A Tiered Validation Strategy for ML-Assisted Non-Target Analysis: From Foundational Concepts to Advanced Applications in Biomedical Research

Abstract

This article provides a comprehensive framework for implementing a robust tiered validation strategy in Machine Learning-assisted Non-Target Analysis (ML-NTA). Tailored for researchers, scientists, and drug development professionals, it bridges the gap between raw analytical data and environmentally or biologically actionable insights. The content systematically progresses from foundational principles of NTA and ML integration to advanced methodological applications, tackling common troubleshooting scenarios. It culminates in a detailed examination of multi-tiered validation, incorporating analytical verification, external dataset testing, and environmental plausibility assessments. By offering a structured pathway to ensure the reliability, interpretability, and real-world relevance of ML-NTA outputs, this guide aims to empower professionals in translating complex datasets into credible findings for drug discovery, environmental monitoring, and risk assessment.

Demystifying ML-Assisted NTA: Core Principles and the Critical Need for Tiered Validation

Core Principles and Definitions

Non-Target Analysis (NTA) represents a paradigm shift in analytical chemistry, moving from hypothesis-driven to discovery-based approaches. Unlike traditional targeted analysis that quantifies predefined compounds, NTA aims to comprehensively detect and identify a wide range of chemical substances without prior knowledge of the sample composition [1]. This capability is particularly valuable for discovering unknown contaminants, transformation products, and metabolites that would otherwise escape detection using conventional methods.

High-Resolution Mass Spectrometry (HRMS) serves as the analytical foundation for NTA by providing the exact molecular mass of compounds with exceptional accuracy. Where conventional mass spectrometry measures nominal mass, HRMS distinguishes between molecules with minute mass differences—such as cysteine (121.0196 Da) and benzamide (121.0526 Da)—enabling precise molecular formula assignment and compound identification [2]. The high resolving power (typically ≥20,000) and mass accuracy (≤5 ppm) of modern HRMS instruments make this distinction possible [3].

The integration of these fields has created a powerful platform for comprehensive chemical characterization across pharmaceutical, environmental, and biological research, particularly for addressing "known unknowns" and "unknown unknowns" in complex mixtures [4] [5].

The HRMS Working Principle

The operational principle of HRMS encompasses three fundamental stages that transform sample molecules into interpretable data, as detailed in Table 1.

Table 1: Fundamental Stages of High-Resolution Mass Spectrometry

Step	Description	Common Techniques	Key Applications
Ionization	Converts neutral molecules to gas-phase ions	Electrospray Ionization (ESI), Matrix-Assisted Laser Desorption/Ionization (MALDI)	ESI for fragile biomolecules; MALDI for proteins and polymers
Mass Analysis	Separates ions by mass-to-charge ratio (m/z)	Time-of-Flight (TOF), Orbitrap, Fourier Transform Ion Cyclotron Resonance (FT-ICR)	TOF for rapid screening; Orbitrap for high resolution; FT-ICR for ultra-high resolution
Detection	Records ion intensity and exact mass	High-precision detectors	Quantification, structural elucidation, formula prediction

The ionization process occurs under vacuum conditions to prevent ion-molecule collisions, using techniques like ESI that preserve molecular integrity for accurate mass determination [2] [6]. Following ionization, mass analyzers separate ions based on their m/z values with high resolution, while detection systems generate mass spectra that reflect ion abundance and precise molecular weights [6].

Integrated NTA Workflow with HRMS

The complete NTA workflow integrates sample preparation, HRMS analysis, and advanced data processing in a systematic approach to uncover previously undetected chemicals. The following diagram illustrates this comprehensive process:

NTA-HRMS Integrated Workflow with Machine Learning Assistance

Sample Preparation and Extraction

Effective sample preparation is crucial for balancing selectivity and sensitivity in NTA. The goal is to remove interfering matrix components while preserving a broad spectrum of analytes [7]. Common extraction techniques include:

• Solid Phase Extraction (SPE): Often employs multi-sorbent strategies (e.g., Oasis HLB with ISOLUTE ENV+, Strata WAX, and WCX) to broaden chemical coverage [7]
• Green Extraction Techniques: QuEChERS, Microwave-Assisted Extraction (MAE), and Supercritical Fluid Extraction (SFE) improve efficiency by reducing solvent usage and processing time [7]

HRMS Data Acquisition and Processing

HRMS platforms, including quadrupole time-of-flight (Q-TOF) and Orbitrap systems coupled with liquid or gas chromatography (LC/GC), generate complex datasets essential for NTA [7]. Post-acquisition processing involves:

• Centroiding: Converting profile data to centroid data
• Peak Detection and Alignment: Identifying chromatographic peaks and aligning retention times across samples
• Componentization: Grouping related spectral features (adducts, isotopes) into molecular entities [7]

Quality assurance measures include confidence-level assignments (Level 1-5) and batch-specific quality control samples to ensure data integrity [7]. The output is a structured feature-intensity matrix serving as the foundation for machine learning analysis.

Experimental Protocols for Key NTA Applications

Protocol 1: ML-Assisted Contaminant Source Tracking

Objective: Identify contamination sources using ML-based NTA with a tiered validation strategy [7].

Table 2: Reagent Solutions for Contaminant Source Tracking

Research Reagent	Function	Application Context
Mixed-mode SPE cartridges	Broad-spectrum analyte enrichment	Water sample preparation for NTA
LC-HRMS quality control samples	Monitor instrument performance	Batch-to-batch normalization
Certified Reference Materials (CRMs)	Verify compound identities	Analytical confidence assessment
Internal standard mixture	Correct retention time drift	Data alignment across batches

Procedure:

• Sample Collection: Collect environmental samples (water, soil, biota) from potential contamination sources
• Extraction: Process using mixed-mode SPE to maximize compound coverage [7]
• HRMS Analysis: Analyze using LC-HRMS with ESI in both positive and negative modes
• Data Preprocessing:
  • Apply retention time correction and m/z recalibration
  • Perform peak matching across batches [7]
  • Generate feature-intensity matrix
• ML Analysis:
  • Conduct exploratory analysis using PCA and hierarchical clustering
  • Train supervised classifiers (Random Forest, SVM) on labeled samples
  • Apply feature selection to identify source-specific chemical indicators [7]
• Validation:
  • Tier 1: Verify identities using CRMs or spectral library matches
  • Tier 2: Assess generalizability with external datasets
  • Tier 3: Evaluate environmental plausibility using geospatial data [7]

Protocol 2: Structural Alert Screening for Hazard Prioritization

Objective: Prioritize potentially hazardous features using ML classification of MS/MS spectra [8].

Procedure:

• Data Acquisition: Collect MS/MS spectra for known compounds with and without structural alerts
• Feature Extraction:
  • Extract fragment masses and neutral losses from MS/MS spectra
  • Bin fragments and neutral losses to nearest 0.1 m/z
  • Create binary matrices indicating presence/absence of features [8]
• Model Training:
  • Train neural network classifier for organophosphorus alerts
  • Train Random Forest classifier for aromatic amine alerts
  • Optimize using cross-validation [8]
• Application to NTS Data:
  • Apply trained models to prioritize LC-HRMS features in environmental samples
  • Focus identification efforts on high-priority features [8]

Machine Learning Integration and Tiered Validation

Machine learning has redefined NTA potential by identifying latent patterns in high-dimensional HRMS data that traditional statistics often miss [7]. The tiered validation framework ensures ML outputs are both chemically accurate and environmentally meaningful, addressing the critical gap between analytical capability and decision-making.

The following diagram illustrates the ML-assisted analysis framework with integrated validation:

ML-Assisted NTA Framework with Tiered Validation

ML-Oriented Data Processing and Analysis

The transition from raw HRMS data to interpretable patterns involves sequential computational steps:

• Data Preprocessing: Addresses data quality through noise filtering, missing value imputation (e.g., k-nearest neighbors), and normalization to mitigate batch effects [7]
• Exploratory Analysis: Identifies significant features via univariate statistics and prioritizes compounds with large fold changes
• Dimensionality Reduction: Techniques like PCA and t-SNE simplify high-dimensional data [7]
• Supervised ML Models: Including Random Forest and Support Vector Classifiers trained on labeled datasets to classify contamination sources [7]

Tiered Validation Strategy

A robust three-tiered validation framework ensures reliable ML-NTA outputs:

• Analytical Confidence: Verified using certified reference materials or spectral library matches to confirm compound identities [7]
• Model Generalizability: Assessed by validating classifiers on independent external datasets with cross-validation techniques to evaluate overfitting risks [7]
• Environmental Plausibility: Correlates model predictions with contextual data, such as geospatial proximity to emission sources or known source-specific chemical markers [7]

Applications and Current Challenges

NTA-HRMS has demonstrated significant utility across multiple domains, though challenges remain for full operationalization.

Table 3: NTA-HRMS Applications Across Sample Matrices

Sample Matrix	Commonly Detected Chemicals	Analytical Platform	Key Applications
Water	PFAS, pharmaceuticals, pesticides	LC-HRMS (51%), GC-HRMS (32%), Both (16%)	Source tracking, emerging contaminant discovery
Soil/Sediment	Pesticides, PAHs, transformation products	GC-HRMS, LC-HRMS	Effect-directed analysis, contamination forensics
Human Biospecimens	Plasticizers, pesticides, halogenated compounds	LC-HRMS (ESI+/ESI-)	Biomarker discovery, exposure assessment
Consumer Products	Flame retardants, plasticizers	GC-HRMS, LC-HRMS	Safety evaluation, regulatory compliance

Key Applications

• Environmental Monitoring: Identifying previously unknown pollutants and transformation products in water, soil, and air [5]
• Exposure Assessment: Characterizing human chemical exposures through biomonitoring [5]
• Pharmaceutical Analysis: Detecting unknown impurities and metabolites in drug development [1]
• Effect-Directed Analysis: Identifying bioactive compounds in complex mixtures through fractionation and bioactivity testing [4]

Current Challenges and Research Gaps

Despite significant advances, NTA-HRMS faces several challenges:

• Quantitative Limitations: Most NTA provides relative quantitation; absolute concentration estimates require additional calibration approaches [4]
• Data Complexity: Large, high-dimensional datasets require advanced computational tools and expertise [7]
• Standardization Needs: Lack of harmonized protocols and reporting standards across laboratories [3]
• Confidence Assessment: Varying levels of identification confidence require systematic reporting frameworks [3]
• Black-Box Models: Complex ML models like deep neural networks lack interpretability, limiting regulatory acceptance [7]

The integration of machine learning with NTA-HRMS continues to evolve, with future developments focusing on improved quantification methods, enhanced model interpretability, and standardized validation frameworks to bridge the gap between analytical capability and environmental decision-making [7] [4] [9].

The Role of Machine Learning in Interpreting Complex NTA Datasets

Non-target analysis (NTA) using high-resolution mass spectrometry (HRMS) has emerged as a vital approach for detecting thousands of chemicals without prior knowledge, proving particularly valuable for identifying emerging environmental contaminants and unknown compounds in complex samples [7] [9]. The principal challenge of NTA now lies not in detection itself, but in developing computational methods to extract meaningful environmental information from the vast chemical datasets generated by HRMS platforms [7]. Machine learning (ML) has redefined the potential of NTA by effectively identifying latent patterns within high-dimensional data, making these algorithms particularly well-suited for contamination source identification and compound characterization [7]. This document outlines a systematic framework and detailed protocols for implementing ML-assisted NTA within a tiered validation strategy for robust research outcomes.

Comprehensive Workflow of ML-Assisted NTA

The integration of ML and NTA for contaminant source identification follows a systematic four-stage workflow: (i) sample treatment and extraction, (ii) data generation and acquisition, (iii) ML-oriented data processing and analysis, and (iv) result validation [7]. Each stage requires careful optimization to ensure data quality and interpretable results.

Stage 1: Sample Treatment and Extraction

Sample preparation requires careful optimization to balance selectivity and sensitivity, achieving a compromise between removing interfering components and preserving as many compounds as possible with adequate sensitivity [7].

Protocol 1.1: Comprehensive Sample Preparation for ML-NTA

• Objective: To extract a broad range of compounds with sufficient recovery while minimizing matrix interference for downstream ML analysis.
• Materials: Solid phase extraction (SPE) systems, multi-sorbent materials (Oasis HLB, ISOLUTE ENV+, Strata WAX, WCX), QuEChERS kits, microwave-assisted extraction (MAE) systems.
• Procedure:
  • Select appropriate extraction technique based on sample matrix (water, soil, biological).
  • For broad-spectrum coverage, employ multi-sorbent SPE strategies combining complementary sorbents.
  • Perform extraction using optimized parameters (pH, solvent composition, volume).
  • Concentrate extracts under gentle nitrogen stream to appropriate volume.
  • Add internal standards for quality control of the extraction process.
• Quality Control: Include procedural blanks, replicates, and spiked samples to monitor contamination, precision, and recovery efficiency.

Stage 2: Data Generation and Acquisition

HRMS platforms, including quadrupole time-of-flight (Q-TOF) and Orbitrap systems, generate complex datasets essential for NTA [7].

Protocol 2.1: HRMS Data Acquisition for ML-Ready Datasets

• Objective: To generate high-quality, consistent HRMS data suitable for ML processing.
• Materials: LC-QTOF or LC-Orbitrap systems, quality control samples, data acquisition software.
• Procedure:
  • Perform chromatographic separation using reversed-phase or HILIC columns with appropriate gradients.
  • Acquire data in both positive and negative ionization modes with mass resolution >25,000.
  • Include data-dependent MS/MS acquisition for compound identification.
  • Inject quality control samples (pooled samples, solvent blanks) regularly throughout sequence.
  • Process raw data using software (e.g., XCMS, MS-DIAL) for peak picking, alignment, and componentization.
• Output: Structured feature-intensity matrix with rows representing samples and columns corresponding to aligned chemical features.

Stage 3: ML-Oriented Data Processing and Analysis

The transition from raw HRMS data to interpretable patterns involves sequential computational steps [7].

Table 1: Data Preprocessing Methods for ML-NTA

Processing Step	Technique Options	Purpose	Key Parameters
Missing Value Imputation	k-nearest neighbors, half-minimum	Handle missing values	k value, imputation method
Normalization	Total Ion Current (TIC), probabilistic quotient	Mitigate batch effects	Reference sample, method
Data Alignment	Retention time correction, m/z recalibration	Standardize features across batches	Alignment tolerance, reference
Noise Filtering	Blank subtraction, coefficient of variation	Remove irreproducible features	Blank threshold, CV cutoff

Protocol 3.1: Data Preprocessing Pipeline

• Objective: To transform raw feature-intensity data into a clean, normalized dataset ready for ML analysis.
• Software: Python/R with appropriate packages (scikit-learn, XCMS, CAMERA).
• Procedure:
  • Perform missing value imputation using k-nearest neighbors (k=5) for values missing in <50% of samples.
  • Apply TIC normalization to correct for overall signal intensity variations.
  • Conduct data alignment across batches using statistical algorithms for retention time correction and m/z recalibration.
  • Filter features present in blanks >30% of sample intensity or with high analytical variability (CV >30% in QCs).
  • Apply data scaling (autoscaling, Pareto scaling) as appropriate for subsequent ML algorithms.

Dimensionality Reduction and Pattern Recognition

Dimensionality reduction techniques simplify high-dimensional data, while clustering methods group samples by chemical similarity [7].

Table 2: ML Algorithms for NTA Data Analysis

Algorithm Category	Specific Methods	NTA Applications	Advantages
Unsupervised Learning	PCA, t-SNE, HCA, k-means	Exploratory data analysis, sample clustering	No labels required, reveals intrinsic patterns
Supervised Classification	Random Forest, SVM, Logistic Regression	Source attribution, sample classification	High accuracy, handles non-linear relationships
Feature Selection	Recursive feature elimination, variable importance	Identify marker compounds, reduce dimensionality	Improves interpretability, reduces overfitting

Protocol 3.2: Dimensionality Reduction and Classification

• Objective: To identify patterns in chemical profiles and build predictive models for sample classification.
• Software: Python/R with scikit-learn, SIMCA, or similar platforms.
• Procedure:
  • Perform Principal Component Analysis (PCA) to visualize sample clustering and identify outliers.
  • Apply t-distributed Stochastic Neighbor Embedding (t-SNE) for nonlinear dimensionality reduction.
  • Implement Random Forest classification with 100-500 trees to differentiate sample classes.
  • Use recursive feature elimination to identify most discriminative chemical features.
  • Validate model performance using cross-validation and independent test sets.

Tiered Validation Strategy for ML-Assisted NTA

Validation ensures the reliability of ML-NTA outputs through a three-tiered approach that bridges analytical rigor with real-world relevance [7].

Tier 1: Analytical Confidence Verification

Protocol 4.1: Analytical Validation Using Reference Materials

• Objective: To verify the accuracy of compound identification and quantification.
• Materials: Certified reference materials (CRMs), internal standards, spectral libraries.
• Procedure:
  • Analyze CRMs relevant to the sample matrix alongside experimental samples.
  • Confirm compound identities using Level 1-5 confidence rankings based on spectral matching.
  • Verify retention time stability and mass accuracy against known standards.
  • Calculate precision and accuracy for quantified compounds.

Tier 2: Model Generalizability Assessment

Protocol 4.2: External Validation of ML Models

• Objective: To evaluate model performance on independent datasets and assess overfitting risks.
• Materials: Independent sample sets from different batches or locations, validation software.
• Procedure:
  • Reserve 20-30% of samples as a hold-out test set before model training.
  • Apply trained models to completely independent external datasets.
  • Use k-fold cross-validation (k=5 or 10) to assess model stability [10].
  • Calculate performance metrics (accuracy, precision, recall, F1-score) on validation sets.

Tier 3: Environmental Plausibility Checks

Protocol 4.3: Contextual Validation with Environmental Data

• Objective: To correlate model predictions with contextual environmental information.
• Materials: Geospatial data, known source-specific chemical markers, historical contamination data.
• Procedure:
  • Compare model-predicted source contributions with known source locations.
  • Verify presence of established chemical markers for specific sources.
  • Assess temporal consistency of predictions with known emission patterns.
  • Evaluate quantitative relationships between predicted sources and measured concentrations.

Table 3: Tiered Validation Framework for ML-NTA

Validation Tier	Validation Components	Acceptance Criteria	Outcome Metrics
Tier 1: Analytical Confidence	CRM analysis, spectral matching, mass accuracy	Mass error < 5 ppm, RT stability < 0.2 min, spectral match > 80%	Identification confidence levels (1-5), quantification accuracy
Tier 2: Model Generalizability	Cross-validation, external validation, hold-out testing	Cross-validation accuracy > 80%, minimal performance drop on external sets	Accuracy, precision, recall, F1-score, ROC curves
Tier 3: Environmental Plausibility	Geospatial correlation, marker consistency, temporal trends	Statistical significance (p < 0.05) with contextual data	Correlation coefficients, spatial clustering, temporal patterns

Essential Research Reagent Solutions

Table 4: Key Research Reagents and Materials for ML-NTA Workflows

Reagent/Material	Function	Application Notes
Multi-sorbent SPE cartridges (Oasis HLB, Strata WAX/WCX)	Broad-spectrum compound extraction	Combine complementary sorbents for increased coverage of polar and non-polar compounds [7]
Certified Reference Materials (CRMs)	Analytical validation	Verify accuracy of identification and quantification for quality assurance [7]
Stable isotope-labeled internal standards	Quantification and process control	Correct for matrix effects and recovery variations during sample preparation
Quality control samples (pooled QCs, solvent blanks)	Monitoring analytical performance	Evaluate system stability, reproducibility, and contamination throughout sequences [7]
Retention time index standards	Chromatographic alignment	Standardize retention times across batches and instruments for consistent feature alignment [7]
MS tuning and calibration solutions	Instrument calibration	Ensure mass accuracy and sensitivity according to manufacturer specifications

Machine learning has transformed non-target analysis from a mere detection tool to an powerful interpretive framework for understanding complex environmental mixtures. The structured workflow and tiered validation strategy presented here provides researchers with a systematic approach for implementing ML-assisted NTA that balances innovation with analytical rigor. By adhering to these protocols and validation frameworks, researchers can generate chemically accurate, environmentally meaningful results that support informed decision-making in environmental monitoring, regulatory actions, and public health protection. Future advancements in explainable AI and integrated computational models will further enhance the applicability of ML-NTA in environmental risk assessment frameworks.

