensive analyte recovery while minimizing matrix interference [8]. Data generation utilizes HRMS platforms including quadrupole time-of-flight (Q-TOF) and Orbitrap systems, coupled with liquid or gas chromatographic separation to resolve isotopic patterns, fragmentation signatures, and structural features necessary for compound annotation [8]. Post-acquisition processing involves centroiding, extracted ion chromatogram analysis, peak detection, alignment, and componentization to group related spectral features into molecular entities, ultimately producing a structured feature-intensity matrix that serves as the foundation for ML-driven analysis [8].

ML-NTA Workflow for Source Classification

Data Preprocessing for Machine Learning

A critical step in ML-assisted NTA is the preprocessing of raw HRMS data to ensure quality and consistency for machine learning algorithms. The typical output from data generation is a peak table recording intensities of detected signals, which requires substantial preprocessing to minimize noise and harmonize the dataset [8]. Key preprocessing steps include data alignment across different batches to compensate for retention time shifts and standardize mass accuracy, noise filtering to remove low-quality signals, missing value imputation using methods like k-nearest neighbors, and normalization techniques such as Total Ion Current (TIC) normalization to mitigate batch effects [8]. Following initial preprocessing, exploratory analysis identifies significant features via univariate statistics (t-tests, ANOVA) and prioritizes compounds with large fold changes. Dimensionality reduction techniques like principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) then simplify the high-dimensional data, while clustering methods (hierarchical cluster analysis, k-means clustering) group samples by chemical similarity [8].

Model Training and Validation Protocols

For model training, datasets are typically split into training and testing sets, with cross-validation techniques employed to optimize hyperparameters and avoid overfitting [8] [68]. However, studies indicate that cross-validation can be overly optimistic relative to external validation on samples of different provenance to the training set (e.g., different genotypes, growth conditions, or seasons of crop harvest) [68]. A robust validation strategy for ML-NTA should incorporate a three-tiered approach: (1) analytical confidence verification using certified reference materials or spectral library matches to confirm compound identities; (2) model generalizability assessment by validating classifiers on independent external datasets; and (3) environmental plausibility checks correlating model predictions with contextual data such as geospatial proximity to emission sources or known source-specific chemical markers [8]. This multi-faceted validation bridges analytical rigor with real-world relevance, ensuring results are both chemically accurate and environmentally meaningful.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Research Materials for ML-NTA Experiments

Category	Item	Function and Application
Sample Preparation	Solid Phase Extraction (SPE)	Enrichment of specific compound classes; multi-sorbent strategies broaden coverage [8]
	QuEChERS	Efficient extraction for large-scale environmental samples; reduces solvent usage [8]
	Pressurized Liquid Extraction (PLE)	Automated extraction with high pressure and temperature for improved efficiency [8]
Instrumentation	HRMS Platforms (Q-TOF/Orbitrap)	High-resolution mass detection for accurate mass measurement and structural elucidation [8] [15]
	Liquid/Gas Chromatography Systems	Separation of complex mixtures prior to mass spectrometry analysis [8] [15]
Data Processing	Compound Discoverer, MZmine	Software for peak detection, alignment, and compound identification [15]
	Python/R with scikit-learn	Programming environments for implementing machine learning algorithms [70]
Reference Materials	Certified Reference Materials (CRMs)	Verification of compound identities and method validation [8]
	Spectral Libraries (NIST, GNPS)	Reference databases for compound identification via spectral matching [71]
Computational Resources	SHAP, Permutation Importance	Tools for model interpretability and feature importance analysis [70]

Implementation Considerations and Best Practices

Data Characteristics and Model Selection

The performance of ML classifiers is significantly influenced by dataset characteristics, particularly the ratio between sample size and feature dimensionality. Omics data are typically heterogeneous, sparse, and affected by the "curse of dimensionality" problem, having far fewer observations (samples) than features [69]. Research indicates that applying supervised feature selection improves the performance of feature extraction methods for classification purposes across various datasets [69]. For high-dimensional data with limited samples, random forests have demonstrated particular resilience, often outperforming other classifiers without requiring extensive feature pre-selection [68]. In contrast, most other machine learning classifiers including SVC show significant improvement when paired with feature selection filters like ReliefF, though even with such enhancements they typically do not outperform random forests in external validation scenarios [68].