In machine learning-assisted non-target analysis (ML-assisted NTA), the journey from raw feature detection to the meaningful identification of contaminants is fraught with a critical bottleneck. This bottleneck encompasses the computational and methodological challenges in transforming high-dimensional chemical feature data into reliable, source-specific identifications that can inform environmental decision-making [7]. The high dimensionality of datasets significantly elevates computational costs and complicates the process of selecting relevant features, often resulting in suboptimal selections [11]. Furthermore, early NTA approaches that prioritized signal intensity risked overlooking low-concentration but high-risk contaminants and failed to account for source-specific chemical interactions [7]. This protocol outlines a systematic, tiered-validation framework designed to address this bottleneck, enhancing the reliability and interpretability of ML-NTA for researchers and drug development professionals.

Quantitative Data in ML-NTA Workflows

The ML-NTA workflow generates and relies on multifaceted quantitative data. The table below summarizes key performance metrics for the machine learning models used in the data analysis stage.

Table 1: Key Performance Metrics for ML Models in Contaminant Source Classification

Metric	Description	Typical Range in NTA Studies	Interpretation in NTA Context
Classification Accuracy	The correctness of the AI model's predictions in classifying contamination sources [12].	Balanced accuracy of 85.5% to 99.5% has been reported for PFAS source classification [7].	Must be balanced against other performance metrics like latency [12].
Latency	The time taken for an AI model to process an input and produce an output [12].	Critical for real-time applications; specific values are hardware and model-dependent.	Important for near-real-time monitoring applications.
Throughput	The number of tasks an AI system can handle within a given time frame [12].	Dependent on data complexity and computational resources.	Indicates the efficiency of processing large batches of HRMS samples.

The initial data acquisition stage produces a foundational quantitative dataset: a feature-intensity matrix. In this matrix, rows represent individual environmental samples, and columns correspond to the aligned chemical features detected by high-resolution mass spectrometry (HRMS), with cell values indicating the intensity or abundance of each feature [7].

Table 2: Quantitative Data Characteristics in HRMS-Based NTA

Data Aspect	Quantitative Measure	Impact on Analysis
Feature Dimensionality	Can encompass thousands to millions of chemical features [7].	Increases computational burden; necessitates robust feature selection.
Signal Intensity	Varies over several orders of magnitude between features.	Requires normalization; high-intensity features can dominate unsupervised analysis.
Confidence Levels	Assignment of Levels 1-5 for compound identification [7].	Provides a quantitative confidence score for identifications.

Experimental Protocols for a Tiered Validation Strategy

A tiered validation strategy is paramount to ensure that model outputs are both chemically accurate and environmentally meaningful. The following protocols provide a methodology for each tier.

Tier 1 Validation: Analytical Confidence Verification

Objective: To confirm the chemical identity of features prioritized by ML models. Materials: Certified reference materials (CRMs), commercial spectral libraries, quality control (QC) samples. Procedure:

• CRM Analysis: Analyze CRMs relevant to the suspected contamination sources alongside environmental samples. This verifies instrument calibration and retention time stability.
• Spectral Library Matching: Compare the acquired MS/MS spectra of prioritized features against commercial (e.g., NIST) and public (e.g, MassBank) spectral libraries. A match is typically considered confident (Level 1 identification) when the forward and reverse spectral similarity scores exceed a threshold (e.g., >70%) and the retention time is consistent with the standard [7].
• Quality Control: Inject batch-specific QC samples (e.g., pooled samples) throughout the analytical sequence to monitor instrument performance and data reproducibility [7].

Tier 2 Validation: Model Generalizability Assessment

Objective: To evaluate the performance and robustness of the trained ML classifier on independent data, ensuring it has not overfitted the training set. Materials: An external dataset not used during model training or hyperparameter tuning. Procedure:

• Data Splitting: Initially, split the annotated dataset into a training set (e.g., 70-80%) and a hold-out test set (e.g., 20-30%).
• Model Training & Cross-Validation: Train the selected ML model (e.g., Random Forest, Support Vector Classifier) on the training set. Use k-fold cross-validation (e.g., k=10) on this training set to optimize model parameters and obtain a preliminary performance estimate [7].
• External Validation: Apply the final, optimized model to the completely unseen hold-out test set. This provides an unbiased estimate of model performance.
• Performance Metrics Calculation: Calculate key metrics from Table 1 (e.g., balanced accuracy, precision, recall) for both the cross-validation and external validation sets. A significant drop in performance on the external set indicates overfitting.

Tier 3 Validation: Environmental Plausibility Check

Objective: To contextualize model predictions within real-world conditions and known source-receptor relationships. Materials: Geospatial data on potential emission sources, historical contamination data, literature on source-specific chemical markers. Procedure:

• Geospatial Correlation: Overlay the model-predicted contamination sources on a map with known locations of industrial, agricultural, or urban sites. High prediction probabilities for a specific source should correlate with proximity to that source type [7].
• Marker Compound Consistency: Investigate whether the features identified as important by the ML model (e.g., via Random Forest's feature importance) are known chemical markers for the predicted source. For example, the model might correctly identify certain PFAS compounds as highly indicative of fire-fighting foam runoff [7].
• Historical Comparison: Compare current model findings with historical data from the same or similar sites to assess the temporal plausibility of the identified sources.

Workflow and Logical Pathway Diagrams

The following diagram illustrates the integrated workflow of ML-assisted NTA, from sample collection to validated identification, highlighting the critical bottleneck and the tiered validation strategy designed to address it.

The logical relationship between the core feature selection bottleneck and the information bottleneck principle is further detailed in the following diagram.

The Scientist's Toolkit: Research Reagent Solutions

Successful execution of the ML-NTA workflow relies on a suite of essential reagents, software, and analytical resources.

Table 3: Essential Research Reagents and Materials for ML-NTA

Item Name	Function / Application	Specific Examples / Notes
Multi-Sorbent SPE Cartridges	Broad-spectrum extraction of analytes with diverse physicochemical properties from complex environmental matrices [7].	Oasis HLB, ISOLUTE ENV+, Strata WAX, Strata WCX [7].
Certified Reference Materials (CRMs)	Analytical confidence verification (Tier 1 Validation); used for instrument calibration and confirming compound identities [7].	Source-specific CRMs (e.g., PFAS mixtures, pesticide mixes).
Quality Control (QC) Samples	Monitoring data integrity and instrumental performance throughout the analytical sequence [7].	Pooled quality control samples, procedural blanks [7].
HRMS Platform with Chromatography	Data generation and acquisition; provides the high-resolution mass spectral data and chromatographic separation needed for NTA [7].	Orbitrap or Q-TOF systems coupled with LC or GC [7].
Information Bottleneck Feature Selection Tool	Addresses the feature selection bottleneck by globally optimizing the selection of a feature subset (Xs) that is maximally informative about the source labels (Y) [11].	Masked Deterministic IB (MDIB) neural network framework [11].
Spectral Libraries	Compound annotation and identification via spectral matching (Tier 1 Validation) [7].	NIST, MassBank, mzCloud.
ML Model Benchmarking Datasets	For training, testing, and benchmarking ML models for visualization and classification tasks [13].	VizNet [13], VizML [13].

In the realms of machine learning (ML)-assisted non-targeted analysis (NTA) and pharmaceutical development, validation transcends mere best practice—it constitutes an operational necessity. The convergence of black-box model complexity and pervasive data quality challenges creates a risk landscape where undiscovered errors can compromise scientific conclusions, regulatory submissions, and ultimately, patient safety. As models grow more sophisticated, traditional validation approaches become insufficient, necessitating a systematic, tiered strategy that spans from data acquisition to model deployment.

The stakes are substantial. In pharmaceutical research, data quality lapses have triggered regulatory application denials and significant market value erosion [14]. Similarly, ML models, particularly deep learning architectures, introduce unique vulnerabilities through their non-deterministic behavior and opacity, making standard validation protocols inadequate [15]. This application note establishes a comprehensive validation framework specifically designed for ML-assisted NTA research, providing experimentally-validated protocols to ensure reliability amidst these complexities.

The Dual Challenge: Black-Box Models and Data Quality

The Opacity Problem of Black-Box Models

Machine learning models, especially complex deep learning networks, often function as "black boxes" where the relationship between inputs and outputs lacks transparency. This opacity presents three critical validation challenges:

• Explainability Deficit: The inability to trace decision pathways hinders scientific acceptance and regulatory approval, particularly when models are used for critical decision-making in drug development [16].
• Unpredictable Failure Modes: Without transparent internal logic, models may fail unexpectedly on edge cases or data distributions not represented in training sets [15].
• Validation Complexity: Traditional sensitivity analysis becomes computationally intensive and difficult to interpret when inputs and outputs in neural network models lack the clear relationships found in statistical models [16].

The Data Quality Imperative

Data quality forms the foundational layer upon which all subsequent analysis rests. In pharmaceutical research and NTA studies, data challenges manifest uniquely:

• Incomplete Patient Records: Lead to potential misdiagnoses and flawed clinical trial outcomes [14].
• Inconsistent Drug Formulation Data: Generate errors in manufacturing and dosage calculations [14].
• Fragmented Data Silos: Hamper collaboration and real-time decision-making across research organizations [14].
• Data Drift: The gradual shift in real-world data distributions compared to training data causes model degradation over time, necessitating continuous monitoring [15].

Table 1: Documented Impacts of Poor Data Quality in Pharmaceutical Research

Issue Documented	Consequence	Source
FDA Application Denial	Clinical trial datasets lacking required nonclinical toxicology studies	[14]
Import Alert List Additions	93 companies flagged for drug quality issues including record-keeping lapses	[14]
Manufacturing Site Penalties	Inadequate documentation and quality control measures delaying drug approval	[14]

Tiered Validation Strategy for ML-Assisted NTA

A tiered validation strategy provides a structured approach to navigate the complexities of modern analytical pipelines. This multi-layered framework ensures comprehensive coverage from basic data quality to model performance in real-world scenarios.

The Four-Stage Workflow

ML-assisted NTA for contaminant source identification follows a systematic workflow comprising four critical stages [7]:

• Sample Treatment and Extraction: Balancing selectivity and sensitivity through purification techniques.
• Data Generation and Acquisition: Utilizing HRMS platforms to generate complex datasets for analysis.
• ML-Oriented Data Processing and Analysis: Applying computational methods to extract meaningful patterns.
• Result Validation: Implementing a multi-faceted approach to verify reliability and relevance.

The following diagram illustrates this comprehensive workflow and its key components:

Tiered Validation Protocol

The validation stage (Stage 4) implements a three-tiered approach to ensure comprehensive verification [7]:

• Tier 1: Analytical Confidence Verification

  • Objective: Confirm compound identities through analytical standards
  • Protocol: Use certified reference materials (CRMs) or spectral library matches
  • Acceptance Criteria: ≥95% match with reference spectra

• Tier 2: Model Generalizability Assessment

  • Objective: Evaluate performance on independent datasets
  • Protocol: Validate classifiers on external datasets using k-fold cross-validation
  • Acceptance Criteria: Balanced accuracy maintained within 5% of training performance

• Tier 3: Environmental Plausibility Checks

  • Objective: Correlate predictions with real-world context
  • Protocol: Compare model outputs with geospatial proximity to emission sources and known source-specific chemical markers
  • Acceptance Criteria: Statistical significance (p < 0.05) in expected directional relationships

Experimental Protocols and Application

Case Study: Toxicity-Based Prioritization Framework

An automated toxicity-based prioritization framework for NTA demonstrates the practical implementation of tiered validation [17]. This integrated workflow combines spectral matching, retention time prediction, and toxicity assessment to prioritize environmental pollutants.

Table 2: Experimental Protocol for Toxicity-Based Prioritization

Step	Methodology	Parameters Measured	Tools/Platforms
Sample Preparation	Solid phase extraction with multi-sorbent strategy	Analyte recovery rates	Oasis HLB, ISOLUTE ENV+
Data Acquisition	LC-QTOF-MS with MSE mode (DIA)	Retention time, m/z, intensity	High-resolution mass spectrometer
Data Processing	Spectral library searching, QSRR-based RT prediction	Spectral matching scores, RT accuracy	EPA ToxCast, ChemSpider, PubChem
Toxicity Assessment	Multi-endpoint toxicity prediction	ToxPi scores, 6 toxicity endpoints	EPA TEST software
Prioritization	Combined algorithm of multiple filters	Tier assignment (1-5)	NTAprioritization.R package

The workflow successfully processed a candidate list of 6,982 compounds from a sludge water sample, reducing it to a prioritized list of 2,779 compounds with 21 out of 28 spiked standards correctly identified and prioritized [17].

Visualization of the Prioritization Workflow

The toxicity-based prioritization framework integrates multiple data sources and analytical steps to efficiently identify compounds of concern:

The Scientist's Toolkit: Essential Research Reagents and Platforms

Table 3: Key Research Reagent Solutions for ML-Assisted NTA

Category	Specific Tools/Platforms	Function	Application Context
HRMS Platforms	Q-TOF, Orbitrap Systems	High-resolution mass detection for compound identification	Structural elucidation of unknown compounds [7]
Chromatography Systems	LC-ESI, GC-EI	Compound separation prior to mass analysis	Expanding chemical coverage for comprehensive analysis [17]
Spectral Libraries	EPA ToxCast, ChemSpider, PubChem	Reference databases for spectral matching	Compound identification and confirmation [17]
Toxicity Prediction	EPA TEST Software, ToxPi	Multi-endpoint toxicity assessment	Prioritization based on potential biological impact [17]
Data Processing	NTAprioritization.R Package	Automated prioritization workflow	Streamlined candidate evaluation and tier assignment [17]
Model Validation	Galileo Platform, Scikit-learn	Performance metrics tracking and model evaluation	Continuous validation and drift detection [18] [19]

Validation Metrics and Performance Assessment

Quantitative Validation Metrics

Establishing comprehensive performance metrics is essential for objective model assessment. The following metrics provide multidimensional evaluation:

Table 4: Performance Metrics for Model Validation

Metric Category	Specific Metrics	Optimal Range	Application Context
Classification Metrics	Accuracy, Precision, Recall, F1 Score	Domain-dependent (e.g., >0.85 for high-stakes)	Model performance evaluation [19] [20]
Model Discrimination	ROC-AUC	>0.8 (excellent), 0.7-0.8 (acceptable)	Binary classification tasks [19]
Regression Metrics	MAE, MSE, RMSE	Context-dependent (lower values preferred)	Continuous outcome prediction [20]
Clustering Quality	Silhouette Score, Davies-Bouldin Index	>0.5 (good clustering), lower values better for DBI	Unsupervised learning applications [20]
Toxicity Prediction	Balanced Accuracy	85.5-99.5% (as demonstrated in PFAS classification)	Contaminant source identification [7]

Addressing Common Validation Challenges

Even with robust metrics, validation faces practical challenges that require specific mitigation strategies:

• Data Scarcity: When labeled data is limited, employ data augmentation, transfer learning from pre-trained models, and active learning to prioritize informative samples [18].
• Class Imbalance: Address unequal class distributions with techniques like SMOTE (Synthetic Minority Over-sampling Technique), weighted loss functions, and stratified sampling [18].
• Conflicting Metrics: Establish a clear metric hierarchy based on business goals and apply multi-objective optimization to find balanced solutions [18].
• Concept Drift: Implement automated monitoring to detect changes in data distributions that require model retraining [15].

Validation in ML-assisted NTA research is not merely a technical checklist but a fundamental scientific principle that ensures research integrity and practical utility. The tiered validation strategy presented here provides a structured approach to navigate the complexities of black-box models and data quality challenges. By implementing these protocols—from analytical verification to environmental plausibility checks—researchers can build confidence in their findings and accelerate the translation of analytical data into actionable environmental and pharmaceutical insights.

The framework demonstrated in the toxicity-based prioritization case study, which successfully processed thousands of compounds while maintaining identification accuracy, showcases the practical implementation of these principles [17]. As ML applications continue to evolve in sophistication, so too must our validation methodologies, ensuring that scientific progress remains grounded in reliability and reproducibility.

Building the Workflow: A Step-by-Step Guide to ML-Oriented NTA and Prioritization Strategies

Within the framework of tiered validation strategy for machine learning (ML)-assisted non-target analysis (NTA), the initial stage of sample treatment and extraction is paramount. This step transforms a raw, complex environmental or biological matrix into a purified analyte mixture suitable for high-resolution mass spectrometry (HRMS). The quality and comprehensiveness of the data generated by HRMS are fundamentally limited by the efficacy of this first sample preparation stage. Consequently, the selection and execution of extraction protocols directly influence the performance of downstream ML models by determining the diversity and integrity of the chemical features available for pattern recognition and source attribution [7]. This document provides detailed application notes and protocols for comprehensive extraction techniques, designed to establish a robust foundation for reliable ML-assisted NTA.

Core Principles of Extraction

The fundamental goal of sample preparation is to isolate target analytes from interfering components in the sample matrix while ensuring high recovery and preserving the chemical integrity of the constituents. The extraction process generally follows these stages: (1) the solvent penetrates the solid matrix; (2) solutes dissolve into the solvent; (3) solutes diffuse out of the solid matrix; and (4) the extracted solutes are collected [21]. Several factors critically influence extraction efficiency and must be optimized for specific applications [21]:

• Solvent Selection: Governed by the principle of "like dissolves like," where solvent polarity should match the polarity of the target solutes. Methanol (MeOH) and ethanol (EtOH) are frequently used as universal solvents for phytochemical investigations [21].
• Particle Size: A smaller particle size enhances solvent penetration and solute diffusion, improving yield. However, excessively fine particles can lead to excessive solute absorption and filtration challenges [21].
• Temperature: Elevated temperatures can increase solubility and diffusion rates but must be balanced against the risk of degrading thermolabile compounds or causing solvent loss [21].
• Extraction Duration: Yield increases with time until equilibrium is reached between the solute concentration inside and outside the solid material [21].
• Solvent-to-Solid Ratio: A higher ratio generally increases extraction yield, but an excessively high ratio is inefficient, requiring large solvent volumes and prolonged concentration times [21].

A variety of techniques, from conventional to modern, are available for sample treatment. The choice of method depends on the sample matrix, the physicochemical properties of the analytes, and the requirements for throughput, selectivity, and solvent consumption. The table below provides a comparative summary of key extraction methods.

Table 1: Comparison of Common Extraction Techniques Used in Sample Preparation

Extraction Technique	Principle	Best For	Advantages	Disadvantages	Key Parameters
Maceration [21]	Solvent-assisted passive diffusion at room temperature.	Thermolabile compounds; simple setup.	Simple, low equipment cost.	Long extraction time, low efficiency.	Solvent type, particle size, soaking duration.
Percolation [21]	Continuous flow of fresh solvent through the sample bed.	Continuous processes; higher efficiency than maceration.	More efficient than maceration.	Can require more solvent than maceration.	Solvent flow rate, particle size, column packing.
Decoction [21]	Heating the sample in solvent, typically water.	Water-soluble, heat-stable compounds.	Efficient for hard plant tissues.	Not suitable for thermolabile or volatile compounds.	Boiling duration, pH, herb-to-water ratio.
Solid Phase Extraction (SPE) [7]	Selective adsorption/desorption of analytes from a liquid sample onto a solid sorbent.	Purification and concentration; selective class extraction.	High selectivity, clean-up, analyte enrichment.	Can be selective for certain properties, limiting broad coverage.	Sorbent chemistry (e.g., Oasis HLB, ENV+), wash/elution solvents.
Pressurized Liquid Extraction (PLE) [7]	Extraction with liquid solvents at elevated temperatures and pressures.	Fast and efficient extraction from solid matrices.	Fast, reduced solvent consumption, automated.	High equipment cost.	Temperature, pressure, solvent type, static/dynamic cycles.
Microwave-Assisted Extraction (MAE) [21] [7]	Heating the sample-solvent mixture via microwave energy.	Rapid heating and extraction.	Rapid, low solvent consumption, high yield.	Potential for non-uniform heating.	Microwave power, temperature, solvent dielectric constant.
Supercritical Fluid Extraction (SFE) [7]	Utilization of supercritical fluids (e.g., CO₂) as the extraction solvent.	Selective extraction of non-polar to moderately polar compounds.	Solvent-free (using CO₂), tunable selectivity, fast.	High equipment cost, limited for polar compounds.	Pressure, temperature, modifier addition.
QuEChERS [7]	"Quick, Easy, Cheap, Effective, Rugged, and Safe" method involving solvent extraction and salt-induced partitioning.	High-throughput multi-residue analysis (e.g., pesticides).	Rapid, high-throughput, minimal solvent.	May require further clean-up for complex matrices.	Salt mixtures, dispersive SPE sorbents for clean-up.

Detailed Experimental Protocols

Protocol 1: Solid Phase Extraction (SPE) for Broad-Range Contaminant Enrichment

This protocol is optimized for the extraction of a wide range of emerging contaminants from water samples, forming a foundational step for ML-NTA workflows [7].

1. Research Reagent Solutions

Table 2: Essential Materials for SPE Protocol

Item	Function
Oasis HLB SPE Cartridge (or equivalent)	Hydrophilic-Lipophilic Balanced copolymer sorbent for broad-spectrum retention.
ISOLUTE ENV+ / Strata WAX/WCX	Mixed-mode or ion-exchange sorbents used in a multi-sorbent strategy for expanded coverage.
HPLC-grade Methanol	Elution solvent for strongly retained analytes.
HPLC-grade Acetone	Elution solvent for a broader range of analytes.
Type 1 Water (LC-MS grade)	For sample preparation and cartridge conditioning.
Ammonium Formate / Acetate Buffer	For pH adjustment and ion-pairing in mobile phases.

2. Procedure

• Conditioning: Sequentially pass 5-10 mL of methanol (or acetone) followed by 5-10 mL of Type 1 water through the SPE cartridge. Do not allow the sorbent bed to dry out.
• Sample Loading: Load the acidified/pretreated water sample (e.g., 100 mL to 1 L) at a controlled, slow flow rate (e.g., 5-10 mL/min).
• Washing: After sample loading, wash the cartridge with 5-10 mL of a water-methanol mixture (e.g., 95:5 v/v) to remove weakly retained matrix interferences.
• Drying: Remove residual water by drawing air or nitrogen through the cartridge for 10-30 minutes, or by centrifugation.
• Elution: Elute the retained analytes with 5-10 mL of a strong organic solvent (e.g., methanol, acetone, or a mixture) into a collection tube.
• Reconstitution: Evaporate the eluent to near dryness under a gentle stream of nitrogen. Reconstitute the dried extract in an appropriate volume (e.g., 100-200 µL) of initial mobile phase (e.g., water/methanol) compatible with the subsequent LC-MS analysis. Vortex thoroughly and transfer to an autosampler vial.

Protocol 2: Pressurized Liquid Extraction (PLE) from Solid Matrices

This protocol is designed for the efficient extraction of organic contaminants from solid samples such as soil, sediment, or biological tissue [7].

1. Research Reagent Solutions

Table 3: Essential Materials for PLE Protocol

Item	Function
PLE System (e.g., Accelerated Solvent Extractor)	Automated system to maintain high temperature and pressure.
Diatomaceous Earth	Dispersant to mix with the sample for improved solvent contact.
Cellulose Filters	Placed at the ends of the extraction cell to prevent particulate clogging.
HPLC-grade Solvents (e.g., Acetone, Hexane, DCM)	Extraction solvents selected based on target analyte polarity.

2. Procedure

• Sample Preparation: Homogenize the solid sample and mix it thoroughly with an inert dispersant like diatomaceous earth in a defined ratio.
• Cell Packing: Place a cellulose filter at the bottom of the stainless-steel extraction cell. Pack the sample-dispersant mixture into the cell, avoiding voids. Top with another cellulose filter.
• Extraction: Place the cell in the PLE system. Set the extraction parameters. A typical method includes:
  • Solvent: Acetone:hexane (1:1 v/v) or other optimized mixtures.
  • Temperature: 100 °C.
  • Pressure: 1500 psi.
  • Heat Time: 5-10 minutes.
  • Static Time: 5-10 minutes.
  • Flush Volume: 60% of cell volume.
  • Purge Time: 60-90 seconds with nitrogen.
  • Cycles: 1-3 static cycles.
• Collection: The extracted analytes are collected in a sealed vial.
• Concentration: If necessary, concentrate the extract under a gentle nitrogen stream and reconstitute in a suitable solvent for analysis.

Workflow Integration and Visualization

The sample treatment and extraction stage is the critical first step in a multi-stage ML-assisted NTA workflow. The following diagram illustrates its position and relationship with subsequent stages, from data generation to final validation.

The specific choice of extraction technique dictates the chemical feature space that will be profiled. The diagram below outlines the decision-making process for selecting an appropriate technique based on the sample matrix.

This document outlines the detailed protocols for the data generation and acquisition stage within a tiered validation strategy for machine learning (ML)-assisted non-target analysis (NTA) using Liquid Chromatography-High-Resolution Mass Spectrometry (LC-HRMS). The quality and integrity of the data acquired at this stage are foundational for all subsequent ML and statistical analysis. Adherence to the standardized protocols described herein ensures the generation of consistent, high-fidelity data suitable for retrospective interrogation and model training [22] [23].

Essential Research Reagent Solutions

The following table details key reagents and materials essential for preparing and running HRMS samples, ensuring analytical reproducibility and accuracy.

Table 1: Key Research Reagent Solutions and Materials

Item	Function / Description
Internal Standards	Isotopically-labeled compounds spiked into all samples and calibration standards to monitor instrument performance, correct for matrix effects, and validate the analytical run [23].
Methanolic Standard Mixtures	Quality control (QC) samples containing approximately 100 reference compounds at known concentrations (e.g., 0.5 mg L–1) used to verify system suitability, sensitivity, and chromatographic performance [23].
Blank Matrices	Samples of the solvent or a blank biological matrix processed without analytes. Critical for identifying background contamination and ensuring the absence of carryover [23].
HPLC Grade Water	Ultra-pure water used as a control matrix and for preparing mobile phases and standard solutions to minimize background interference [22].
Structured Query Language (SQL) Database (e.g., ScreenDB)	A digital archive for parsed, peak-deconvoluted LC-HRMS data. Enables scalable, long-term storage and flexible, retrospective querying of vast datasets for NTA and method monitoring [23].
Laboratory Information Management System (LIMS)	A system, such as STARLIMS, for managing sample metadata, case characteristics, and complementary quantitative data, ensuring traceability and connectivity with HRMS data [23].

HRMS Instrumentation and Data Acquisition Specifications

Precise configuration of the LC-HRMS platform is critical for acquiring comprehensive data. The parameters below are based on established, scalable NTA workflows [23].

Table 2: Example LC-HRMS Instrument Configuration for NTA

Parameter	Specification
Chromatography	Reversed-phase liquid chromatography (RPLC)
Gradient Mode	Linear gradient (specific solvents and proportions should be defined per method)
Total Run Time	15 minutes [23]
Mass Spectrometer	Q-TOF (e.g., Xevo G2-S) [23]
Ionization Source	Electrospray Ionization (ESI), positive and/or negative mode
Acquisition Mode	Data-independent acquisition (DIA / MSE), collecting low and high collision energy spectra concurrently [23]
Data Archiving	Parsing of peak-deconvoluted data to an SQL database (e.g., ScreenDB) for long-term, queryable storage [23]

Experimental Protocols

Sample Preparation Protocol

Objective: To reproducibly extract and prepare samples for LC-HRMS analysis while maintaining analyte integrity. Materials: Samples, internal standards mixture, appropriate solvents (e.g., methanol, acetonitrile), HPLC-grade water, centrifuges, vortex mixer. Procedure:

• Sample Aliquoting: Precisely aliquot a defined volume or mass of each sample (e.g., 100 µL of plasma, 100 mg of tissue homogenate).
• Internal Standard Addition: Spike a known concentration of internal standards into every sample, calibration standard, and QC sample immediately prior to any preparation steps. This controls for variability in extraction and ionization [23].
• Protein Precipitation / Extraction: Add a precipitating solvent (e.g., cold acetonitrile, 3:1 v/v) to the sample. Vortex vigorously for 1-2 minutes.
• Centrifugation: Centrifuge at a high speed (e.g., 10,000-15,000 x g) for 10 minutes to pellet precipitated proteins and particulates.
• Supernatant Collection: Carefully transfer the clear supernatant to a new LC-MS compatible vial.
• Evaporation & Reconstitution: Evaporate the supernatant to dryness under a gentle stream of nitrogen or using a centrifugal evaporator. Reconstitute the dried residue in a defined volume of initial mobile phase or a suitable solvent mixture (e.g., 95:5 water:methanol). Vortex to ensure complete dissolution.
• Vialing: Transfer the final extract to an LC vial with insert for analysis.

Quality Control and System Suitability Protocol

Objective: To ensure the analytical system is performing within specified parameters and that the generated data is reliable. Materials: Methanolic standard mixtures, blank matrices, internal-standard blank injections [23]. Procedure:

• Sequence Design: Integrate QC samples at regular intervals throughout the analytical sequence (e.g., at the beginning, after every 10-12 experimental samples, and at the end).
• QC Sample Set: For each batch, analyze:
  • A blank matrix to check for contamination.
  • An internal-standard blank injection (solvent) to monitor instrumental background.
  • Methanolic standard mixtures containing ~100 compounds at 0.5 mg L^-1 to assess sensitivity, retention time stability, and mass accuracy [23].
• Data Review: Before proceeding with full data analysis, review the QC data. The batch should be accepted only if predefined forensic QC criteria are met (e.g., stable retention time and intensity for internal standards, acceptable mass error, absence of significant contamination) [23].

Data Acquisition and Pre-processing Protocol

Objective: To acquire raw HRMS data in a manner that captures maximum information and to convert it into a structured, queryable format. Materials: Prepared samples, configured LC-HRMS system, data processing software (e.g., UNIFI, XCMS), SQL database. Procedure:

• Data Acquisition: Inject samples according to the established sequence and LC-HRMS method (Table 2), acquiring data in a data-independent (M^E) mode to fragment all ions without precursor selection.
• Initial Data Processing: Process the raw data files through a peak-picking and componentization algorithm within the vendor's software (e.g., UNIFI). This step performs peak detection, groups co-eluting ions (including isotopes, adducts, and fragment ions) into components, and filters out non-chromatographic noise [23].
• Data Parsing to SQL Database: Export the componentized data and parse it into a structured SQL database (e.g., ScreenDB). This involves storing each ion signal (with its accurate mass, retention time, intensity, and link to its component) as a separate, queryable entry. This "decomponentized" storage is crucial for flexible, retrospective NTA [23].

Experimental Workflow and Data Flow Visualization

The following diagrams, generated using Graphviz, illustrate the logical flow of the experimental process and the subsequent data lifecycle.

Diagram 1: End-to-end workflow for HRMS-based sample analysis and data generation.

Diagram 2: Data flow from the SQL database to various applications, including ML training.

Within the systematic framework of Machine Learning-assisted Non-Target Analysis (ML-NTA) for contaminant source identification, Stage 3: ML-Oriented Data Processing serves as the critical computational bridge between raw analytical data and interpretable environmental insights [7]. This stage transforms the high-dimensional, complex data generated by high-resolution mass spectrometry (HRMS) into a structured format suitable for pattern recognition and machine learning modeling [7]. The primary objective is to extract meaningful chemical patterns and reduce data complexity while preserving diagnostically significant information essential for accurate contaminant source attribution [7]. The process is methodically sequenced into three core components: Data Preprocessing to ensure data quality and consistency, Dimensionality Reduction to mitigate the curse of dimensionality and enhance model generalization, and Clustering to uncover inherent group structures within the data without prior knowledge of sample labels [7]. The effective execution of this stage is a prerequisite for developing robust, interpretable, and generalizable ML models that can withstand rigorous tiered validation and provide actionable intelligence for environmental decision-making [7].

Data Preprocessing: Foundational Data Quality Assurance

Data preprocessing encompasses the initial set of operations designed to address data quality issues inherent in raw HRMS feature-intensity matrices, where rows represent samples and columns correspond to aligned chemical features [7]. This phase ensures the reliability and consistency of downstream analyses.

Core Preprocessing Techniques

The principal techniques employed in ML-NTA workflows include [24] [7] [25]:

• Missing Value Imputation: Data collections are frequently incomplete. Strategies include removing records with excessive missingness or estimating missing values using methods like k-nearest neighbors (KNN) imputation, which leverages similarities between samples to fill gaps intelligently [7].
• Noise Filtering: Low-abundance signals that are indistinguishable from instrumental noise are identified and removed to prevent them from obscuring genuine chemical patterns [7].
• Data Normalization: Techniques such as Total Ion Current (TIC) normalization are applied to correct for variations in overall signal intensity between samples, mitigating batch effects and making feature intensities comparable across the entire dataset [7].
• Data Alignment: Variations in analytical platforms or acquisition dates can cause misalignment. This process involves retention time correction, mass-to-charge ratio (m/z) recalibration, and peak matching to ensure chemical features are accurately aligned across all samples [7].

Table 1: Standardized Data Preprocessing Protocol for HRMS Data in ML-NTA

Processing Step	Standard Method/Protocol	Key Parameters & Considerations
Missing Value Imputation	k-Nearest Neighbors (KNN) Imputation [7]	- n_neighbors: Typically 5. - Distance metric: Euclidean. - Applied separately to each batch in cross-batch studies.
Noise Filtering	Abundance-based Thresholding [7]	- Remove features with intensity < 3x blank sample signal [7]. - Filter features present in < 10% of QC samples [7].
Data Normalization	Total Ion Current (TIC) Normalization [7]	- Normalize each sample's feature intensities to its total ion count. - Robust to high missing value rates.
Data Alignment	Retention Time Correction & Peak Matching [7]	- Algorithms: XCMS [7]. - Critical for cross-batch/lab studies. - Orbitrap may require more stringent alignment than Q-TOF [7].
Outlier Handling	Interquartile Range (IQR) Method [25]	- Identify outliers: Values < Q1 - 1.5IQR or > Q3 + 1.5IQR. - Decision: Remove, cap, or retain based on domain context [25].

Experimental Protocol: k-Nearest Neighbors (KNN) Imputation

Purpose: To replace missing values in the feature-intensity matrix with estimates derived from the most similar samples, preserving dataset structure and statistical power [7].

Procedure:

• Input: A peak table (samples × features) with missing values, denoted as NaN.
• Parameter Setting: Define the number of neighbors (k); a default of k=5 is often effective.
• Distance Calculation: For each sample containing a missing value in a specific feature, compute the Euclidean distance to all other samples based on the non-missing features.
• Neighbor Identification: Identify the k samples with the smallest Euclidean distance (the nearest neighbors).
• Value Imputation: Calculate the mean (for continuous data) or mode (for categorical data) of the target feature's values from the k nearest neighbors. Use this calculated value to replace the missing datum.
• Iteration: Repeat steps 3-5 until all missing values have been imputed.
• Output: A complete feature-intensity matrix ready for subsequent analysis.

Considerations: KNN imputation is computationally intensive for very large datasets. The choice of k and the distance metric can influence results and should be reported for reproducibility [7].

KNN Imputation Workflow: This protocol replaces missing values using similar samples.

Dimensionality Reduction: Addressing the Curse of Dimensionality

HRMS-based NTA datasets are characteristically high-dimensional, containing thousands of chemical features (dimensions) per sample. This creates the "curse of dimensionality," leading to data sparsity, increased computational cost, and a high risk of model overfitting [26]. Dimensionality reduction techniques counteract this by transforming the data into a lower-dimensional space while preserving its essential structure [26] [27].

Techniques and Selection Criteria

Two primary approaches are feature selection and feature extraction [26].

• Feature Selection identifies and retains a subset of the most relevant original features. This is valuable when interpretability is crucial, as the original feature meanings are retained. Methods include:

  • Filter Methods: Use statistical measures (e.g., correlation, ANOVA) to rank features independently of the model [26].
  • Wrapper Methods: Evaluate feature subsets based on their performance on a specific model (e.g., Recursive Feature Elimination) [26] [7].
  • Embedded Methods: Integrate feature selection within the model training process (e.g., Random Forest importance) [26] [27].

• Feature Extraction creates new, fewer features by transforming or combining the original ones. These new features often better capture underlying patterns, though they may lack direct interpretability [26].

Table 2: Comparative Analysis of Dimensionality Reduction Techniques for ML-NTA

Technique	Type	Key Principle	Advantages	Limitations	Ideal Use Case in ML-NTA
Principal Component Analysis (PCA) [26] [7] [27]	Linear Feature Extraction	Finds orthogonal axes (PCs) of maximum variance in the data.	- Simple, fast, deterministic. - Preserves global structure. - Good for initial exploration.	Assumes linear relationships. - Poor with complex nonlinear patterns.	Exploratory data analysis, visualization, preprocessing for linear models [26] [7].
t-SNE [26] [7]	Nonlinear Feature Extraction	Preserves local similarities by modeling pairwise probabilities.	- Excellent for visualizing complex clusters. - Captures nonlinear structures.	- Computational heavy. - Results depend on perplexity parameter. - Global structure not preserved.	Visualizing cluster separation and local sample relationships [26] [7].
Linear Discriminant Analysis (LDA) [26]	Supervised Feature Extraction	Maximizes separation between pre-defined classes.	- Optimal for classification. - Enhances class separability.	Requires labeled data. - Assumes normal data distribution.	Creating features for a classifier when sample sources are known [26].
Autoencoders [26]	Nonlinear Feature Extraction	Neural network that learns compressed data representation.	- Powerful for complex, nonlinear data. - Can handle very high dimensionality.	- "Black-box" nature. - Computationally intensive. - Requires large datasets.	Extracting features from highly complex NTA datasets when other methods fail [26].

Experimental Protocol: Principal Component Analysis (PCA)

Purpose: To reduce data dimensionality by transforming features into a set of linearly uncorrelated principal components (PCs) that capture the maximum variance in the data [26] [27].

Procedure:

• Input: A preprocessed and complete feature-intensity matrix.
• Standardization: Standardize the dataset so that each feature has a mean of 0 and a standard deviation of 1. This prevents features with larger scales from dominating the variance.
• Covariance Matrix Computation: Calculate the covariance matrix to understand how the features vary from the mean with respect to each other.
• Eigendecomposition: Compute the eigenvectors and eigenvalues of the covariance matrix. The eigenvectors represent the directions of the new feature space (Principal Components), and the eigenvalues represent the magnitude of the variance carried by each PC.
• Sorting: Sort the eigenvectors by decreasing eigenvalue. The eigenvector with the highest eigenvalue is the first principal component.
• Projection: Select the top k eigenvectors (where k is the desired number of dimensions) and project the original data onto this new subspace to obtain the lower-dimensional representation.

Considerations: The number of components to retain (k) is a critical choice. It can be determined by looking for an "elbow" in a Scree Plot (plot of eigenvalues) or by retaining enough components to explain a sufficiently high proportion (e.g., >95%) of the total cumulative variance [26].

PCA Procedure: This workflow reduces data dimensionality by identifying key variance directions.