Optimization Strategies for Enhanced Performance

To maximize classifier performance in NTA applications, several optimization strategies prove valuable. For random forests, semi-automatic parameter adjustment methods can identify optimal parameters, with studies demonstrating that RF algorithms with proper tuning achieve high accuracy and excellent resistance to overfitting [67]. For SVC, careful selection of kernel functions and regularization parameters is crucial, along with robust feature selection to handle high-dimensional chemical space. PLS-DA performance can be enhanced through appropriate data scaling and consideration of the optimal number of latent variables to avoid overfitting. For all classifiers, studies emphasize the importance of external validation using samples with known source-receptor relationships, as this provides a more realistic assessment of real-world performance compared to internal validation methods alone [8] [68].

Model Selection Decision Guide

This comparative analysis demonstrates that Random Forest, SVC, and PLS-DA each offer distinct advantages and limitations for source classification in non-target analysis research. Random Forest emerges as the most robust classifier for external validation scenarios, demonstrating superior performance with high-dimensional data and providing native feature importance metrics valuable for interpreting contamination sources or structure-activity relationships. SVC offers strong performance potential, particularly for complex non-linear relationships, but requires careful feature selection and parameter tuning while suffering from limited native interpretability. PLS-DA, despite its popularity in chemometrics, shows limitations in external validation contexts but remains valuable for more straightforward classification tasks with moderate-dimensional data. The selection of an appropriate classifier should be guided by specific research objectives, dataset characteristics, and validation requirements, with random forests representing a particularly compelling choice for real-world applications where generalizability to new sample populations is essential. As ML-assisted NTA continues to evolve, emphasis on model interpretability, robust validation strategies, and integration with domain knowledge will be crucial for advancing chemical confidence levels in non-target analysis assignment research.

Integrating NTA Findings with Toxicological Risk Assessment and Regulatory Frameworks

Non-targeted analysis (NTA) has emerged as a transformative approach for identifying unknown and unanticipated chemicals in environmental and biological samples, thereby addressing critical gaps in traditional risk assessment paradigms. Unlike conventional targeted methods that analyze predefined compounds, NTA employs high-resolution mass spectrometry (HRMS) to detect thousands of chemicals without prior knowledge, providing a comprehensive view of the chemical landscape [15]. This capability is particularly valuable for understanding complex exposure scenarios involving emerging contaminants, transformation products, and chemical mixtures that traditional monitoring often misses [8]. The integration of NTA findings into risk assessment and regulatory frameworks represents a paradigm shift from targeted chemical analysis to comprehensive exposure characterization, enabling more proactive and protective public health decision-making.

The fundamental challenge in contemporary chemical risk management lies in the vast and expanding chemical universe. With over 350,000 chemicals and substances in global use and more than 204 million chemicals in the largest registries, traditional targeted monitoring approaches capable of detecting only a small fraction of these compounds are increasingly inadequate for comprehensive risk assessment [14]. NTA bridges this gap by allowing retrospective screening and early identification of emerging contaminants without upfront selection and purchase of standards, thus providing a mechanism for continuous environmental monitoring and intervention [14]. This article examines current methodologies, computational frameworks, and validation strategies for integrating NTA-derived data into toxicological risk assessment and regulatory decision-making processes.

Fundamental Concepts: Targeted, Suspect, and Non-Targeted Analysis

Understanding the distinction between different analytical approaches is essential for contextualizing NTA's role in risk assessment. These approaches exist on a spectrum of chemical investigation, each with distinct applications, strengths, and limitations in regulatory contexts [72].

Targeted analysis represents the conventional approach in regulatory monitoring, focusing on precise quantification of predefined chemicals using reference standards. This method provides high-quality quantitative data for specific compounds but offers no information about other chemicals present in the sample [72]. Suspect screening analysis (SSA) occupies a middle ground, where chemicals are identified by comparison against predefined lists or libraries of suspected compounds. While broader than targeted analysis, SSA remains constrained by the scope of the suspect list employed [72]. Non-targeted analysis (NTA) represents the most comprehensive approach, aiming to characterize sample composition without prior knowledge of chemical content. True NTA attempts to identify unknown compounds not included in established libraries and not previously suspected in the samples [72].

In practice, many workflows integrate these approaches, using comprehensive data acquisition followed by tiered data analysis that sequentially applies targeted, suspect, and non-targeted identification strategies [72]. This integrated approach maximizes both quantitative precision and comprehensiveness in chemical exposure assessment.