Clustering: Discovering Inherent Group Structures

Clustering is an unsupervised machine learning technique that groups similar data points together based on their characteristics without using pre-defined labels [28]. In ML-NTA, this is pivotal for discovering natural groupings in the data, such as identifying samples that share a common contamination source or similar chemical profile [7].

Clustering Algorithms and Their Applications

The choice of algorithm depends on the expected data structure and the research question.

• Centroid-based (e.g., k-means): Organizes data around central points (centroids). It is efficient and scalable but requires pre-specifying the number of clusters (k) and is sensitive to outliers and non-spherical clusters [28] [29].
• Density-based (e.g., DBSCAN): Defines clusters as contiguous regions of high data density. Its key advantage is that it does not require k to be specified beforehand and can identify clusters of arbitrary shapes and noise points [28] [29].
• Hierarchical (e.g., HCA): Builds a tree of clusters (dendrogram), allowing visualization of relationships at multiple levels of granularity. It provides a rich view of data structure but can be computationally intensive for large datasets [28] [7] [29].
• Distribution-based (e.g., GMM): Assumes data is generated from a mixture of probability distributions. It provides probabilistic cluster assignments, which is useful for handling ambiguity, but requires specifying the number of distributions and can be computationally expensive [28] [29].

Table 3: Clustering Method Selection Guide for Environmental Sample Grouping

Algorithm	Core Mechanism	Key Parameters	Pros	Cons	NTA Application Context
k-Means [28] [29]	Iteratively assigns points to nearest of k centroids.	k (number of clusters).	- Simple, fast, scalable (O(n)) [29]. - Easy to interpret.	- Sensitive to initial centroid guess & outliers [28] [29]. - Assumes spherical, similar-sized clusters.	Initial, efficient grouping of samples where the approximate number of source types is known.
DBSCAN [28] [29]	Groups dense regions; labels sparse areas as noise.	eps (neighborhood radius), min_samples (core point definition).	- Finds arbitrary shapes. - Robust to outliers. - No need to specify k.	- Struggles with varying densities [28] [29]. - Parameter choice is critical.	Identifying core and outlier samples in spatial/temporal gradients with unknown cluster count [7].
Hierarchical (HCA) [28] [7] [29]	Builds a tree of clusters via merging/splitting.	Distance metric, linkage criterion.	- No need to specify k upfront. - Provides intuitive dendrogram. - Reveals data hierarchy.	- High computational cost (O(n²) typical) [29]. - Merging/splitting is irreversible.	Analyzing hierarchical source relationships (e.g., major source type -> sub-types) [7].
Gaussian Mixture Model (GMM) [28] [29]	Fits data as a mixture of Gaussian distributions.	n_components (number of distributions).	- Provides soft (probabilistic) clustering. - Flexible cluster shape (covariance).	- Sensitive to initialization. - Can overfit if n_components is too high.	Modeling samples with partial membership to multiple contamination sources.

Experimental Protocol: k-Means Clustering

Purpose: To partition n samples into k clusters, where each sample belongs to the cluster with the nearest mean (centroid), minimizing within-cluster variance [28].

Procedure:

• Input: A dataset (often the output from a dimensionality reduction step like PCA for better performance).
• Parameter Initialization: Choose the number of clusters k. Methods to inform this choice include the Elbow Method (plotting within-cluster sum of squares vs. k) or domain knowledge.
• Centroid Initialization: Randomly select k data points from the dataset as the initial centroids.
• Assignment Step: Assign each data point to the closest centroid based on a distance metric (typically Euclidean distance).
• Update Step: Recalculate the centroids as the mean of all data points assigned to that cluster.
• Iteration: Repeat steps 4 and 5 until the centroids no longer change significantly (convergence) or a maximum number of iterations is reached.
• Output: Cluster labels for each sample and the final centroid locations.

Considerations: k-means is sensitive to the initial random selection of centroids. It is good practice to run the algorithm multiple times with different initializations (n_init parameter) and use the result with the lowest within-cluster variance. The Elbow Method is a heuristic, not a definitive test for k [28].

k-Means Clustering: This algorithm partitions data into k clusters by minimizing variance.

The Scientist's Toolkit: Essential Research Reagents & Computational Solutions

This section details the critical software, libraries, and analytical tools required to implement the protocols described in this application note.

Table 4: Essential Research Reagents & Computational Solutions for ML-NTA Data Processing

Tool/Category	Specific Examples	Function in ML-Oriented Data Processing
Programming Languages & Core Libraries	Python, R	Primary languages for implementing the entire data processing pipeline, from data manipulation to model training and visualization.
Data Manipulation & Analysis	Pandas, NumPy (Python); dplyr (R)	Used for loading, cleaning, filtering, and transforming the feature-intensity matrix (e.g., handling missing values, normalization).
Machine Learning & Preprocessing	Scikit-learn (Python); caret (R)	Provides a unified interface for all major preprocessing techniques (imputation, scaling), dimensionality reduction algorithms (PCA, LDA), and clustering methods (k-means, DBSCAN, HCA). Essential for building reproducible pipelines.
Nonlinear Dimensionality Reduction & Advanced ML	t-SNE, UMAP, Autoencoders (e.g., using TensorFlow or PyTorch)	Specialized libraries for implementing complex feature extraction techniques that capture nonlinear patterns in the data, crucial for visualizing and understanding complex NTA datasets.
Data Visualization	Matplotlib, Seaborn (Python); ggplot2 (R)	Used to create diagnostic plots (e.g., boxplots for outliers, scree plots for PCA, dendrograms for HCA) and publication-quality figures to communicate results.
HRMS Data Processing Suites	XCMS [7]	Open-source software for the pre-processing of raw HRMS data, including peak detection, retention time alignment, and feature grouping, generating the initial input table for ML analysis.

Integrated Workflow and Tiered Validation Considerations

The three components of Stage 3 form a cohesive and sequential workflow. Data preprocessing ensures a high-quality, consistent dataset, which is a prerequisite for effective dimensionality reduction. Dimensionality reduction, in turn, simplifies the data and often reveals the underlying structure that clustering algorithms seek to group [7].

The outputs of this stage—whether a set of principal components, cluster assignments, or features selected by a supervised algorithm—directly feed into the final modeling and validation stages of the ML-NTA framework. It is paramount that the transformations and models developed in Stage 3 are validated rigorously within the proposed tiered strategy [7]. This includes using internal validation metrics (e.g., silhouette score for clustering), validating on external datasets, and, crucially, assessing the environmental plausibility of the discovered patterns (e.g., do the clusters correspond to known point source locations?) [7]. Furthermore, to ensure reproducibility and avoid data leakage, the entire data processing pipeline (including all parameters for imputation, scaling, and dimensionality reduction) must be fit exclusively on the training data and then applied to the validation and test sets [25].

ML-Oriented Data Processing Flow: This integrated workflow structures data for modeling.

Within a Machine Learning-assisted Non-Target Analysis (ML-NTA) framework, the stage of supervised learning represents a critical transition from exploratory data patterning to predictive modeling for definitive source identification. This phase leverages labeled sample data to train algorithms that can classify unknown contaminants to their originating environmental or industrial sources. The application of these models transforms high-dimensional chemical fingerprint data into actionable, attributable insights, which is a cornerstone for informed environmental decision-making and risk assessment [7]. The integration of this stage within a tiered validation strategy is paramount to ensure that model predictions are not only statistically robust but also environmentally plausible and reliable for regulatory purposes.

Core Principles and Data Prerequisites

Supervised learning models operate on the fundamental principle of learning a mapping function from input features (chemical signals from HRMS) to output labels (contamination sources) based on a set of training examples. The input is typically a feature-intensity matrix, where rows represent environmental samples and columns correspond to the aligned chemical features (e.g., m/z values at specific retention times) detected via HRMS [7]. The quality of the output is contingent on the quality and structure of the input data.

Table 1: Prerequisite Data Structure for Supervised Learning in ML-NTA

Data Component	Description	Role in Supervised Learning
Feature-Intensity Matrix	A structured table with samples as rows and aligned chemical features (intensities) as columns [7].	Serves as the input data (X) for the model training and prediction.
Source Labels	Categorical identifiers (e.g., "industrial effluent," "agricultural runoff") assigned to each sample based on known origin [7].	Serves as the target output (y) for classification models.
Training Set	A subset of the data with known source labels used to train the model.	Enables the algorithm to learn the unique chemical patterns associated with each source.
Test Set	A held-out subset of the data with known source labels used to evaluate model performance.	Provides an unbiased assessment of the model's generalizability to new, unseen data.

A critical preparatory step involves feature selection, which reduces the dimensionality of the data by identifying and retaining the most informative chemical features. Techniques such as recursive feature elimination enhance model performance by mitigating overfitting, improving computational efficiency, and increasing model interpretability. The selected features act as the diagnostic chemical fingerprint for each contamination source [7].

Algorithm Selection and Experimental Protocols

The choice of algorithm depends on the research goal, dataset size, and the desired balance between performance, interpretability, and computational complexity.

Commonly Used Classification Algorithms

Table 2: Supervised Learning Algorithms for Source Identification in NTA

Algorithm	Key Characteristics	Typical Use Case in NTA	Reported Performance
Random Forest (RF)	Ensemble method using multiple decision trees; robust to overfitting; provides feature importance metrics [7].	Identifying complex, non-linear interactions in source signatures; high-dimensional data [7].	Balanced accuracy of 85.5–99.5% for PFAS source classification [7].
Support Vector Classifier (SVC)	Finds the optimal hyperplane to separate classes in high-dimensional space; effective with clear margins of separation.	Distinguishing between sources with distinct chemical profiles.	Balanced accuracy of 85.5–99.5% for PFAS source classification [7].
Logistic Regression (LR)	A linear model that predicts class probabilities; highly interpretable.	Baseline modeling and when a linear relationship between features and source is assumed.	Balanced accuracy of 85.5–99.5% for PFAS source classification [7].
Partial Least Squares Discriminant Analysis (PLS-DA)	A dimensionality reduction technique combined with classification; effective for collinear data.	Identifying source-specific indicator compounds through variable importance metrics [7].	Widely used for biomarker and indicator compound discovery.

Protocol for Model Training and Evaluation

The following protocol outlines a standardized procedure for developing a supervised classification model for source identification.

Protocol: Building a Classifier for Contaminant Source Identification

Objective: To train and validate a supervised learning model that accurately classifies environmental samples to their contamination sources based on HRMS-derived chemical features.

Step 1: Data Preprocessing

• Input: Feature-intensity matrix from HRMS data processing [7].
• Procedure:
  • Missing Value Imputation: Address missing values using appropriate methods (e.g., k-nearest neighbors imputation) [7].
  • Normalization: Apply normalization techniques (e.g., Total Ion Current (TIC) normalization) to correct for systematic variations and batch effects [7].
  • Feature Filtering: Remove features with excessive missing values or low variance that are unlikely to be informative.

Step 2: Feature Selection

• Objective: Reduce data dimensionality and focus on the most discriminatory variables.
• Procedure: Apply feature selection algorithms such as Recursive Feature Elimination (RFE) or leverage model-specific importance metrics (e.g., from Random Forest) to identify a subset of key chemical features for classification [7].

Step 3: Dataset Splitting

• Procedure: Randomly split the preprocessed and feature-selected dataset into a training set (e.g., 70-80%) and a hold-out test set (e.g., 20-30%). The test set must not be used in any model training or parameter tuning to ensure an unbiased performance estimate.

Step 4: Model Training

• Procedure:
  • Select one or more algorithms from Table 2 (e.g., Random Forest, SVC).
  • Train each model using the training set. Employ cross-validation (e.g., 10-fold) on the training set to tune hyperparameters and prevent overfitting [7].

Step 5: Model Evaluation

• Procedure:
  • Use the trained model to predict the source labels for the hold-out test set.
  • Evaluate performance using metrics such as balanced accuracy (crucial for imbalanced datasets), confusion matrix, and Area Under the Curve of the Receiver Operating Characteristics (AUC-ROC) [7] [8].
  • For the best-performing model, analyze feature importance rankings to identify the chemical features most diagnostic of each source.

Integration within a Tiered Validation Strategy

The predictions from a supervised learning model must be rigorously validated within the broader tiered validation strategy of the ML-NTA workflow [7]. This moves beyond mere statistical validation to environmental and chemical plausibility.

Tier 1: Analytical Confidence Validation

• Verify the chemical identity of key discriminatory features identified by the model using certified reference materials (CRMs) or spectral library matches (confidence Levels 1-2) [7].

Tier 2: Model Generalizability Validation

• Assess the model on independent external datasets collected from different locations or time periods [7].
• Validate model predictions with other analytical techniques when possible.

Tier 3: Environmental Plausibility Validation

• Correlate model predictions with contextual data, such as geospatial proximity to known emission sources or the presence of source-specific chemical markers documented in the literature [7].
• This step bridges the gap between statistical output and real-world contamination scenarios, providing the rationale required for regulatory actions.

The following workflow diagram illustrates how supervised learning integrates into the complete ML-NTA process and is subjected to the tiered validation strategy.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Key Research Reagent Solutions for ML-NTA Workflows

Item	Function/Application
Quality Control (QC) Samples	Pooled quality control samples are analyzed intermittently with the environmental samples to monitor instrument stability, ensure data integrity, and correct for batch effects during data preprocessing [7].
Certified Reference Materials (CRMs)	Used in the tiered validation strategy (Tier 1) to confirm the identity and concentration of key discriminatory compounds identified by the model, providing analytical confidence [7].
Solid Phase Extraction (SPE) Cartridges	Used for sample clean-up and analyte enrichment during preparation. Multi-sorbent strategies (e.g., Oasis HLB, Strata WAX) are employed for broad-spectrum extraction of diverse contaminants [7].
LC-HRMS Instrument with Chromatography	Quadrupole Time-of-Flight (Q-TOF) or Orbitrap systems coupled with liquid chromatography (LC) are fundamental for generating the high-resolution mass spectrometric data required for NTA [7].
Structural Alert Databases	Computational resources like ToxAlerts, which contain known toxicophores, can be used to label and prioritize features with potential toxicity during model interpretation or as a separate filtering step [8].

Integrating Seven Key Prioritization Strategies for Efficient Feature Selection

In Machine Learning (ML)-assisted non-target analysis (NTA), the process of identifying unknown chemicals in complex environmental or biological samples generates high-dimensional datasets. Feature selection is a critical step in this workflow, as it helps to reduce data dimensionality, mitigate overfitting, and enhance model interpretability by identifying the most chemically relevant signals [30] [7]. This document outlines a structured framework integrating seven key prioritization strategies for efficient feature selection, specifically contextualized within a broader thesis on tiered validation strategies for ML-assisted NTA research. The protocols herein are designed for researchers, scientists, and drug development professionals working with high-resolution mass spectrometry (HRMS) data.

The Seven Key Prioritization Strategies: A Comparative Analysis

The following table summarizes the seven core feature prioritization strategies, adapting established frameworks from data science and product management to the context of ML-assisted NTA [30] [31] [32]. These strategies are categorized to help practitioners select the most appropriate method based on their specific research goal, data size, and computational resources.

Table 1: Seven Key Feature Prioritization Strategies for ML-Assisted NTA

Strategy Name	Core Principle	Best Suited NTA Scenario	Key Advantages	Key Limitations
Value vs. Complexity [31] [33]	Ranks features based on their perceived business value (e.g., importance for classification) versus the complexity to obtain or use them.	Preliminary filtering to identify "low-hanging fruit" – highly informative features that are easy to interpret.	Intuitive; facilitates quick initial data reduction.	Requires expert domain knowledge to assess value and complexity.
Weighted Scoring [31]	Uses a pre-defined, weighted scoring system across multiple criteria (e.g., abundance, fold-change, m/z uniqueness) to compute a composite feature score.	Prioritizing features when multiple, competing criteria are important for the research objective.	Enables objective, multi-faceted evaluation of features.	Defining weights and criteria can be subjective and requires careful calibration.
Kano Model [31] [33]	Classifies features into categories: Basic (must-haves), Performance (linear value increase), and Delighters (high-impact surprises).	Interpreting model results to understand which features are fundamentally important versus those that offer predictive advantages.	Shifts focus from mere presence to feature role and impact on model performance.	Better for post-hoc analysis and interpretation than for initial selection.
Minimum Redundancy Maximum Relevance (mRMR) [30]	Selects features that have high relevance to the target variable (e.g., source class) while maintaining low redundancy amongst themselves.	Building parsimonious models where multicollinearity is a concern and a compact, diverse feature set is desired.	Directly optimizes for relevance and diversity, mitigating correlation issues.	Computationally intensive for very large feature sets.
Univariate Statistical Filtering [32]	Evaluates features one at a time based on univariate statistical tests (e.g., ANOVA F-value, mutual information) against the target variable.	Rapid, large-scale screening of thousands of features to remove obvious non-informative ones.	Computationally fast and simple to implement; scales to very high dimensions.	Ignores feature interactions and correlations.
Recursive Feature Elimination (RFE) [32]	A wrapper method that recursively removes the least important features and re-builds the model to find the optimal subset.	Identifying a highly performant feature subset for a specific, chosen ML algorithm (e.g., SVM, Random Forest).	Often yields high-performing feature sets tailored to a specific classifier.	Computationally very expensive; prone to overfitting if not properly validated.
L1 Regularization (Lasso) [32]	An embedded method that uses L1 regularization during model training to push feature coefficients to zero, effectively performing selection.	Sparse model construction, especially with linear models or specific deep learning adaptations (e.g., First-Layer Lasso).	Integrates selection directly into the model training process.	The choice of the regularization parameter is critical and data-dependent.

Experimental Protocols for Strategy Implementation

Protocol for mRMR-Based Feature Selection

This protocol is designed for the mRMR R package or the mrmr_selection function in Python libraries like sklearn-feature-selection.

1. Data Preprocessing:

• Input: A feature-intensity matrix (samples x features) from HRMS data processing [7].
• Missing Value Imputation: Apply k-nearest neighbors (KNN) imputation to handle missing peak intensities. Use a small k (e.g., k=5) to preserve local data structure.
• Normalization: Perform total ion current (TIC) normalization to correct for sample-to-sample analytical variation [7].
• Transformation: Log-transform the data to stabilize variance and make the data distribution more symmetrical.

2. Strategy Execution:

• Parameter Setting: Set the max_features parameter to the maximum number of features to be selected.
• Algorithm Execution: Run the mRMR algorithm, which will compute the mutual information between each feature and the target variable (e.g., contamination source) as well as the mutual information between all feature pairs.
• Feature Ranking: The algorithm outputs an ordered list of features, starting with the one with the highest relevance and lowest redundancy, with subsequent features added to maximize relevance and minimize redundancy to the already selected set [30].

3. Output & Validation:

• The output is a ranked list of non-redundant, informative features.
• Validate the selected feature subset by training a downstream classifier (e.g., Random Forest) and evaluating its performance on a held-out test set using metrics like balanced accuracy [7].

Protocol for L1 Regularization (Lasso) in Deep Learning

This protocol outlines the application of Lasso, specifically the Deep Lasso variant, for feature selection with deep tabular models [32].

1. Model Configuration:

• Downstream Model: Select a deep learning architecture, such as an MLP or FT-Transformer.
• Integration of L1 Penalty: For Deep Lasso, apply an L1 penalty to the weights of the first layer of the network. This induces sparsity, effectively zeroing out the contributions of irrelevant features [32].
• Hyperparameter Tuning: Use a Bayesian optimization engine like Optuna to tune critical hyperparameters, including:
  • The regularization strength (λ) for the L1 penalty.
  • The learning rate and weight decay for the AdamW optimizer.
  • The architecture-specific parameters (e.g., hidden layer dimensions for an MLP).

2. Training & Selection:

• Training Regime: Train the model for a fixed number of epochs (e.g., 200) with an early stopping callback (e.g., patience=20) based on validation loss.
• Feature Importance: After training, the absolute values of the weights in the first layer of the network serve as the feature importance scores. Features connected to weights that have been forced to zero are considered irrelevant.
• Subset Selection: Rank features by their first-layer weight magnitudes and select the top k features for the final model.

3. Benchmarking:

• Compare the performance of the Deep Lasso-selected features against features selected by other methods (e.g., univariate filtering) by evaluating the downstream neural network's accuracy or RMSE on a reserved test set [32].

Workflow Visualization for ML-Assisted NTA

The following diagram illustrates the integrated workflow for ML-assisted Non-Targeted Analysis, highlighting the stages where the seven prioritization strategies are applied.

The Scientist's Toolkit: Essential Research Reagents and Computational Solutions

The following table details key reagents, software, and algorithms essential for implementing the described feature selection protocols in an ML-assisted NTA study.

Table 2: Key Research Reagent Solutions for ML-Assisted NTA

Item Name	Specification / Example	Primary Function in Workflow
Multi-Sorbent SPE Cartridges	e.g., Oasis HLB, Strata WAX/WCX, ISOLUTE ENV+ [7]	Broad-spectrum extraction of compounds with diverse physicochemical properties from complex matrices.
HRMS Instrumentation	e.g., Q-TOF, Orbitrap systems coupled with UHPLC [7] [34]	Generation of high-resolution, high-mass-accuracy data for detecting thousands of chemical features.
Data Preprocessing Software	e.g., XCMS, MZmine [7] [34]	Automated peak picking, retention time alignment, and componentization to create a feature-intensity matrix.
Programming Environment	Python with scikit-learn, XGBoost, PyTorch/TensorFlow [35] [7]	Provides the computational ecosystem for implementing ML models, feature selection algorithms, and custom scripts.
Hyperparameter Optimization Engine	e.g., Optuna [32]	Efficiently searches the hyperparameter space for both feature selection methods and downstream models to maximize performance.
Spectral Database	e.g., HMDB, NIST Tandem Mass Spectral Library [34]	Provides reference data for compound annotation and assigning confidence levels to identifications.
Certified Reference Materials (CRMs)	Source-specific analytical standards [7]	Used in the tiered validation strategy to confirm the identity and concentration of key marker compounds.

Integration with a Tiered Validation Strategy

The selected features must be validated within a robust, multi-tiered framework to ensure their chemical and environmental relevance [7].

• Tier 1: Analytical Confidence: Confirm the chemical identity of high-priority features using Certified Reference Materials (CRMs) or high-quality spectral library matches. This verifies that the ML model is prioritizing real, identifiable chemicals [7].
• Tier 2: Model Generalizability: Validate the entire ML model, trained on the selected features, using an independent external dataset. Perform k-fold cross-validation (e.g., 10-fold) to assess and mitigate the risk of overfitting [7] [32].
• Tier 3: Environmental Plausibility: Correlate the model's predictions and the selected features with contextual environmental data, such as geospatial information or known source-specific chemical markers. This step bridges the gap between statistical output and actionable environmental insight [7].

Navigating Pitfalls and Enhancing Performance in ML-NTA Workflows

In ML-assisted non-target analysis (NTA) for drug discovery, the reliability of biological insights is fundamentally constrained by pervasive data challenges. The integration of high-dimensional multi-omic data—a cornerstone of modern tiered validation strategies—is particularly vulnerable to technical noise, missing values, and batch effects, which can confound biological signals and lead to spurious predictions [36]. Technical noise and batch effects introduce non-biological variation that obscures true cellular expression patterns and complicates cross-dataset integration, while missing values can severely bias statistical estimates and model performance [37] [38]. Effectively mitigating these artifacts is therefore not a mere preprocessing step but a critical prerequisite for generating biologically valid, reproducible findings in computational biology and drug development. This protocol provides a comprehensive framework for identifying and correcting these data imperfections, establishing a robust foundation for downstream machine learning analyses and experimental validation within an NTA research paradigm.

Selecting optimal data correction strategies requires evidence-based decisions. The tables below summarize quantitative performance metrics for various imputation and batch-effect correction methods from recent benchmarking studies.

Table 1: Performance Comparison of Missing Value Imputation Methods

Imputation Method	Reported Performance (Dataset Context)	Key Metric(s)	Considerations for NTA
k-Nearest Neighbors (kNN)	Best for real-world product development data [39]	Model performance with Gradient Boosting	Robust for heterogeneous, real-world data structures.
MissForest	Best performance on healthcare diagnostic datasets [40]	RMSE, MAE	Effective for clinical/biological data; computationally intensive.
MICE	Second-best after MissForest on healthcare data [40]	RMSE, MAE	Flexible; good alternative; performance depends on chosen subroutine.
Bayes/Lasso	Best for generated (simulated) datasets [39]	Model performance with Gradient Boosting	May be optimal for data with well-defined underlying distributions.
Random Forest (in mice)	Weakest performance [39]	Model performance with Gradient Boosting	Not recommended as a primary choice based on current evidence.

Table 2: Performance of Batch-Effect Correction Strategies

Correction Method / Level	Application Context	Key Finding	Recommendation
Protein-Level Correction	MS-based Proteomics [41]	Most robust strategy across balanced and confounded designs.	Correct at the protein level after quantification for proteomics data.
iRECODE (with Harmony)	Single-cell RNA-seq [38]	Simultaneously reduces technical and batch noise; 10x more efficient than sequential correction.	Ideal for single-cell transcriptomics and other sparse, high-dimensional data.
Harmony	Single-cell RNA-seq & Multi-omics [38] [41]	Effective batch correction with good cell-type mixing (high iLISI) and identity preservation (cLISI).	A highly versatile and effective integration algorithm.
Ratio-based Scaling	MS-based Proteomics [41]	Superior prediction performance in large-scale plasma proteomics (T2D cohort).	Recommended for large-scale studies, especially with reference materials.

Experimental Protocols

Protocol 1: Comprehensive Missing Data Imputation with MissForest and kNN

This protocol details the steps for handling missing data using two top-performing methods, MissForest and kNN, suitable for healthcare and biological datasets commonly used in NTA research [39] [40].

Materials and Reagents:

• Software Environment: Python with pandas, numpy, scikit-learn, and missingpy packages [40].
• Input Data: A dataset (e.g., .csv format) with missing values encoded as NaN.
• Computational Resources: Standard desktop computer for small datasets (<10,000 features); high-performance computing (HPC) cluster may be required for large-scale omics data.

Procedure:

• Data Preprocessing and Partitioning:
  • Load the dataset and explicitly mark missing values.
  • For performance evaluation, if the original dataset is complete, simulate missing values under the Missing Completely at Random (MCAR) mechanism. Randomly introduce missingness at levels such as 10%, 20%, and 25% to create evaluation subsets [40].
• Imputation Execution:
  • For MissForest Imputation:
    • Use the MissForest algorithm from the missingpy package.
    • Utilize default parameters: criterion='mean_squared_error' and max_iter=10 [40].
    • Fit the model on the dataset with missing values and transform the data to obtain the imputed dataset.
  • For k-Nearest Neighbor Imputation:
    • Use the KNNImputer from scikit-learn.
    • Standardize the features prior to imputation to ensure equal weighting in the distance calculation.
    • Select the number of neighbors (e.g., k=5 is a common starting point) and a distance metric (Euclidean is default) [39].
    • Fit the imputer and transform the data.
• Model Validation and Selection:
  • On the datasets where missing values were simulated, compare the imputed values against the original ground truth.
  • Calculate performance metrics including Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE). Lower values indicate better accuracy [40].
  • Select the imputation method that yields the lowest error metrics for the final application.

Note on Workflow Order: Always perform data imputation before conducting feature selection. Imputing first ensures that the feature selection algorithm operates on a complete dataset, which leads to more stable and reliable selected feature sets [40].

Protocol 2: Integrated Noise and Batch-Effect Correction for Single-Cell Data with iRECODE

This protocol leverages the iRECODE platform to simultaneously address technical noise (dropouts) and batch effects in single-cell RNA sequencing data, a common challenge in NTA workflows [38].

Materials and Reagents:

• Software/Tool: RECODE/iRECODE platform (available from the original publication or repository).
• Input Data: A count matrix (genes x cells) from scRNA-seq experiments, with associated metadata specifying batch labels (e.g., sequencing run, patient, or lab).
• Computational Environment: R or Python environment as required by the tool. Ensure sufficient memory (≥16 GB RAM recommended for large datasets).

Procedure:

• Data Input and Preprocessing:
  • Format the gene expression count matrix and batch information according to iRECODE's input requirements.
  • The algorithm begins by applying Noise Variance-Stabilizing Normalization (NVSN) and mapping the data to an "essential space" using singular value decomposition [38].
• Integrated Correction Execution:
  • Within this essential space, iRECODE integrates a batch-correction algorithm, such as Harmony, to correct for non-biological variation across batches [38].
  • This dual approach allows for the simultaneous reduction of technical noise (addressing sparsity and dropouts) and batch effects while preserving the full dimensionality of the data.
• Output and Validation:
  • The output is a denoised and batch-corrected gene expression matrix.
  • Validation: Use the following metrics to assess correction efficacy:
    • Integration Score (iLISI): Measures batch mixing. A higher score indicates better integration of cells from different batches [38].
    • Cell-type LISI (cLISI): Measures biological signal preservation. A high score confirms distinct cell types remain separable post-correction [38].
    • Silhouette Score: Can be used to confirm that cell-type identities are preserved while batch clusters are merged [38].

Protocol 3: Protein-Level Batch-Effect Correction for MS-Based Proteomics

For mass spectrometry-based proteomics data within a tiered validation pipeline, this protocol outlines a robust protein-level correction strategy, as benchmarked on large-scale studies [41].

Materials and Reagents:

• Input Data: Protein abundance matrices derived from MS data, with batch information. Reference materials (e.g., Quartet project materials) are highly recommended for optimal correction [41].
• Software Environment: R or Python with appropriate packages (e.g., sva for ComBat, Harmony for Harmony).
• BECAs: Selected from algorithms like ComBat, Ratio, Harmony, or Median Centering [41].

Procedure:

• Data Quantification and Aggregation:
  • Generate protein-level abundance values from precursor or peptide-level intensities using a quantification method such as MaxLFQ, TopPep3, or iBAQ [41].
  • Crucial Note: Perform batch-effect correction after protein quantification. Evidence demonstrates that protein-level correction is more robust than correcting at the precursor or peptide level [41].
• Batch-Effect Correction Execution:
  • Apply the chosen BECA to the protein abundance matrix.
  • Example with Ratio-based Scaling: If using the Ratio method, normalize the intensity of each protein in a study sample by its corresponding intensity in a universally profiled reference sample [41].
  • Example with ComBat: Use the ComBat function from the sva R package, providing the protein matrix and batch covariate to remove batch-specific mean shifts [41].
• Performance Assessment:
  • For datasets with technical replicates: Calculate the Coefficient of Variation (CV). Successful correction will reduce the median CV across proteins [41].
  • For datasets with known sample groups: Use Principal Variance Component Analysis (PVCA) to quantify the proportion of variance explained by biological factors versus batch factors. A successful correction will minimize the variance component attributed to batch [41].
  • Signal-to-Noise Ratio (SNR) based on PCA can also be used to evaluate the resolution between known biological groups post-correction [41].

Workflow and Pathway Diagrams

Tiered Validation Strategy for ML-Assisted NTA

This diagram illustrates the overarching workflow, positioning data correction as the critical first step in a robust ML-assisted non-target analysis pipeline.

Integrated Single-Cell Data Correction with iRECODE

This diagram details the computational pathway of the iRECODE algorithm for simultaneous noise and batch-effect reduction.

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Computational Tools for Data Correction

Tool/Resource	Function	Application Context
RECODE/iRECODE Platform	Reduces technical noise and batch effects simultaneously.	Single-cell RNA-seq, scHi-C, spatial transcriptomics [38].
Harmony	Batch effect correction algorithm that iteratively clusters cells to remove technical variation.	Single-cell data, multi-omics data integration [37] [38] [41].
MissForest Algorithm	Imputes missing values using a random forest model.	Healthcare diagnostic data, biological datasets with complex correlations [40].
k-Nearest Neighbors (kNN) Imputer	Imputes missing values by averaging the k-most similar observations.	Real-world product development data, general-purpose imputation [39].
MICE (Multiple Imputation by Chained Equations)	Generates multiple imputed datasets to account for uncertainty.	Flexible framework for various data types; a robust alternative to single imputation [42].
ComBat	Empirical Bayes method for adjusting for batch effects in data.	Microarray, proteomics, and other genomic data [41].
Quartet Reference Materials	Commercially available reference samples for multi-omics.	Benchmarking and optimizing batch-effect correction in proteomics and other omics assays [41].

Balancing Sample Size and Feature Dimensionality for Robust Modeling

In machine learning (ML)-assisted non-target analysis (NTA) for drug discovery, achieving robust models requires a critical balance between sample size (N) and feature dimensionality (P). The "curse of dimensionality" is a pervasive challenge; high-dimensional data increases sparsity and computational demands, slowing algorithms and raising overfitting risks [43]. Simultaneously, insufficient samples yield models with high variance, lower statistical power, and reduced probability of reproducing true effects [44]. This application note details a tiered validation strategy, providing practical protocols and criteria to navigate this balance, ensuring model reliability and generalizability for researchers and drug development professionals.

Core Concepts and Challenges

The Small Sample Imbalance (S&I) Problem

The Small Sample Imbalance (S&I) problem occurs when a dataset has an insufficient number of samples (N ≪ M, where M is the standard dataset size for the application) and a significantly unequal class distribution [45]. This dual challenge leads to models that overfit and fail to generalize. In NTA research, where novel compound identification is key, this can manifest as an inability to distinguish true signals from noise or to identify rare but critical biological activities.

The Role of Dimensionality Reduction

Dimensionality reduction transforms high-dimensional data into a lower-dimensional space, preserving essential structures while mitigating overfitting and enhancing computational efficiency [43]. Techniques are broadly classified into:

• Feature Selection: Identifies and retains the most relevant features. This includes filter methods (using statistical measures), wrapper methods (assessing feature subsets via model performance), and embedded methods (integration within model training, e.g., LASSO) [46] [43].
• Feature Projection: Creates new, lower-dimensional features by transforming original data (e.g., PCA, t-SNE, UMAP) [47] [43].

Tiered Validation Strategy: A Framework for Robust Modeling

The following tiered strategy provides a structured approach to validate model robustness against sample size and dimensionality challenges.

Diagram 1: Tiered validation strategy workflow for robust ML modeling.

Experimental Protocols and Application Notes

Protocol 1: Evaluating Sample Size Adequacy

Principle: Systematically assess the impact of sample size on model performance and effect size to determine an adequate sample count [44].

Materials:

• A large, well-annotated dataset (e.g., from public repositories like NCBI GEO or the dataset described in [48]).
• Computing environment with Python (scikit-learn, NumPy) or R.

Procedure:

• Data Preparation: Begin with the largest available dataset. For a classification task, ensure class labels are available.
• Subsampling: Create a sequence of random subsamples without replacement, starting from a small size (e.g., N=16) and increasing incrementally (e.g., to N=2500) [44].
• Effect Size Calculation: For each subsample size, calculate both grand and average effect sizes.
  • Grand Effect Size: Uses the overall mean and variance of the entire dataset.
  • Average Effect Size: Uses the average of the means and variances of the individual classes [44].
• Model Training & Validation: On each subsample, train multiple ML classifiers (e.g., Support Vector Machine (SVM), Random Forest (RF), Neural Networks (NN), Logistic Regression (LR)) using a 10-fold cross-validation scheme [44].
• Performance Recording: Record the classification accuracy and the variance in accuracy for each subsample size and classifier.
• Criteria Application: A sample size is deemed adequate when it achieves:
  • Effect Size Criterion: Both average and grand effect sizes are ≥ 0.5.
  • Accuracy Criterion: Model classification accuracy is ≥ 80% [44].

Application Note: The point where increasing the sample size no longer yields a significant improvement in effect size or accuracy represents a good cost-benefit ratio for data collection.

Protocol 2: Dimensionality Reduction for scRNA-seq and High-Dimensional Biological Data

Principle: Apply and evaluate dimensionality reduction methods to project high-dimensional data (e.g., gene expression from scRNA-seq) into a lower-dimensional space for downstream analysis like clustering or lineage reconstruction [47].

Materials:

• scRNA-seq dataset (e.g., from Smart-Seq2 or 10X Genomics platforms) [47].
• Software/R packages: Seurat (for PCA), SC3 (for PCA), UMAP, FIt-SNE, ZINB-WaVE, scVI [47].

Procedure:

• Data Preprocessing: Perform standard quality control, normalization, and log-transformation on the gene expression count matrix.
• Method Selection: Select a suite of dimensionality reduction methods for comparison. A comprehensive evaluation suggests including methods from different categories [47]:
  • Linear: PCA, GLMPCA
  • Non-linear: Diffusion Map, t-SNE, UMAP, Isomap
  • Count & Zero-Inflation Aware: ZINB-WaVE, Poisson NMF, DCA
• Application: Apply each method to the preprocessed data to obtain a low-dimensional embedding (typically 2 to 50 dimensions).
• Evaluation Metrics: Evaluate the performance of each embedding using downstream analytical tasks:
  • Cell Clustering: Apply a clustering algorithm (e.g., k-means, Louvain) on the low-dimensional space and compare the resulting clusters to known cell-type labels using adjusted Rand index (ARI) or normalized mutual information (NMI).
  • Lineage Reconstruction: Use the embedding as input to trajectory inference tools (e.g., Monocle, TSCAN) and evaluate the accuracy of the inferred lineage against known biological pathways.
  • Neighborhood Preservation: Assess the method's ability to preserve the structure of the original high-dimensional space using metrics like neighborhood hit.
• Computational Scalability: Record the computational time and memory usage for each method, which is critical for large-scale datasets (e.g., >10,000 cells) [47].

Application Note: The choice of method involves a trade-off. PCA is highly scalable and often effective for initial clustering [47]. For visualizing complex cellular populations, non-linear methods like UMAP or t-SNE are superior [47] [43]. For data with significant dropout events, methods like ZINB-WaVE or DCA that explicitly model the count and zero-inflated nature of scRNA-seq data are recommended [47].

Protocol 3: Ensemble Feature Selection for Multi-Biometric Healthcare Data

Principle: Implement a scalable ensemble feature selection strategy to reduce dimensionality while retaining clinically relevant features for classification in heterogeneous healthcare datasets [46].

Materials:

• High-dimensional healthcare dataset (e.g., multi-modal data from the BioVRSea or SinPain datasets) [46].
• Python environment with scikit-learn.

Procedure:

• Feature Ranking: Use a tree-based model (e.g., Random Forest) to generate an initial ranking of all features based on their importance scores.
• Greedy Backward Elimination: Sequentially remove the least important feature from the ranked list. At each step, evaluate the performance of a classifier (e.g., SVM) using the current feature set via cross-validation.
• Subset Generation: This process produces multiple candidate feature subsets of varying sizes.
• Subset Merging: Apply a merging strategy (e.g., union or intersection of top-performing subsets) to consolidate the candidate subsets into a single, robust set of selected features.
• Validation: Train and validate final ML models (SVM, Random Forest) using the reduced feature set. Compare performance (e.g., F1 scores) and computational efficiency against models using the full feature set [46].

Application Note: This "waterfall selection" method effectively reduces feature count by over 50% while maintaining or improving classification metrics, making the models more computationally efficient and clinically interpretable [46].

Data Presentation and Analysis

Quantitative Comparison of Dimensionality Reduction Methods

Table 1: Evaluation of dimensionality reduction methods for scRNA-seq data analysis (adapted from [47]). Performance ratings are based on comprehensive benchmarking across 30 datasets for clustering and 14 datasets for lineage reconstruction.