Table 1: Comparison of Analytical Approaches in Chemical Monitoring

Aspect	Targeted Analysis	Suspect Screening	Non-Targeted Analysis
Scope	Limited to predefined compounds	Limited by suspect list	Comprehensive, no upfront limitations
Quantification	Precise with standards	Semi-quantitative	Qualitative to semi-quantitative
Identification Confidence	High (with standards)	Moderate to high	Variable (Levels 1-5)
Primary Application	Regulatory compliance	Chemical prioritization	Exposure discovery
Data Volume	Low	Moderate	High
Standards Required	Before analysis	For confirmation	For highest confidence ID

Analytical Workflows and Methodologies

Standardized NTA Workflows

The integration of NTA into risk assessment begins with robust analytical workflows that ensure data quality and interpretability. A systematic four-stage framework for NTA encompasses sample treatment and extraction, data generation and acquisition, ML-oriented data processing and analysis, and result validation [8]. Sample preparation requires careful optimization to balance selectivity and sensitivity, often employing techniques such as solid phase extraction (SPE), Soxhlet extraction, gel permeation chromatography (GPC), and pressurized liquid extraction (PLE) to maximize compound recovery while minimizing matrix interference [8]. For liquid samples with sufficient concentrations, direct injection is often recommended, while solid samples typically require extraction with organic solvents such as methanol or acetonitrile for LC and hexane or acetone for GC analysis [14].

Data generation relies on HRMS platforms, including quadrupole time-of-flight (Q-TOF) and Orbitrap systems, coupled with liquid or gas chromatographic separation (LC/GC) to resolve isotopic patterns, fragmentation signatures, and structural features necessary for compound annotation [8]. Post-acquisition processing involves centroiding, extracted ion chromatogram (EIC/XIC) analysis, peak detection, alignment, and componentization to group related spectral features into molecular entities [8]. Quality assurance measures, including confidence-level assignments and batch-specific quality control samples, are critical throughout this process to ensure data integrity for subsequent risk assessment applications [8].

Chromatographic and Ionization Considerations

The detectable chemical space in NTA is heavily influenced by analytical platform selection. Liquid chromatography (LC) coupled with electrospray ionization (ESI) is particularly effective for polar, water-soluble compounds and larger molecules, while gas chromatography (GC) with electron ionization (EI) covers more non-polar, volatile compounds [15] [14]. Studies indicate that approximately 51% of NTA investigations use only LC-HRMS, 32% use only GC-HRMS, and 16% use both platforms to expand chemical coverage [15]. The selection of ionization techniques further influences detectable chemical space, with many LC-HRMS studies employing both negative and positive electrospray ionization (43% of studies) to broaden compound detection [15].

The chemical domain covered by any NTA method represents the intersection of all method steps, from sample preparation through instrumental analysis [14]. Understanding these methodological boundaries is essential for proper interpretation of NTA results in risk assessment contexts, as certain compound classes with specific properties may require specialized approaches. For example, highly hydrophilic ionic compounds like glyphosate or very non-polar high-molecular weight compounds such as large polycyclic aromatic hydrocarbons may not be effectively captured by generic screening methods [14].

Diagram 1: Integrated NTA and Risk Assessment Workflow. This workflow illustrates the sequential stages from sample collection to regulatory decision, highlighting the transition from analytical phases (yellow) to data interpretation (green) and risk assessment integration (red).

Computational Tools and Data Analysis Frameworks

Machine Learning Applications in NTA

The complexity and volume of data generated by HRMS-based NTA necessitates advanced computational approaches for meaningful interpretation. Machine learning (ML) algorithms have demonstrated particular utility for identifying latent patterns in high-dimensional NTA data, enabling more accurate contamination source identification and chemical prioritization [8]. ML classifiers such as Support Vector Classifier (SVC), Logistic Regression (LR), and Random Forest (RF) have been successfully implemented to screen hundreds of per- and polyfluoroalkyl substances (PFAS) across different sources, achieving classification balanced accuracy ranging from 85.5% to 99.5% [8]. These approaches represent a significant advancement over traditional statistical methods that often struggle to disentangle complex source signatures.

The ML-oriented data processing pipeline typically involves sequential computational steps beginning with data preprocessing to address quality issues through noise filtering, missing value imputation, and normalization to mitigate batch effects [8]. Exploratory analysis then identifies significant features via univariate statistics and prioritizes compounds with large fold changes. Dimensionality reduction techniques like principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) simplify high-dimensional data, while clustering methods group samples by chemical similarity [8]. Supervised ML models are subsequently trained on labeled datasets to classify contamination sources, with feature selection algorithms refining input variables to optimize model accuracy and interpretability [8].