Method	Modeling Counts	Modeling Zero Inflation	Non-Linear Projection	Computational Efficiency	Clustering Performance	Lineage Reconstruction Performance
PCA	No	No	No	High	Good	Fair
Poisson NMF	Yes	No	No	High	Good	Good
ZINB-WaVE	Yes	Yes	No	Low	Good	Good
UMAP	No	No	Yes	High	Very Good	Good
t-SNE	No	No	Yes	Medium	Very Good	Fair
Diffusion Map	No	No	Yes	Medium	Fair	Very Good
DCA	Yes	Yes	Yes	Medium	Very Good	Very Good

Sample Size and Model Performance Metrics

Table 2: Impact of sample size on classifier performance and effect size in a well-behaved arrhythmia dataset (adapted from [44]). Accuracy values are approximate and represent trends observed across multiple classifiers.

Sample Size (N)	Average Accuracy (%)	Variance in Accuracy	Grand Effect Size	Average Effect Size
16	68 - 98%	High	~0.8 (High Variance)	~0.8 (High Variance)
120	85 - 99%	Medium	~0.8	~0.8
1000	90 - 99%	Low	~0.8	~0.8
2500	90 - 99%	Very Low	~0.8	~0.8

The Scientist's Toolkit: Key Reagents and Materials

Table 3: Essential research reagents and computational tools for ML-assisted NTA studies.

Item Name	Function/Application	Example/Reference
PANC-1 Cell Line	A human pancreatic cancer cell line used for in vitro experimental validation of predicted drug synergies.	[48]
NCATS Dataset	A publicly available dataset containing single-agent and combination screening data for anti-cancer compounds.	[48]
Avalon/Morgan Fingerprints	Chemical structure descriptors used to represent compounds for machine learning models predicting drug synergy.	[48]
Seurat / SC3	Software toolkits for the analysis of single-cell RNA-sequencing data, including standard dimensionality reduction (PCA) and clustering.	[47]
Ensemble Feature Selection Pipeline	A scalable, open-source tool for reducing dimensionality in multi-biometric healthcare datasets while preserving clinical relevance.	[46]
Graph Convolutional Networks (GCNs)	A deep learning architecture demonstrated to achieve high hit rates in predicting synergistic drug combinations.	[48]
SHAP (Shapley Additive Explanations)	An interpretability technique used to quantify the contribution of each input feature to a model's prediction, aiding in biomarker discovery.	[49]

Workflow Visualization for Key Experiments

ML-Driven Drug Repurposing Workflow

The following diagram illustrates a multi-level validation strategy for identifying non-lipid-lowering drugs with lipid-lowering potential, integrating machine learning with clinical and experimental validation [50].

Diagram 2: Multi-tiered drug repurposing workflow.

The integration of machine learning (ML) into non-target analysis (NTA) and drug discovery has introduced a fundamental tension: the choice between highly accurate complex models and transparent, interpretable ones. This trade-off arises because simpler models, such as linear regression or decision trees, offer clear insights into their decision-making processes through easily understandable parameters but often achieve lower predictive performance. In contrast, complex models like deep neural networks and ensemble methods can capture intricate patterns in high-dimensional data at the cost of operating as "black boxes," where the rationale behind predictions is obscure [51]. In scientific fields such as environmental monitoring and drug development, where model predictions inform critical decisions about contaminant source identification or candidate drug selection, understanding why a model reaches a particular conclusion is not merely advantageous—it is often essential for regulatory acceptance, mechanistic validation, and building scientific trust [7] [52].

The challenge is particularly acute in ML-assisted NTA research, where models are tasked with identifying unknown contaminants or predicting compound properties from high-resolution mass spectrometry (HRMS) data. Here, the "incompleteness in problem formalization" means that achieving high classification accuracy is only part of the solution [52]. The broader scientific goal includes learning about environmental processes, identifying toxic chemical structures, and providing defensible evidence for regulatory actions. Consequently, a model's ability to explain its reasoning becomes as valuable as its predictive power, necessitating careful consideration of the interpretability trade-off within a robust tiered validation framework [7].

Conceptual Framework: Defining Interpretability and Explainability

In machine learning, interpretability refers to the degree to which a human can understand the cause of a model's decision, typically through direct inspection of the model's structure and parameters. An interpretable model, such as a shallow decision tree, provides transparent insight into its internal workings, mapping inputs to outputs in a way that is logically traceable [53] [52]. For instance, the coefficients in a linear regression model clearly indicate the direction and magnitude of each feature's influence on the prediction.

Explainability, on the other hand, often describes the ability of a model—even a complex one—to provide post-hoc, human-intelligible rationales for its specific predictions without necessarily revealing its underlying computational mechanisms [53]. Explainable AI (XAI) techniques, such as SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations), can generate approximate explanations for black-box models, but these are approximations that may not fully capture the model's true reasoning and add computational overhead [51].

The trade-off between model complexity and explainability is not merely a technical consideration but a foundational one that affects how scientific knowledge is extracted from data. As models become more complex to handle high-dimensional HRMS data or intricate biological interactions, their inner workings become less transparent, creating a gradient from inherently interpretable models to those requiring external explanation methods [51] [7].

Quantitative Comparison of ML Models in Scientific Applications

The performance characteristics of different machine learning models vary significantly across scientific applications, highlighting the practical implications of the interpretability-accuracy trade-off. The following table summarizes quantitative findings from environmental science and drug discovery case studies.

Table 1: Performance Metrics of ML Models in Scientific Applications

Application Domain	Model Type	Interpretability Level	Key Performance Metrics	Reference Study
PFAS Source Identification	Random Forest (RF)	Low (Black-box)	Balanced Accuracy: 85.5-99.5%	[7]
Aromatic Amine Structural Alert Prediction	Random Forest	Low (Black-box)	AUC-ROC: 0.82, True Positive Rate: 0.58	[8]
Organophosphorus Structural Alert Prediction	Neural Network	Very Low (Black-box)	AUC-ROC: 0.97, True Positive Rate: 0.65	[8]
Lipid-Lowering Drug Discovery	Multiple ML Models	Variable (Model-Dependent)	Clinical Validation Success (4 candidate drugs)	[54]

The data reveals that complex models can achieve high performance—the neural network for organophosphorus detection demonstrated excellent AUC-ROC—but this comes at the cost of transparency. In contrast, the drug discovery study employed a multi-model approach with subsequent experimental validation, emphasizing that model selection must align with the ultimate goal of generating scientifically valid and actionable results [54] [8].

Table 2: Tiered Validation Framework for ML-Assisted NTA

Validation Tier	Primary Objective	Key Methods & Techniques	Considerations for Interpretability
Analytical Confidence Verification	Confirm chemical identity of features	Certified Reference Materials (CRMs), spectral library matching	High interpretability enables direct mapping of model features to known chemical structures
Model Generalizability Assessment	Evaluate performance on independent data	External dataset validation, k-fold cross-validation	Simple models are less prone to overfitting and yield more reliable performance estimates
Environmental/Drug Action Plausibility	Correlate predictions with real-world context	Geospatial analysis, known source-specific markers, clinical data, animal studies	Interpretable models provide chemically plausible attribution rationale required for regulatory acceptance

Experimental Protocols for Model Development and Validation

Protocol 1: Developing a Structural Alert Classifier for NTA

This protocol outlines the methodology for creating ML models to predict hazardous structural alerts from tandem mass spectrometry (MS2) data, as demonstrated in PMC11924234 [8].

Materials and Software Requirements

• R programming environment (v4.2.1+) with caret package
• HRMS data from Q-TOF or Orbitrap systems (positive ionization mode, [M+H]+ adducts)
• MassBank Europe or equivalent spectral database
• ToxAlerts web server for structural alert identification

Experimental Procedure

• Data Curation and Labeling
  • Filter MassBank for MS2 spectra with Level 1 identification and available SMILES
  • Screen SMILES structures for presence of target structural alerts (e.g., aromatic amines, organophosphorus) using ToxAlerts
  • Label spectra as "alert present" or "alert absent" for binary classification

• Feature Engineering from MS2 Spectra

  • Extract fragment ions (m/z) with relative intensity >50 (base peak normalized to 999)
  • Calculate neutral losses by subtracting fragment m/z from precursor ion m/z
  • Bin fragments and neutral losses to nearest 0.1 m/z to reduce dimensionality
  • Create binary feature matrices (1=presence, 0=absence) for fragments and neutral losses

• Model Training and Optimization

  • Implement multiple algorithms (Random Forest, Neural Networks, etc.)
  • Use k-fold cross-validation (e.g., 10-fold) during training to optimize hyperparameters
  • Evaluate using AUC-ROC and true positive rate on held-out test set

• Application to Environmental Samples

  • Process unknown LC-HRMS features through the same preprocessing pipeline
  • Apply trained classifiers to prioritize features containing structural alerts
  • Validate identifications with orthogonal analytical techniques

Protocol 2: Tiered Validation Strategy for ML Models in Drug Discovery

This protocol describes the multi-level validation approach for ML-predicted drug candidates, as implemented in the lipid-lowering drug discovery study [54].

Materials and Reagents

• Clinical data repositories for retrospective analysis
• Animal models appropriate for the disease context
• Molecular docking software (e.g., AutoDock, GROMACS for dynamics)
• Cell-based assay systems for functional validation

Experimental Workflow

• Computational Prediction Phase
  • Curate training set of known active and inactive compounds from literature
  • Train multiple ML models (e.g., SVM, RF, Neural Networks) using chemical descriptors
  • Predict novel candidates from FDA-approved drug libraries

• Large-Scale Clinical Data Validation

  • Analyze electronic health records or insurance claims databases
  • Compare outcomes between patients taking candidate drugs versus controls
  • Adjust for confounding factors using propensity score matching

• Standardized Animal Studies

  • Administer candidate drugs to hyperlipidemic animal models (e.g., ApoE-/- mice)
  • Measure blood lipid parameters (LDL-C, HDL-C, triglycerides) at regular intervals
  • Conduct histopathological examination of relevant tissues

• Mechanistic Studies via Molecular Simulations

  • Perform molecular docking of candidate drugs to putative targets (e.g., PCSK9, HMG-CoA reductase)
  • Run molecular dynamics simulations (100+ ns) to assess binding stability
  • Validate binding affinities through surface plasmon resonance or isothermal titration calorimetry

The following workflow diagram illustrates the integrated process of model development and tiered validation within the context of ML-assisted NTA research.

Successful implementation of ML-assisted NTA research requires specialized materials and computational resources. The following table details key solutions for experimental and computational workflows.

Table 3: Essential Research Reagent Solutions for ML-Assisted NTA

Category	Specific Tool/Reagent	Function/Purpose	Application Context
Sample Preparation	Solid Phase Extraction (SPE)	Compound enrichment & matrix interference removal	Environmental sample cleanup prior to HRMS [7]
	QuEChERS	Rapid multi-residue extraction	High-throughput sample processing for large-scale studies [7]
HRMS Platforms	Q-TOF Mass Spectrometry	High-resolution accurate mass measurement	Structural elucidation of unknown compounds [7] [8]
	Orbitrap Mass Spectrometry	Ultra-high resolution & mass accuracy	Detection of complex mixture components [7]
Data Processing	XCMS	LC-HRMS data alignment & peak picking	Preprocessing of raw MS data for ML analysis [7]
	patRoon R package	NTA data processing workflow management	Streamlined data analysis from raw data to annotations [8]
ML Libraries	caret R package	Unified interface for ML model training & validation	Standardized implementation of multiple algorithms [8]
	SHAP/LIME	Post-hoc explanation of black-box models	Providing interpretability for complex models [51]
Validation Tools	Certified Reference Materials	Analytical confidence verification	Confirmation of compound identities [7]
	Molecular Docking Software	Binding mode prediction & mechanistic studies	Understanding drug-target interactions [54]

Strategic Framework for Model Selection in Tiered Validation

Navigating the interpretability trade-off requires a strategic framework that aligns model selection with research objectives, validation resources, and regulatory considerations. The following diagram outlines a decision pathway for choosing between complex and explainable models within a tiered validation context.

This framework emphasizes that model selection is not merely a technical optimization problem but a strategic decision with implications throughout the validation pipeline. In regulated environments or when mechanistic insight is paramount, inherently interpretable models provide the transparency needed for scientific validation and regulatory acceptance [7] [52]. When predictive performance is prioritized and adequate computational resources exist for explainable AI techniques, complex models with post-hoc explanations may be appropriate, provided their limitations are acknowledged within the validation framework [51] [55].

The interpretability trade-off in machine learning represents a fundamental consideration for scientific applications in non-target analysis and drug discovery. While complex models often achieve superior predictive performance on benchmark datasets, their black-box nature poses challenges for scientific validation, regulatory acceptance, and extracting mechanistic insights. A tiered validation strategy that incorporates analytical confidence verification, model generalizability assessment, and environmental or drug action plausibility testing provides a structured framework for evaluating ML predictions regardless of model complexity. By aligning model selection with research objectives and employing appropriate explanation techniques when needed, researchers can navigate the interpretability trade-off while maintaining scientific rigor and generating actionable results that advance environmental monitoring and therapeutic development.

In Machine Learning (ML)-assisted non-target analysis (NTA) research, the development of robust and generalizable models is paramount for translating complex chemical data into actionable environmental insights. This process hinges on two critical optimization pillars: feature selection and hyperparameter tuning [7] [56]. Feature selection mitigates the "curse of dimensionality"—a common challenge in high-dimensional data like mass spectrometry—by identifying and retaining only the most informative chemical features, thereby reducing noise and computational cost while enhancing model interpretability [56]. Hyperparameter tuning, conversely, systematically optimizes the configuration settings of the learning algorithm itself, which control the learning process and are set prior to training [57]. Effective tuning is essential to prevent overfitting, where a model performs well on training data but fails on unseen data, and to ensure the model can handle real-world variability [58] [59]. Within a tiered validation strategy for ML-NTA, these optimization tactics are not isolated steps but are deeply integrated into a iterative cycle of model validation and refinement, ensuring that the final model is both accurate and chemically plausible [7].

The following workflow diagram outlines the systematic integration of these optimization tactics within the broader ML-assisted NTA framework, emphasizing the iterative refinement cycle between feature selection, model training, hyperparameter tuning, and validation.

Feature Selection Algorithms

Feature selection is a critical data preprocessing step designed to reduce dimensionality by excluding irrelevant, redundant, or noisy features from the dataset [56]. In the context of NTA, this translates to selecting the most diagnostic chemical signals (e.g., specific mass-to-charge ratios or fragmentation patterns) that are indicative of a contamination source, while discarding thousands of non-informative signals [7]. This process enhances model performance, increases computational efficiency, and, crucially, improves model interpretability by isolating the features most relevant to the underlying environmental problem [56]. Feature selection methods can be broadly categorized into three groups, each with distinct mechanisms and advantages.

Table 1: Comparison of Feature Selection Method Categories

Category	Mechanism	Advantages	Disadvantages	Common Algorithms
Filter Methods	Selects features based on statistical measures of correlation or association with the target variable, independent of the ML model [56].	Computationally fast and scalable; less prone to overfitting; simple to implement [56].	Ignores feature interactions and dependencies; may select redundant features [56].	Chi-square (χ²) test, Analysis of Variance (ANOVA), Pearson correlation [7] [56].
Wrapper Methods	Evaluates feature subsets by using the performance of a specific ML model as the selection criterion. Involves a search strategy to find the best subset [56].	Considers feature interactions; typically results in high-performing feature sets for the chosen model [56].	Computationally expensive, especially with many features; higher risk of overfitting to the model [56].	Recursive Feature Elimination (RFE), Forward Selection, Backward Elimination [7].
Embedded Methods	Performs feature selection as an integral part of the model training process. The model itself inherently performs feature selection [56].	Balances performance and computation; considers feature interactions within the model [56].	Tied to the specific learning algorithm; may not be as transferable [56].	Random Forest (Gini importance/Mean Decrease in Impurity), LASSO (L1 regularization) [7] [56].

Application Protocol: Recursive Feature Elimination (RFE) with Cross-Validation

Recursive Feature Elimination (RFE) is a powerful wrapper method that is particularly effective for building parsimonious models in NTA. It works by recursively removing the least important features (as determined by a model's coef_ or feature_importances_ attribute) and re-evaluating the model until the optimal number of features is identified [7]. The following protocol details its application.

• Objective: To identify the minimal set of chemical features (e.g., aligned peaks from HRMS data) that yields the highest predictive accuracy for a source classification task.
• Prerequisites: A preprocessed feature-intensity matrix where samples are rows and aligned chemical features are columns. The target variable (e.g., contamination source) is defined and encoded.
• Materials & Computational Tools:
  • Programming Environment: Python with scikit-learn library.
  • ML Estimator: A base classifier such as Support Vector Classifier (SVC) or Random Forest.
  • Evaluation Metric: Accuracy or Balanced Accuracy for multi-class problems.

Step-by-Step Procedure:

• Initialize the RFE Model: Choose a base estimator and the desired number of features to select. A typical initial choice for the estimator is a linear SVC with a linear kernel.

• Integrate with Cross-Validation: To robustly determine the optimal number of features, use RFECV (RFE with cross-validation). It automatically tunes the number of features based on cross-validation performance.

• Train the Selector: Fit the RFECV object on the training dataset. It is critical to fit only on the training split to avoid data leakage.

• Analyze Results: After fitting, the object will contain the optimal number of features and a mask of the selected features.

• Model Training & Validation: Train your final model using only the selected features (X_train_selected) and evaluate its performance on the held-out test set (X_test_selected) as part of the tiered validation strategy.

Hyperparameter Tuning Techniques

Hyperparameter tuning is the process of finding the optimal configuration for a machine learning model's hyperparameters—the parameters set before the training process begins that control the learning itself [57]. In ML-NTA, proper tuning is not a luxury but a necessity to ensure model robustness and generalizability to new, unseen environmental samples [58] [7]. Techniques range from exhaustive searches to more intelligent, probabilistic methods.

Table 2: Comparison of Hyperparameter Tuning Techniques

Technique	Mechanism	Pros	Cons	Best-Suited Scenarios
Grid Search (GridSearchCV)	Brute-force method that tests all possible combinations within a pre-defined hyperparameter grid [57] [59].	Guaranteed to find the best combination within the grid; straightforward to implement and understand [57].	Computationally expensive and slow; becomes infeasible with large grids or high-dimensional spaces [57] [59].	Small, well-defined hyperparameter spaces where computational resources are not a primary constraint.
Random Search (RandomizedSearchCV)	Randomly samples a fixed number of hyperparameter combinations from specified statistical distributions [57] [59].	Faster than Grid Search; often finds good combinations with fewer computations; good for high-dimensional spaces [57].	Does not guarantee finding the absolute optimum; may miss the best combination if insufficient iterations are run [57].	Larger hyperparameter spaces where a good-enough solution is needed efficiently.
Bayesian Optimization	A smart, sequential model-based optimization (SMBO) that uses past evaluation results to choose the next hyperparameters to test, modeling P(score	hyperparameters) [57].	More efficient than grid/random search; requires fewer iterations to find high-performing combinations [57] [59].	More complex to set up and implement; higher computational cost per iteration [59].	Complex models with costly training cycles (e.g., deep learning), where every trial is expensive.

Application Protocol: Bayesian Optimization with Optuna

Bayesian Optimization represents the state-of-the-art in hyperparameter tuning, offering a superior trade-off between computational cost and performance. This protocol outlines its implementation using the Optuna library.

• Objective: To efficiently find the hyperparameter combination for a Random Forest classifier that maximizes cross-validation accuracy on an NTA dataset.
• Prerequisites: A training dataset (post-feature selection is recommended) and a defined objective metric.
• Materials & Computational Tools:
  • Programming Environment: Python with Optuna, scikit-learn.
  • Model: Any scikit-learn-style estimator (e.g., RandomForestClassifier, SVC).
  • Evaluation: K-fold cross-validation.