In Silico Identification Tools

Compound identification represents a significant bottleneck in NTA workflows, with in silico tools playing an increasingly important role in addressing this challenge. MetFrag, an open in silico identification approach, exemplifies this trend by retrieving potential candidates with matching masses from compound databases and scoring them according to how well experimental spectra match in silico fragments [73]. The integration of regulatory chemical databases has significantly enhanced the utility of such tools for risk assessment applications. For example, connecting MetFrag with the US EPA's CompTox Chemicals Dashboard provides access to over 850,000 compounds of environmental and toxicological relevance while allowing users to leverage the "MS-Ready" concept and various forms of chemical metadata [73].

Critical information from international regulatory bodies can now be exploited through computational platforms toward identifying environmental chemicals. These resources include the US EPA's Chemicals and Products database (CPDat), hazard and exposure information from the Swedish Chemicals Agency KEMI, European chemicals registration data (REACH), and the NORMAN Network's merged suspect list of chemicals of emerging concern [73]. This integration of disparate regulatory resources creates an interconnected information platform that supports more chemically relevant identification of environmental unknowns, effectively helping researchers and regulators collaborate through shared computational infrastructure.

Table 2: Key Computational Tools for NTA Data Analysis

Tool Name	Primary Function	Data Sources	Regulatory Relevance
MetFrag	In silico fragmentation and compound identification	PubChem, ChemSpider, CompTox, NORMAN SusDat	Integrates multiple regulatory agency data sources
US EPA CompTox Dashboard	Chemical data aggregation and curation	~850,000 substances with environmental relevance	EPA regulatory priorities and toxicity data
NORMAN SusDat	Suspect screening list	Chemicals of emerging concern from EU monitoring	European regulatory focus chemicals
XCMS	LC/MS data preprocessing	Raw mass spectrometry data	Open-source tool for cross-platform data analysis
Shinyscreen	Automated quality control of mass spectra	HRMS raw data	Streamlines data quality assessment for regulatory applications

Confidence Assessment and Identification Frameworks

Tiered Identification Confidence

The translation of NTA findings into regulatory actions requires clear communication of identification confidence. The scientific community has established confidence levels for NTA identification, ranging from Level 1 (confirmed structure) to Level 5 (unequivocal molecular formula) [74]. This tiered framework provides transparency regarding the evidence supporting each identification, enabling risk assessors to appropriately weight NTA findings in decision-making processes [74]. Level 1 confirmation requires matching retention time and spectral data with authentic standards, providing the highest confidence suitable for regulatory action [74]. In contrast, Level 5 identifications based solely on molecular formula may be suitable for hypothesis generation but require further confirmation for risk assessment applications.

The BP4NTA Working Group has developed standardized terminology and reporting frameworks to improve consistency in confidence assignment across laboratories and studies [74]. These efforts address the historical variability in identification criteria that has hampered the regulatory adoption of NTA data. The NTA Study Reporting Tool (SRT) provides a structured framework for transparent reporting of methodological details and confidence assignments, facilitating critical evaluation of data quality and appropriate interpretation in risk assessment contexts [75]. Implementation of these harmonized approaches is essential for building regulatory trust in NTA-derived data.

Validation Strategies

Robust validation strategies are essential for establishing the reliability of NTA outputs intended for risk assessment applications. A comprehensive, tiered validation approach integrates analytical verification, model generalizability assessment, and environmental plausibility evaluation [8]. Analytical confidence is first verified using certified reference materials or spectral library matches to confirm compound identities [8]. Model generalizability is then assessed by validating classifiers on independent external datasets, complemented by cross-validation techniques to evaluate overfitting risks [8]. Finally, environmental plausibility checks correlate model predictions with contextual data, such as geospatial proximity to emission sources or known source-specific chemical markers [8].

This multi-faceted validation bridges analytical rigor with real-world relevance, ensuring NTA results are both chemically accurate and environmentally meaningful for risk assessment applications [8]. The emphasis on environmental plausibility is particularly important for regulatory acceptance, as it demonstrates that NTA findings align with known contamination patterns and exposure scenarios. Validation should also address quantitative aspects when NTA data are used for exposure assessment, employing techniques such as quantitative structure–retention relationship models and ionization efficiency-based quantification to derive concentration estimates without authentic standards [12].