Step-by-Step Procedure:

• Define the Objective Function: This is the core function that Optuna will optimize. It takes a Optuna Trial object and returns the evaluation score.

• Create and Run the Study: A Study object orchestrates the optimization. The direction is set to 'maximize' for metrics like accuracy.

• Retrieve and Apply Best Parameters: After optimization, the best hyperparameters can be accessed and used to train the final model.

The Scientist's Toolkit: Research Reagent Solutions

The experimental and computational workflow for ML-assisted NTA relies on a suite of essential tools and reagents. The following table details key components, from chemical standards to software libraries, that form the foundation of a reproducible NTA study.

Table 3: Essential Research Reagents and Materials for ML-NTA

Item Name	Function/Application	Example Specifications/Notes
Mixed Sorbent SPE Cartridges	Broad-spectrum extraction of analytes with diverse physicochemical properties from complex environmental matrices (e.g., water, soil) [7].	Oasis HLB in combination with ISOLUTE ENV+, Strata WAX, or WCX to maximize chemical space coverage [7].
Certified Reference Materials (CRMs)	Analytical confidence verification; used for retention time alignment, mass accuracy calibration, and quantitative validation in Tier 1 of the validation strategy [7].	Commercially available mixes relevant to the study focus (e.g., PFAS, pharmaceuticals).
HRMS Instrumentation	Data generation and acquisition; provides high-resolution mass and fragmentation data for compound annotation and structural elucidation [7].	Quadrupole Time-of-Flight (Q-TOF) or Orbitrap mass spectrometers coupled with LC or GC [7].
Data Preprocessing Software	Converts raw HRMS data into a structured feature-intensity matrix through peak picking, alignment, and componentization [7].	Vendor-specific software (e.g., Agilent MassHunter) or open-source platforms (e.g., XCMS, MS-DIAL).
Python with Scikit-learn	Core programming environment for implementing feature selection algorithms, machine learning models, and hyperparameter tuning [57] [56].	Essential libraries: scikit-learn, pandas, numpy, optuna.
Visualization Libraries (Matplotlib, Graphviz)	Generation of diagnostic plots (e.g., confusion matrices, feature importance) and workflow diagrams for model interpretation and communication [60].	matplotlib, seaborn, and graphviz facilitate the creation of publication-quality figures.

Ensuring Credibility: Implementing a Robust Multi-Tiered Validation Framework

Tier 1 Confidence Framework and Quantitative Criteria

Table 1: Method Criteria and Threshold Metrics for Tier 1 (Confirmed) Confidence Level [61]

Component	Requirement
Native Standards	Analyte-specific
Labeled Internal Standards	Analyte-specific
Calibration Curve	Multipoint (≥6 levels), internal
Accuracy	± Up to 20% RSD
Intrabatch Variability	≤15%
Interbatch Variability	≤15%
Quantification Confidence Indicator	Confirmed

Tier 1 represents the highest confidence level in analytical measurement, applicable when authentic reference standards are analyzed concurrently with samples, with matching exact mass, isotope pattern, retention time, and MS/MS spectrum [61]. The calibration curve must have an r² > 0.95 and cover the range of study samples [61]. Accuracy is calculated from standard reference materials (SRM), such as National Institute of Standards and Technology (NIST) samples, proficiency testing materials, or well-characterized pools used by multiple labs [61].

Experimental Protocol for Tier 1 Verification

Sample Preparation and Analysis

• Sample Treatment: Employ balanced purification techniques such as solid phase extraction (SPE) to remove interfering components while preserving a wide range of compounds. Multi-sorbent strategies (e.g., combining Oasis HLB with ISOLUTE ENV+, Strata WAX, and WCX) can achieve broader-range extractions [7].
• Internal Standard Application: Use isotope dilution with a labeled version of the analytes of interest in every batch [61].
• Calibration Standards Preparation: Prepare a matrix-matched (recommended) or non-matrix-matched calibration curve with at least six nonzero levels. The non-matrix-matched curve must perform within 5% deviation of the matrix-matched slope [61].
• Instrumental Analysis: Perform analysis using High-Resolution Mass Spectrometry (HRMS) platforms, such as quadrupole time-of-flight (Q-TOF) or Orbitrap systems, coupled with liquid or gas chromatographic separation (LC/GC) [7].

Data Acquisition and Processing

• Confirmation Parameters: Confirm analyte identity by matching the exact mass, isotope pattern, retention time, and MS/MS spectrum of the reference standard to the sample analyte [61].
• Quality Control: Inject quality control (QC) samples into every analytical run to monitor precision. Calculate intrabatch and interbatch variability from any QC sample measured repeatedly throughout the analysis [61] [7].
• Data Processing: Process acquired data through steps including centroiding, extracted ion chromatogram (EIC/XIC) analysis, peak detection, alignment, and componentization to group related spectral features into molecular entities [7].

Workflow Visualization for ML-Assisted Tier 1 Verification

Tier 1 Verification Workflow

Research Reagent Solutions for Tier 1 Analysis

Table 2: Essential Research Reagents and Materials [61] [7]

Reagent/Material	Function
Analyte-Specific Native Standards	Unlabeled authentic chemical standards used for calibration curve preparation and definitive identification via RT and MS/MS matching.
Analyte-Specific Labeled Internal Standards	Isotope-labeled (e.g., ¹³C, ²H) versions of the analyte; used for isotope dilution to correct for matrix effects and losses during sample preparation.
Standard Reference Materials (SRM)	Certified materials (e.g., from NIST) used to independently verify method accuracy and performance.
Matrix-Matched Calibration Standards	Calibration standards prepared in a sample-like matrix to correct for matrix-induced ionization effects.
Quality Control (QC) Samples	Pooled samples or control materials analyzed repeatedly within and across batches to monitor precision (intrabatch and interbatch variability).
Multi-Sorbent SPE Cartridges	Solid-phase extraction materials with different functional groups (e.g., HLB, WAX, WCX) for broad-spectrum extraction of diverse analytes from complex matrices.

In Tier 2 of a machine learning (ML)-assisted non-target analysis (NTA) validation strategy, the focus shifts from internal model performance to rigorous assessment of model generalizability—the ability of a model to maintain predictive accuracy on new, external data it has not encountered during training [7]. This tier is critical for establishing confidence that models will perform reliably in real-world deployment scenarios, beyond the controlled conditions of initial development. Generalizability testing guards against the deployment of overfitted models that excel on training data but fail with new samples, a common challenge in analytical applications [62] [63].

Within the broader tiered validation framework for ML-assisted NTA research, Tier 2 acts as a crucial bridge between internal validation (Tier 1) and comprehensive external testing (Tier 3). It employs two complementary approaches: cross-validation techniques that maximize the utility of available data for robustness estimation, and external dataset validation that provides the most realistic assessment of how models handle truly novel data [7] [64]. For drug development professionals and environmental scientists relying on NTA for contaminant source identification or chemical fingerprinting, establishing proven generalizability is a prerequisite for regulatory acceptance and operational deployment.

Core Principles of Model Generalizability

Definition and Importance

Model generalizability refers to a machine learning model's capacity to make accurate predictions on data drawn from the same underlying population as the training data but not used in model development [65]. In ML-assisted NTA, this translates to reliably identifying and quantifying unknown compounds in new environmental samples, clinical specimens, or pharmaceutical products. A highly generalizable model maintains performance when confronted with variations in sample matrices, instrumental conditions, and chemical profiles that inevitably occur in practice [7].

The importance of generalizability assessment stems from the fundamental risk of overfitting, where models learn patterns specific to training data—including noise and irrelevant correlations—rather than underlying relationships that hold universally [62] [64]. This is particularly problematic in NTA applications where models may be deployed across multiple analytical platforms, geographic locations, or temporal periods. Without proper generalizability assessment, models may produce misleading results with significant consequences for environmental health decisions or drug development processes [63].

Common Threats to Generalizability

Multiple methodological pitfalls can compromise model generalizability, often remaining undetectable during internal evaluation while causing significant performance degradation in real-world use [63]:

• Violation of Independence Assumption: Data leakage occurs when information from the test set inadvertently influences model training, creating overoptimistic performance estimates [63]. Common causes include performing data preprocessing, normalization, feature selection, or augmentation before splitting data into training and test sets [62] [63].
• Inappropriate Performance Metrics: Selecting evaluation metrics that do not align with the analytical task or failing to establish appropriate performance baselines can misrepresent true model capability [63].
• Batch Effects and Dataset Shift: Systematic differences between training and deployment data distributions—due to changes in instrumentation, sample preparation protocols, or population characteristics—can severely degrade performance [63] [66]. For example, a clinical ML model for pneumonia detection achieved an F1 score of 98.7% on its original test set but correctly classified only 3.86% of samples from a new dataset of healthy patients due to batch effects [63].
• Non-Representative Data Splits: When training and test sets do not adequately represent the full data distribution—particularly problematic with small datasets or hidden subclasses—performance estimates become unreliable [64].

Cross-Validation Techniques for Generalizability Assessment

Cross-validation (CV) comprises a set of resampling techniques that systematically partition available data to simulate how models will perform on unseen data [67] [65]. By repeatedly holding out portions of data for testing while training on remaining samples, CV provides a more robust estimate of generalization error than a single train-test split [62]. In ML-assisted NTA, CV is employed for three primary purposes: (1) performance estimation—predicting how a model will generalize to new data; (2) algorithm selection—comparing different modeling approaches; and (3) hyperparameter tuning—optimizing model configuration settings [64].

Key Cross-Validation Methods

Table 1: Comparison of Major Cross-Validation Techniques

Method	Procedure	Advantages	Disadvantages	Recommended Use Cases
k-Fold Cross-Validation [62] [65]	Data randomly partitioned into k equal folds; each fold serves as validation once while k-1 folds train	Uses all data for training and validation; lower variance than holdout; computa-tionally efficient	Training folds overlap; performance may vary with different random partitions	Standard choice for most NTA applications with sufficient sample size
Stratified k-Fold [62] [65]	Preserves class distribution percentages in each fold	Maintains repre-sentative splits with imbalanced data	More complex implementation; requires class labels	NTA with rare compounds or unequal class distributions
Leave-One-Out Cross-Validation (LOOCV) [62] [67]	Each single sample serves as validation set once	Virtually unbiased; uses maximum data for training	High computational cost; high variance in estimation	Small datasets (<100 samples) in screening applications
Repeated k-Fold [62]	Multiple rounds of k-fold with different random partitions	More reliable performance estimate	Increased computation time	When dataset variability concerns exist
Nested Cross-Validation [62] [64]	Inner loop for hyperparameter tuning, outer loop for performance estimation	Unbiased performance estimation with hyperparameter tuning	Computationally intensive	Model selection and tuning when data permits
Hold-Out Validation [65] [64]	Single split into training and test sets (typically 70-80%/20-30%)	Simple, fast implementation	High variance; dependent on single split	Very large datasets (>10,000 samples)

Implementing Cross-Validation in ML-Assisted NTA

The following protocol describes the implementation of k-fold cross-validation, the most widely applicable approach for ML-assisted NTA applications:

Protocol 1: k-Fold Cross-Validation for Model Assessment

• Data Preparation: Preprocess the complete dataset following established NTA pipelines, including peak detection, alignment, and normalization [7]. Ensure data integrity through quality control measures.
• Fold Generation: Randomly partition the dataset into k folds (typically k=5 or k=10) of approximately equal size [62] [64]. For stratified k-fold, maintain similar distribution of key characteristics (e.g., sample type, contamination level) in each fold.
• Iterative Training and Validation:
  • For i = 1 to k:
  • Designate fold i as the validation set
  • Combine remaining k-1 folds as the training set
  • Train the model on the training set
  • Apply the trained model to the validation set
  • Calculate performance metrics on the validation set
• Performance Aggregation: Compute the mean and standard deviation of performance metrics across all k iterations [67] [65].
• Final Model Training: After completing cross-validation and selecting optimal hyperparameters, train the final model using the entire dataset for deployment [68].

Table 2: Performance Metrics for Cross-Validation in NTA Applications

Metric	Formula	Application Context
Accuracy	(TP+TN)/(TP+TN+FP+FN)	Balanced classification tasks
F1-Score	2×(Precision×Recall)/(Precision+Recall)	Imbalanced data situations
Area Under ROC Curve (AUC)	Integral of ROC curve	Binary classification performance across thresholds
Mean Squared Error (MSE)	Σ(yᵢ-ŷᵢ)²/n	Regression tasks (concentration prediction)
R² Score	1 - Σ(yᵢ-ŷᵢ)²/Σ(yᵢ-ȳ)²	Proportion of variance explained

Tier 2 Cross-Validation Workflow

External Validation with Independent Datasets

The Critical Role of External Validation

While cross-validation provides valuable insights into model robustness, external validation using completely independent datasets represents the gold standard for assessing true generalizability [63] [66]. External validation tests a model's ability to perform on data collected at different times, by different instruments, or from different populations than the training data—directly simulating real-world deployment conditions [69]. This approach is particularly crucial for ML-assisted NTA applications intended for regulatory decision-making or cross-institutional use.

Research demonstrates that models exhibiting excellent performance during internal validation can fail dramatically when applied to external datasets. For instance, a clinical ML model for predicting hospital admission from emergency department data showed AUC performance ranging from 0.84 to 0.94 across different sites when trained on pooled data, but performance varied more widely (AUC 0.71 to 0.93) when site-specific models were tested across all sites [69]. These site-specific performance differences highlight the critical importance of external validation.

Designing External Validation Studies

Protocol 2: External Validation for Generalizability Assessment

• Dataset Curation:

  • Collect external validation datasets that represent the anticipated deployment environments
  • Ensure external datasets remain completely separate from training/validation data throughout model development
  • Document key metadata including sampling protocols, analytical instrumentation, and processing methods

• Model Application:

  • Apply the fully-trained model (trained on the complete development dataset) to the external dataset
  • Use identical preprocessing pipelines for both training and external data
  • Avoid any model adjustments or parameter tuning based on external validation results

• Performance Benchmarking:

  • Calculate identical performance metrics as used during model development
  • Compare external performance against internal cross-validation results
  • Establish performance degradation thresholds for acceptable deployment

• Error Analysis:

  • Identify specific failure modes and performance variations across subsets
  • Analyze whether performance differences correlate with dataset characteristics (e.g., sample matrix, instrument type)

• Iterative Refinement:

  • If performance is unsatisfactory, expand training data diversity or adjust model architecture
  • Maintain a completely separate external test set for final validation after refinements

Quantitative Assessment of Generalizability

Table 3: Framework for Quantifying Generalizability Performance Gaps

Performance Metric	Internal Validation Performance (Mean ± SD)	External Validation Performance	Performance Gap	Acceptance Threshold
Classification Accuracy	94.2% ± 2.1%	87.5%	-6.7%	≤10% decrease
F1-Score	0.92 ± 0.03	0.84	-0.08	≤0.10 decrease
AUC-ROC	0.96 ± 0.02	0.89	-0.07	≤0.08 decrease
Mean Squared Error	0.15 ± 0.04	0.23	+0.08	≤50% increase
R² Score	0.88 ± 0.05	0.79	-0.09	≤0.12 decrease

Advanced Topics in Generalizability Assessment

The SPECTRA Framework for Comprehensive Evaluation

Recent advances in generalizability assessment include the SPECTRA (Spectral Framework for Model Evaluation) approach, which systematically evaluates model performance as a function of decreasing similarity between training and test data [66]. Rather than relying on single performance estimates, SPECTRA plots model performance against cross-split overlap (similarity between train and test splits) and calculates the area under this curve as a comprehensive generalizability metric [66].

In evaluations of 19 state-of-the-art deep learning models across 18 molecular sequencing datasets, SPECTRA revealed that traditional sequence similarity- and metadata-based splits provide incomplete assessments of model generalizability [66]. The framework demonstrated that as cross-split overlap decreases, even sophisticated models consistently show reduced performance, though the degree of degradation varies substantially by task and model architecture.

Addressing Data Leakage in Cross-Validation

Data leakage during cross-validation represents a significant threat to reliable generalizability assessment [62] [63]. Leakage occurs when information from the validation set inadvertently influences the training process, creating overoptimistic performance estimates [63]. Common sources in ML-assisted NTA include:

• Normalization Leakage: Applying normalization across the entire dataset before splitting, allowing validation distribution information to influence training [62]
• Feature Selection Leakage: Performing feature selection using the complete dataset before cross-validation [63]
• Augmentation Leakage: Applying data augmentation techniques before data splitting [63]

To prevent data leakage, all data preprocessing steps—including normalization, imputation, and feature selection—should be performed independently within each cross-validation fold using only training data [63].

Cross-Validation Data Flow: Preventing Leakage

Table 4: Research Reagent Solutions for Generalizability Assessment

Resource Category	Specific Tools/Approaches	Function in Generalizability Assessment
Cross-Validation Implementations	scikit-learn (KFold, StratifiedKFold) [65], MLJ (Julia), CARET (R)	Standardized implementation of resampling methods
Performance Metrics Libraries	scikit-learn metrics, TorchMetrics, TensorFlow Model Analysis	Comprehensive calculation of validation metrics
Data Preprocessing Tools	SCONE (R), PyMS (Python), XCMS [7]	Reproducible data preprocessing pipelines
Molecular Feature Alignment	XCMS retention time correction [7], MZmine 3	Cross-batch data alignment for external validation
Benchmark Datasets	MoleculeNet [66], Proteinglue [66], FLIP [66]	Standardized external validation resources
Generalizability Frameworks	SPECTRA [66], WILDS [66]	Specialized tools for generalizability quantification

Tier 2 validation through cross-validation and external dataset assessment provides the methodological foundation for establishing model generalizability in ML-assisted non-target analysis. By implementing rigorous resampling techniques and testing models against truly independent datasets, researchers can distinguish between models that merely memorize training data and those that learn transferable patterns applicable to new samples and conditions. The protocols and frameworks presented here enable quantitative assessment of generalizability, identification of performance boundaries, and documentation of model limitations—all essential for responsible deployment of ML models in pharmaceutical development, environmental monitoring, and clinical applications.

Within a comprehensive tiered validation strategy for Machine Learning-assisted Non-Target Analysis (ML-NTA), Tiers 1 (analytical confidence) and 2 (model generalizability) establish the foundational reliability of chemical data and predictive models. Tier 3 validation is the critical final step that contextualizes these findings within the real world, assessing their environmental and biological plausibility [7]. This tier answers the crucial question: Do the model's predictions and the identified chemical patterns make sense given the known context of the contamination source and the biological or environmental systems affected? The absence of such plausibility assessments can render analytically sound results environmentally meaningless, leading to flawed environmental decision-making [7] [70]. This document provides detailed application notes and protocols for implementing robust Tier 3 plausibility checks.

Theoretical Framework and Key Concepts

Defining Plausibility in ML-NTA

For ML-NTA, plausibility is the degree to which the model's outputs—such as identified contamination sources, spatial gradients, or temporal trends—align with pre-established, evidence-based expectations of environmental or biological systems [70].

• Biological Plausibility: Concerns the consistency of findings with known disease processes and the mechanisms of action of identified contaminants. It asks whether a proposed biological impact is credible given current knowledge [70].
• Clinical Plausibility: Focuses on human interaction with biological processes, often relating to patient outcomes, treatment pathways, and population health effects in public health contexts [70].
• Environmental Plausibility: Concerns the consistency of chemical patterns with known source characteristics, environmental fate and transport processes, and geospatial relationships between sources and sampling points [7].

Operationally, biologically and clinically plausible extrapolations (or predictions) are defined as "predicted survival estimates that fall within the range considered plausible a-priori, obtained using a-priori justified methodology" [70]. This definition is directly transferable to ML-NTA, where "survival estimates" can be replaced with "source contributions," "contamination gradients," or "risk assessments."

The Role of Tier 3 in a Tiered Validation Strategy

A tiered validation strategy ensures that ML-NTA findings are not just statistically sound but also environmentally actionable [7]. The following workflow illustrates how Tier 3 integrates with and completes the overall validation process.