Integration with Risk Assessment Frameworks

Next-Generation Risk Assessment (NGRA)

The incorporation of NTA findings into risk assessment is most advanced within next-generation risk assessment (NGRA) frameworks that integrate exposure science, toxicokinetics, and toxicodynamics using new approach methodologies (NAMs) [76]. NGRA represents a shift from traditional risk assessment approaches by leveraging in vitro bioactivity data, computational toxicology, and targeted testing to evaluate chemical safety [76]. A tiered NGRA framework applied to pyrethroids demonstrates how bioactivity indicators derived from high-throughput screening can be combined with exposure estimates to evaluate cumulative risks, addressing limitations of traditional assessment methods that rely heavily on acceptable daily intakes and default extrapolation models [76].

The five-tiered NGRA approach begins with bioactivity data gathering and progress through combined risk assessment, margin of exposure analysis with toxicokinetic modeling, refinement of bioactivity indicators, and confirmation of risk characterization [76]. This structured framework provides a scientifically robust yet resource-efficient strategy for evaluating data-rich and data-poor chemicals, with NTA serving as a critical tool for identifying previously unrecognized exposures requiring assessment. The integration of NTA with NGRA is particularly valuable for evaluating real-world exposure to complex chemical mixtures, as it provides comprehensive exposure data that can be combined with bioactivity information from ToxCast and similar programs [76].

Evidence-Based Risk Assessment Framework

A systematic framework for evidence-based risk assessment provides a structured approach for integrating diverse data streams, including NTA findings, into chemical safety decisions [77]. This framework incorporates principles from evidence-based medicine and toxicology to ensure risk decisions are based on the best available scientific evidence, identified and evaluated through transparent, objective processes [77]. The approach emphasizes systematic review methodologies to comprehensively assemble and evaluate relevant evidence, with explicit consideration of the strengths and limitations of different data sources [77].

The evidence-based risk assessment framework encompasses four key phases: (1) defining the causal question and developing criteria for study selection; (2) developing and applying criteria for review of individual studies; (3) evaluating and integrating evidence; and (4) drawing conclusions based on inferences [77]. This structured approach is applicable to both data-rich and data-poor risk decision contexts, making it particularly valuable for evaluating emerging contaminants identified through NTA where traditional toxicity data may be limited. The framework facilitates appropriate weighting of NTA-derived evidence relative to other data streams, such as epidemiological studies, in vivo toxicology, and mechanistic data [77].

Diagram 2: NTA Data Integration in Risk Assessment Framework. This diagram illustrates how NTA-derived exposure data integrates with hazard information in next-generation risk assessment paradigms, ultimately supporting risk characterization and management decisions.

Regulatory Applications and Implementation Challenges

Environmental Monitoring Applications

Regulatory applications of NTA in environmental monitoring are advancing, with several jurisdictions developing formal guidance for implementation. The NORMAN Network has established comprehensive guidance for suspect and non-target screening in environmental monitoring, covering all steps from sampling and sample preparation through analysis and data evaluation to reporting [14]. This guidance acknowledges that while NTS methods strive to cover the largest possible compound domain, it is essential to understand methodological limitations, particularly regarding what chemicals may not be covered [14]. Such transparency is critical for appropriate regulatory interpretation and application of NTA findings.

Retrospective NTA applications demonstrate how existing monitoring data can be repurposed to address regulatory priorities, such as identifying pollutants with industrial point sources occurring at high intensities across multiple time points [73]. These approaches leverage in silico workflows to prioritize "masses of interest" and identify potential "known unknown" pollutants by incorporating regulator-supplied chemical information [73]. The successful application of such workflows in regulatory contexts highlights the potential for NTA to enhance monitoring programs without requiring complete methodological overhaul, instead building upon existing targeted approaches through data mining and retrospective analysis.

Methodological and Interpretation Challenges

Despite significant advances, challenges remain in the widespread regulatory implementation of NTA for risk assessment. Method validation approaches remain fragmented and overly reliant on laboratory-based tests, potentially underperforming in real-world conditions involving field-validated source-receptor relationships [8]. Model interpretability also presents challenges, as complex models like deep neural networks may achieve high classification accuracy but offer limited transparency regarding attribution rationale, reducing regulatory trust [8]. Additionally, insufficient emphasis on environmental plausibility assessment may limit the real-world relevance of NTA findings for risk decision-making [8].

The translation of NTA findings into risk-based prioritization requires careful consideration of quantitative aspects, particularly when authentic standards are unavailable for concentration determination. Quantitative structure–property relationship models and ionization efficiency-based quantification offer promising approaches for estimating concentrations without authentic standards, but these methods introduce additional uncertainty that must be accounted for in risk characterization [12]. Furthermore, integration of NTA data with bioactivity information requires careful consideration of concentration relevance, as detected compounds may be present at levels below biological activity thresholds [12]. Addressing these challenges requires continued method development and interdisciplinary collaboration between analytical chemists, toxicologists, and risk assessors.

Table 3: Key Reagents and Materials for NTA Workflows

Category	Specific Examples	Application in NTA	Regulatory Relevance
Extraction Materials	Oasis HLB, ISOLUTE ENV+, Strata WAX/WCX, QuEChERS	Broad-spectrum extraction, matrix cleanup	Standardized protocols improve interlaboratory comparability
Chromatography Columns	C18 (LC), phenylmethylpolysiloxane (GC), HILIC	Compound separation by physicochemical properties	Method reproducibility across monitoring networks
Ionization Sources	ESI, APCI, EI	Compound-dependent ionization efficiency	Complementary coverage of chemical space
Reference Materials	Certified reference materials, stable isotope-labeled standards	Identification confirmation, quantification	Essential for highest confidence identifications (Level 1)
Quality Control Materials	Procedure blanks, solvent blanks, pooled samples	Monitoring contamination, signal drift	Required for data quality assessment in regulatory contexts

The integration of NTA findings with toxicological risk assessment and regulatory frameworks represents a paradigm shift in chemical safety evaluation, moving from targeted investigation of predefined chemicals to comprehensive characterization of real-world exposures. Machine learning and artificial intelligence are poised to further transform this field, enhancing pattern recognition, source attribution, and toxicity prediction capabilities [12]. These computational advances will improve the efficiency and accuracy of contaminant source identification, ultimately contributing to more effective environmental protection measures [8]. However, realizing this potential requires addressing current limitations in model interpretability and validation [8].

Future developments will likely focus on refining integrated workflows that combine NTA with effect-directed analysis to prioritize biologically active contaminants, thereby bridging the gap between exposure identification and hazard characterization [12]. Additionally, harmonized reporting standards and quality control approaches, such as those developed by the BP4NTA Working Group and NORMAN Network, will be essential for building regulatory confidence in NTA-derived data [74] [75] [14]. As these frameworks mature, NTA is positioned to transition from a research tool to a routine component of regulatory monitoring programs, providing comprehensive exposure data to support next-generation, evidence-based risk assessment [77]. This evolution will enable more proactive chemical management that keeps pace with the expanding chemical universe and protects public health through identification of emerging contaminants before they become widespread problems.

Conclusion

Establishing rigorous chemical confidence levels is paramount for transforming non-target analysis from an exploratory tool into a reliable source for regulatory and clinical decision-making. The integration of machine learning and structured workflows is key to managing the complexity of HRMS data, while multi-tiered validation ensures findings are both chemically sound and environmentally or clinically relevant. Future progress hinges on expanding spectral libraries, harmonizing analytical protocols, and further integrating computational toxicology. For drug development, these advances will be crucial for characterizing complex biologics, novel modalities, and ensuring the safety of products, ultimately accelerating the delivery of safe and effective therapies.

References

• 1. 重磅发布｜非靶代谢组-队列精准版：颠覆海量数据 [https://testhome.aptbiotech.com/market/7046]
• 2. 基于色谱-质谱联用的暴露组学分析方法和研究范式的新进展- PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC10877474/]
• 3. timsTOF Pro 2 [https://www.bruker.com/zh/products-and-solutions/mass-spectrometry/timstof/timstof-pro-2.html]
• 4. 规模以上非靶向代谢组学剖析血清超高效液相色谱 [https://www.jove.com/cn/t/50242/large-scale-non-targeted-metabolomic-profiling-serum-ultra]
• 5. 代谢组学分析全流程革新——科研人员的LC-MS数据处理与 ... [https://blog.csdn.net/gitblog_00742/article/details/154560959]
• 6. 顶刊聚焦：Orbitrap Astral凭什么成为代谢组学“新宠”？两大核心 ... [https://www.biotree.com.cn/xingyedongtai/1051.html]
• 7. Non-Targeted Analysis Research [https://www.epa.gov/chemical-research/non-targeted-analysis-research]
• 8. Integrating non-target analysis and machine learning [https://www.nature.com/articles/s41545-025-00504-z]
• 9. News – BP4NTA [https://nontargetedanalysis.org/news/]
• 10. Critical review on in silico methods for structural annotation of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11700063/]
• 11. Machine Learning for Enhanced Identification Probability in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12392258/]
• 12. Advancing non-target analysis of emerging environmental ... [https://www.sciencedirect.com/science/article/pii/S0160412025001552]
• 13. Demonstrating the Use of Non-targeted Analysis for ... - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC10198433/]
• 14. NORMAN guidance on suspect and non-target screening in ... [https://enveurope.springeropen.com/articles/10.1186/s12302-023-00779-4]
• 15. Non-targeted analysis (NTA) and suspect screening ... [https://www.nature.com/articles/s41370-023-00574-6]
• 16. Do experimental projection methods outcompete retention ... [https://pubs.rsc.org/en/content/articlehtml/2025/an/d5an00323g]
• 17. Software Tools and Approaches for Compound Identification of LC-MS/MS Data in Metabolomics - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC6027441/]
• 18. Enhancing the Detectable Chemical Space in an Effluent ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12351525/]
• 19. Interpretation of Isotope Peaks in Small Molecule LC?MS | LCGC International [https://www.chromatographyonline.com/view/interpretation-isotope-peaks-small-molecule-lcms]
• 20. Hydrogen Rearrangement Rules: Computational MS/MS Fragmentation and Structure Elucidation Using MS-FINDER Software - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC7063832/]
• 21. LC-MS/MS data analysis with xcms [https://www.bioconductor.org/packages/release/bioc/vignettes/xcms/inst/doc/xcms-lcms-ms.html]
• 22. In Silico Optimization of Mass Spectrometry Fragmentation Strategies in Metabolomics - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC6836109/]
• 23. A tutorial in small molecule identification via electrospray ionization‐mass spectrometry: The practical art of structural elucidation - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC6099382/]
• 24. Recent Trends and Challenges on the Non-Targeted ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12299502/]
• 25. Comprehensive benchmarking of computational tools for ... [https://jcheminf.biomedcentral.com/articles/10.1186/s13321-024-00931-z]
• 26. Computational mass spectrometry for exposomics in non‑ ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12583337/]
• 27. Pattern Recognition: Latest Techniques and Applications [https://labelyourdata.com/articles/machine-learning/pattern-recognition]
• 28. 5 Breakthrough Machine Learning Research Papers ... [https://machinelearningmastery.com/5-breakthrough-machine-learning-research-papers-already-in-2025/]
• 29. The future of AI regulation in drug development: a comparative ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12598624/]
• 30. Artificial Intelligence for Drug Development [https://www.fda.gov/about-fda/center-drug-evaluation-and-research-cder/artificial-intelligence-drug-development]
• 31. AI and ML in Drug Development Market Size and Key Players [https://www.towardshealthcare.com/insights/ai-and-ml-in-drug-development-market-sizing]
• 32. HPLC 2025 Preview: Prioritization Strategies in Non-Target ... [https://www.chromatographyonline.com/view/hplc-2025-preview-prioritization-strategies-in-non-target-screening-of-environmental-samples-by-chromatography-with-high-resolution-mass-spectrometry]
• 33. Identifying non-intentionally added substances in food ... [https://www.kemira.com/stories/identifying-non-intentionally-added-substances-food-contact-materials/]
• 34. Plastic food contact materials: declaration of compliance ... [https://www.mayerbrown.com/en/insights/publications/2025/11/plastic-food-contact-materials-declaration-of-compliance-soon-to-reflect-compliance-with-new-purity-requirements]
• 35. EU Revises Food Contact Material Regulations [https://www.sgs.com/en/news/2025/03/safeguards-03625-eu-revises-food-contact-material-regulations]
• 36. Emerging New Drug Modalities in 2025 [https://www.bcg.com/publications/2025/emerging-new-drug-modalities]
• 37. Unknown confidence in chemical characterization ... [https://www.sciencedirect.com/science/article/pii/S0273230025000923]
• 38. The critical role that spectral libraries play in capturing ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10284100/]
• 39. Recent Trends and Challenges on the Non-Targeted ... [https://www.mdpi.com/2305-6304/13/7/543]
• 40. USGS Spectral Library Version 7 Data [https://data.usgs.gov/datacatalog/data/USGS:5807a2a2e4b0841e59e3a18d]
• 41. Generative Modeling for Accurate Molecule Retrieval from ... [https://arxiv.org/html/2511.06259v1]
• 42. Carafe enables high quality in silico spectral library ... [https://www.nature.com/articles/s41467-025-64928-4]
• 43. A comparative study on noise filtering of imbalanced data ... [https://www.sciencedirect.com/science/article/pii/S0950705124008700]
• 44. Benchmarking machine learning methods for the ... [https://link.springer.com/article/10.1007/s10462-025-11293-9]
• 45. A Comparative Analysis of Preprocessing Filters for Deep ... [https://www.mdpi.com/2076-3417/15/20/11277]
• 46. An HPLC 2025 Video Interview with Bob Pirok [https://www.chromatographyonline.com/view/method-robustness-retention-alignment-and-machine-learning-in-2d-lc-an-hplc-2025-video-interview-with-bob-pirok]
• 47. Data Fusion in Action: Integrating Different Vibrational and ... [https://www.spectroscopyonline.com/view/data-fusion-in-action-integrating-different-vibrational-and-atomic-spectroscopy-data]
• 48. Comparison of missing value imputation tools for machine ... [https://www.sciencedirect.com/science/article/pii/S0023643825002695]
• 49. Bridging the Gap: Missing Data Imputation Methods and ... [https://www.mdpi.com/2076-3425/15/6/639]
• 50. Imputation for Lipidomics and Metabolomics (ImpLiMet) [https://pubmed.ncbi.nlm.nih.gov/39867531/]
• 51. Resource Allocation vs. Project Prioritization [https://growthshuttle.com/resource-allocation-vs-project-prioritization/]
• 52. Demystifying the Accuracy-Interpretability Trade-Off [https://arxiv.org/html/2503.07914v1]
• 53. An Evaluation of Interpretable Machine Learning Models [https://link.springer.com/article/10.1007/s12599-024-00922-2]
• 54. vs. Balancing in Data Science Model Complexity Interpretability [https://www.linkedin.com/advice/1/how-do-you-balance-complexity-interpretability-your-psjvc]
• 55. Simpler is better: Lifting interpretability-performance trade ... [https://www.sciencedirect.com/science/article/pii/S016792362100066X]
• 56. Artificial intelligence in drug discovery and development [https://link.springer.com/article/10.1007/s44395-025-00007-3]
• 57. Accuracy and Balancing Trade-offs in Interpretability ... Model [https://moldstud.com/articles/p-balancing-accuracy-and-interpretability-trade-offs-in-model-complexity-for-machine-learning]
• 58. Performance and Clarity in Balancing Machine Learning [https://scisimple.com/en/articles/2025-11-14-balancing-performance-and-clarity-in-machine-learning--ak6pj5g]
• 59. Applications of machine learning in drug discovery and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6552674/]
• 60. Explainable Artificial Intelligence in the Field of Drug Research [https://pmc.ncbi.nlm.nih.gov/articles/PMC12129466/]
• 61. Biological plausibility in environmental health systematic ... [https://assessments.epa.gov/risk/document/&deid%3D364639]
• 62. A Record-Breaking Dataset to Train AI Models has Launched [https://newscenter.lbl.gov/2025/05/14/computational-chemistry-unlocked-a-record-breaking-dataset-to-train-ai-models-has-launched/]
• 63. Exploring Meta's Open Molecules 2025 (OMol25) & Universal ... [https://www.rowansci.com/blog/exploring-open-molecules-2025]
• 64. Evaluation of source–receptor relationship for atmospheric ... [https://www.sciencedirect.com/science/article/pii/S1309104215304669]
• 65. TM5-FASST: a global atmospheric source–receptor model ... [https://acp.copernicus.org/articles/18/16173/2018/]
• 66. Publications [https://nontargetedanalysis.org/category/news/publications/]
• 67. Application of random forest based on semi-automatic ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9353770/]
• 68. Merits of random forests emerge in evaluation ... [https://www.sciencedirect.com/science/article/abs/pii/S0003267013012142]
• 69. Research Article Benchmarking feature selection and ... [https://www.sciencedirect.com/science/article/pii/S2001037024000692]
• 70. Feature Importance with Random Forests [https://www.geeksforgeeks.org/machine-learning/feature-importance-with-random-forests/]
• 71. Recent Advances in Implementation of Machine Learning for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12645304/]
• 72. Study Design [https://nontargetedanalysis.org/reference-content/methods/study-design/]
• 73. Retrospective non-target analysis to support regulatory water ... [https://enveurope.springeropen.com/articles/10.1186/s12302-021-00475-1]
• 74. An Introduction to the Benchmarking and Publications for Non ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8848292/]
• 75. Nontargeted Analysis Study Reporting Tool: A Framework ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9408805/]
• 76. A tiered next-generation risk assessment framework ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12198285/]
• 77. Development of an Evidence-Based Risk Assessment ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10080579/]