Experimental Protocols for Plausibility Assessment

This section outlines a standardized, five-step protocol for assessing the environmental and biological plausibility of ML-NTA findings, adapting the DICSA framework used in health technology assessment for environmental science [70].

The DICSA Protocol for Plausibility Assessment

Objective: To prospectively define and quantitatively assess the biological and clinical plausibility of model outputs.

Step 1: Define the Target Setting

• Action: Describe the system of interest in terms of all relevant aspects that influence chemical distribution and impact.
• Documentation:
  • Source Characteristics: Industrial processes, product formulations, or agricultural practices.
  • Pathway Factors: Prevailing wind/water currents, soil composition, hydrogeology.
  • Receptor Factors: Affected ecosystem type, sensitive species, or human population demographics.
  • Temporal Scope: Timeframe of chemical release and sampling.

Step 2: Collect Information from Relevant Sources

• Action: Systematically gather existing evidence to inform expectations about chemical patterns and concentrations.
• Data Sources:
  • Scientific literature on source-specific chemical markers (e.g., PFAS from fire-fighting foams) [7].
  • Historical monitoring data from the study area.
  • Regulatory databases of industrial emissions.
  • Input from domain experts (e.g., hydrologists, toxicologists, site managers).

Step 3: Compare Survival-Influencing Aspects Across Sources

• Action: Critically evaluate and harmonize the collected information to account for differences that may affect cross-study comparisons.
• Evaluation Criteria:
  • Similarity in environmental settings and sample matrices.
  • Consistency in analytical methods.
  • Comparability of population or ecosystem characteristics.

Step 4: Set A-Priori Plausibility Expectations

• Action: Before final model development, define quantitative or qualitative ranges for what constitutes a plausible result.
• Outputs:
  • Expected Chemical Fingerprints: A priori list of suspected source-specific indicator compounds and their relative abundances [7] [9].
  • Plausible Concentration Ranges: Minimum, maximum, and expected concentration levels for key contaminants.
  • Spatial/Temporal Gradients: Expected patterns of chemical distribution from putative sources.

Step 5: Assess Model Alignment with Expectations

• Action: Compare the final ML-NTA model outputs and predictions against the pre-defined plausibility ranges from Step 4.
• Assessment:
  • Do identified marker compounds match a priori expectations?
  • Do predicted source contributions and spatial patterns align with known source locations and environmental pathways?
  • Is the magnitude of the predicted effect (e.g., risk) within a biologically plausible range?

Protocol for Correlative Analysis with Contextual Data

Objective: To statistically strengthen plausibility assessments by correlating ML-NTA outputs with independent, contextual datasets.

Methodology:

• Data Collection: Gather geospatial and contextual data for each sample point.
• Variable Compilation: Assemble these variables into a structured dataset for correlation analysis.
• Statistical Testing: Perform correlation analysis (e.g., Spearman's rank) between the ML model's predictions and the contextual variables.

Table 1: Key Contextual Data Variables for Correlative Plausibility Analysis

Data Category	Specific Variable	Measurement Method	Plausible Correlation with NTA Findings
Geospatial Data	Distance to Potential Source (e.g., factory, WWTP)	GIS Mapping	Negative correlation with contaminant levels [7]
	Land Use Type (e.g., industrial, agricultural)	Land Use Classification	Specific chemical profiles associated with each land use type
Source Inventory	Known Industrial Emissions Inventory	Regulatory Filings	Positive correlation with specific industrial compounds
	Agricultural Pesticide Usage Reports	Government Surveys	Positive correlation with pesticide and metabolite levels
Hydrogeological Data	Upstream/Downstream Position	Hydrological Modeling	Gradient consistent with water flow direction
	Groundwater Flow Models	Hydrogeological Survey	Contaminant plume aligns with predicted flow path

The Scientist's Toolkit: Essential Reagents and Materials

Successful implementation of Tier 3 validation requires specific reagents and materials for data collection, analysis, and interpretation.

Table 2: Key Research Reagent Solutions and Materials for Tier 3 Validation

Item Name	Function / Purpose	Example Specification / Notes
Certified Reference Materials (CRMs)	To verify compound identities and provide analytical confidence for key markers [7].	PFAS mix, Pesticide mix; source-specific marker compounds.
Internal Standards (Isotope-Labeled)	For quality control during sample analysis and quantification.	( ^{13}C ), ( ^{15}N ), or ( ^{2}H )-labeled analogs of target compounds.
Multi-Sorbent SPE Cartridges	Broad-spectrum extraction of contaminants with diverse physicochemical properties [7].	Oasis HLB, ISOLUTE ENV+, Strata WAX/WCX combinations.
GIS Software License	To manage, analyze, and visualize geospatial contextual data.	ArcGIS, QGIS.
Statistical Analysis Software	To perform correlation analyses between NTA results and contextual data.	R, Python (with pandas, scikit-learn).
Chemical Database Access	To research source-specific chemical markers and properties.	PubChem, NORMAN Suspect List, STOFF-IDENT.

Workflow Visualization: Integrating Tier 3 into the ML-NTA Process

The following diagram details the complete ML-NTA workflow, highlighting the integration of Tier 3 plausibility checks and the critical data inputs required at each stage.

Case Study Application: PFAS Source Identification

Scenario: An ML classifier (e.g., Random Forest) has been trained on HRMS data from 92 water samples to classify sources of Per- and Polyfluoroalkyl Substances (PFAS) [7].

Application of DICSA Protocol:

• Define: Target setting is a watershed with a known fire-training area (potential source of long-chain PFCAs) and an industrial park (potential source of fluorotelomers).
• Collect: Literature confirms PFOS and PFOA as markers for aqueous film-forming foams (AFFFs). Industry reports suggest the use of fluorotelomer-based polymers.
• Compare: Historical groundwater data from the site shows a plume of PFOS/PFOA originating from the fire-training area.
• Set Expectations: A priori, samples near the fire-training area are expected to show high levels of PFOS/PFOA, while samples down-gradient from the industrial park should be dominated by fluorotelomer transformation products.
• Assess: The ML model successfully classifies samples into two distinct clusters. The chemical fingerprints of each cluster are examined. Cluster 1, associated with the fire-training area, is indeed dominated by PFOS and PFOA. Cluster 2 shows a high prevalence of fluorotelomer sulfonates, aligning with the industrial source. Geospatial analysis confirms that Cluster 2 samples are hydrologically down-gradient from the industrial park, strengthening the environmental plausibility of the model's classification.

Within a tiered validation strategy for Machine Learning (ML)-assisted non-target analysis research, a central challenge is selecting models that deliver both high predictive power and understandable decision logic. The prevailing assumption of a strict trade-off between accuracy and interpretability often forces researchers to choose between performance and transparency. However, recent empirical evidence challenges this notion, demonstrating that modern interpretable models can achieve competitive accuracy while providing the transparency essential for high-stakes domains like drug development [71]. This document outlines application notes and experimental protocols for comparing ML models, enabling the selection of models that align with the distinct validation tiers of a research strategy, from initial screening to confirmatory analysis.

The Accuracy-Interpretability Landscape in Machine Learning

The core challenge in model selection lies in balancing two competing objectives. Interpretability refers to a model's ability to explain or present its decision logic in a human-understandable way [71]. This is distinct from post-hoc explainability, which uses secondary models to approximate the behavior of a complex "black-box" model. In contrast, intrinsically interpretable models are transparent by design, providing an exact description of how a prediction is computed [71]. From a functionally-grounded perspective, this often involves structural constraints like linearity, additivity, or sparsity [71].

In critical applications such as biomedical time series analysis or drug effect estimation, understanding the model's rationale is as vital as its predictive power [72] [73]. While deep learning models often achieve top accuracy in tasks like EEG or ECG classification, their opacity is a significant drawback [72]. Conversely, simpler models like decision trees or K-nearest neighbors are fully interpretable but may lack the required predictive performance [72]. This complex relationship is not strictly monotonic; there are instances where more interpretable models can match or even surpass the performance of black-box alternatives on specific tasks, particularly with structured, tabular data [74] [71].

Model Performance Across Domains

The following table summarizes the predictive performance of various model types across different application domains, as reported in the literature.

Table 1: Comparative Model Performance Across Different Domains

Domain	Task	Best-Performing Model(s)	Performance Metric & Score	Interpretability of Top Model
General Tabular Data [71]	Classification/Regression	Generalized Additive Models (GAMs), Tree-Based Models	Comparable accuracy to black-box models on many datasets	High (Intrinsically Interpretable)
Biomedical Time Series (e.g., ECG, EEG) [72]	Classification (e.g., heart disease, epilepsy)	Convolutional Neural Networks (CNN) with RNN or Attention layers	Highest accuracy	Low (Black-Box)
Power Demand Prediction [75]	Load Forecasting	Deep Learning (RNN, GRU, LSTM) & Tree-Based (XGBoost, LightGBM)	Lower power scenarios: Tree-based CV-RMSE 13.62%, DL 12.17%	Tree-Based: High, DL: Low
NLP: Rating Inference [74]	Sentiment Analysis/Prediction	Neural Networks (NN), BERT	Highest accuracy (exact scores dataset-dependent)	Low (Black-Box)

A Quantitative Interpretability Score

To move beyond a simple binary classification, one study proposed a Composite Interpretability (CI) Score, which quantifies interpretability based on expert assessments of simplicity, transparency, explainability, and model complexity [74]. The scores for various NLP models are detailed below.

Table 2: Composite Interpretability Scores for a Selection of Models [74]

Model	Simplicity	Transparency	Explainability	Number of Parameters	CI Score
VADER (Rule-Based)	1.45	1.60	1.55	0	0.20
Logistic Regression (LR)	1.55	1.70	1.55	3	0.22
Naive Bayes (NB)	2.30	2.55	2.60	15	0.35
Support Vector Machine (SVM)	3.10	3.15	3.25	20,131	0.45
Neural Network (NN)	4.00	4.00	4.20	67,845	0.57
BERT	4.60	4.40	4.50	183.7M	1.00

Note: A lower CI Score indicates higher interpretability. Simplicity, Transparency, and Explainability are expert rankings on a 1-5 scale (lower is more interpretable).

Detailed Experimental Protocols for Model Comparison

To ensure reproducible and fair comparisons in ML-assisted non-target analysis, the following protocols are recommended.

Protocol 1: Benchmarking Model Performance and Interpretability

Objective: To systematically evaluate and compare the predictive accuracy, interpretability, and robustness of candidate ML models on a specific dataset.

Materials:

• Datasets: Curated, pre-processed, and representative datasets for the non-target analysis task (e.g., mass spectral data). Data should be split into training, validation, and test sets.
• Software: Python environments with scikit-learn, XGBoost, PyTorch/TensorFlow, and interpretability libraries (e.g., SHAP, interpretML).
• Hardware: Computational resources with adequate CPU/GPU support for training complex models.

Methodology:

• Model Selection: Choose a diverse set of models spanning the interpretability spectrum.
  • High Interpretability: Logistic Regression, Decision Trees, Generalized Additive Models (GAMs) [71].
  • Medium Interpretability: Random Forests, Gradient Boosted Trees (XGBoost, LightGBM) [75].
  • Low Interpretability: Deep Neural Networks (Multilayer Perceptrons, CNNs, RNNs), pre-trained transformers [72].
• Hyperparameter Tuning: Conduct an extensive hyperparameter search for each model using a defined strategy (e.g., grid search, random search, Bayesian optimization) with cross-validation on the training set. This is critical for a fair comparison [71].
• Model Training: Train each model with its optimal hyperparameters on the full training set.
• Performance Evaluation: Calculate standard performance metrics (e.g., Accuracy, Precision, Recall, F1-Score, AUC-ROC, RMSE) on the held-out test set.
• Interpretability Assessment:
  • For Intrinsically Interpretable Models: Directly visualize and analyze the model's internal mechanics (e.g., coefficient plots for GAMs [71], decision rules for trees).
  • For Black-Box Models: Apply post-hoc explanation tools like SHAP or LIME to generate feature importance scores [71].
• Robustness Testing: Evaluate model stability in the presence of noise and data shifts, a key component of a comprehensive benchmarking framework [76].

Outputs:

• A table of performance metrics for all models (see Table 1).
• Interpretability visualizations (e.g., shape functions for GAMs, SHAP summary plots).
• A qualitative assessment of the clarity and utility of the explanations for a domain expert.

Protocol 2: Causal Machine Learning with Real-World Data (RWD)

Objective: To leverage RWD and Causal ML (CML) to estimate treatment effects, identify responsive patient subgroups, and generate hypotheses for indication expansion.

Materials:

• Real-World Data (RWD): Electronic Health Records (EHRs), insurance claims, patient registries, or data from wearable devices [73].
• Clinical Trial Data: Data from Randomized Controlled Trials (RCTs) for validation.
• Software: Causal inference libraries (e.g., EconML, DoWhy, CausalML).

Methodology:

• Define Causal Question: Formulate a clear question (e.g., "What is the causal effect of Drug X on outcome Y?").
• Data Preprocessing & Harmonization: Clean RWD and address missing values. Harmonize variable definitions between RWD and RCT data.
• Address Confounding: Select and apply appropriate CML methods to mitigate confounding bias in observational RWD.
  • Propensity Score Methods: Estimate propensity scores using ML models (e.g., boosting, random forests) and apply inverse probability weighting or matching [73].
  • Doubly Robust Methods: Use methods like Targeted Maximum Likelihood Estimation (TMLE) that combine outcome and propensity models for more robust estimates [73].
• Estimate Heterogeneous Treatment Effects: Use models like Causal Forests to identify and characterize patient subgroups with varying responses to the treatment [73].
• Validation: Where possible, compare CML estimates derived from RWD with results from gold-standard RCTs to assess validity [73].

Outputs:

• Estimates of average treatment effect from RWD.
• Characterization of patient subgroups that show enhanced or diminished response.
• Assessment of transportability of treatment effects from trial populations to broader real-world populations.

Workflow and Pathway Visualizations

Tiered Model Selection Workflow

The following diagram outlines a decision workflow for selecting models within a tiered validation strategy, based on project requirements for accuracy and interpretability.

Tiered model selection workflow.

Causal ML Analysis Pathway

This diagram illustrates the key stages in applying Causal Machine Learning to real-world data for drug development.

Causal ML analysis pathway.

For researchers implementing the aforementioned protocols, the following tools and benchmarks are essential.

Table 3: Essential Research Reagents and Resources

Resource Name	Type	Primary Function in Research	Relevance to Tiered Validation
Generalized Additive Models (GAMs) [71]	Interpretable Model Class	Models non-linear relationships with full transparency via additive shape functions.	Ideal for tiers requiring high interpretability without significant accuracy loss.
Causal Forest [73]	Causal ML Algorithm	Estimates heterogeneous treatment effects from observational data, identifying patient subgroups.	Crucial for analyzing RWD to generate hypotheses for new indications or subgroups.
SHAP (SHapley Additive exPlanations) [71]	Post-hoc Explanation Tool	Provides unified, consistent feature importance values for any model's predictions.	Useful for explaining black-box models in later validation tiers, though not a substitute for intrinsic interpretability.
MIB Benchmark [77]	Interpretability Benchmark	Evaluates mechanistic interpretability methods on their ability to recover true causal circuits in models.	Provides a standard for evaluating explanation methods themselves, supporting method selection.
Doubly Robust Estimators (e.g., TMLE) [73]	Causal Inference Method	Combines propensity score and outcome models for robust causal effect estimation even if one model is misspecified.	Enhances the reliability of causal conclusions drawn from RWD in non-randomized settings.
Composite Interpretability (CI) Score [74]	Quantitative Metric	Quantifies the interpretability of a model based on expert assessments and model complexity.	Aids in the objective ranking and selection of models based on their transparency.

The detection and identification of unknown chemicals in environmental, biological, and product-based samples through non-targeted analysis (NTA) has traditionally focused on qualitative characterization. However, the growing need to understand contaminant concentrations for risk assessment has driven the development of quantitative non-targeted analysis (qNTA). While traditional NTA answers the question "What is present?", qNTA addresses the critical follow-up question: "How much is there?" [78]. This transition enables practitioners to generate chemical concentration estimates that directly inform provisional risk-based decisions and prioritize targets for follow-up confirmation analysis [78]. The integration of machine learning (ML) frameworks further enhances this quantitative paradigm by transforming complex high-resolution mass spectrometry (HRMS) data into environmentally actionable parameters for contaminant source identification and risk assessment [7] [9].

The fundamental distinction between qualitative and quantitative data frameworks underpins this methodological evolution. Qualitative data describes characteristics, types, or categories through names or labels—think quality or attribute. In NTA, this manifests as chemical identification, classification by use-type, or categorization by source pattern. In contrast, quantitative data involves measurements or counts recorded using numbers—think quantity. In the qNTA context, this translates to concentration values, peak intensities, fold-changes, and statistical probability scores [79]. Effective risk assessment requires both data types: qualitative identification provides the "what," while quantitative measurement provides the "how much," together creating a complete picture for decision-making [79].

Core Principles of Quantitative Non-Targeted Analysis

Surrogate-Based Calibration Approaches

Most qNTA and "semi-quantitative" approaches rely on surrogate chemicals for calibration and model predictions. The selection of appropriate surrogates is therefore critical for analytical accuracy. Traditionally, surrogates have been chosen based on intuition and availability rather than rational, structure-based selection. This practice limits the objective assessment and improvement of qNTA methods [78]. Structure-based surrogate selection strategies systematically leverage chemical space information to improve quantitative accuracy. Key molecular descriptors relevant to electrospray ionization efficiency can be used to embed chemicals in a defined space where leverage calculations identify optimal surrogates [78].

The Leveraged Averaged Representative Distance (LARD) metric has been proposed to quantify surrogate coverage within a defined chemical space, providing a rational framework for surrogate selection [78]. Research indicates that while qNTA models benefit significantly from rational surrogate selection strategies, a sufficiently large random surrogate sample can perform as well as a smaller, chemically informed surrogate sample. This finding provides practical guidance for researchers designing qNTA studies with limited prior knowledge of the chemical space [78].

Machine Learning-Enhanced QNTA Frameworks

Machine learning redefines qNTA potential by identifying latent patterns within high-dimensional data that traditional statistical methods often miss. ML classifiers such as Support Vector Classifier (SVC), Logistic Regression (LR), and Random Forest (RF) have demonstrated remarkable effectiveness in source attribution tasks. In one application, these classifiers screened 222 targeted and suspect per- and polyfluoroalkyl substances (PFASs) across 92 samples, achieving balanced classification accuracy ranging from 85.5% to 99.5% across different sources [7]. Similarly, Partial Least Squares Discriminant Analysis (PLS-DA) has proven effective in identifying source-specific indicator compounds through variable importance metrics [7].

The integration of ML with qNTA follows a systematic four-stage workflow: (i) sample treatment and extraction, (ii) data generation and acquisition, (iii) ML-oriented data processing and analysis, and (iv) result validation [7]. This framework ensures that raw HRMS data is transformed through sequential computational steps into interpretable patterns and quantifiable concentrations suitable for risk-based decision making.

Application Notes: Implementing ML-Assisted QNTA

Research Reagent Solutions and Essential Materials

Successful implementation of qNTA requires carefully selected reagents and materials optimized for broad-spectrum chemical analysis. The following table details key research solutions and their functions within the ML-assisted qNTA workflow:

Table 1: Essential Research Reagents and Materials for ML-Assisted QNTA

Item Name	Function/Application	Key Considerations
Multi-sorbent SPE (e.g., Oasis HLB, ISOLUTE ENV+, Strata WAX/WCX)	Broad-range compound extraction from environmental matrices	Provides complementary selectivity; enhances coverage of diverse physicochemical properties [7]
Green Extraction Solvents (QuEChERS, MAE, SFE)	Efficient analyte recovery while minimizing matrix interference	Reduces solvent usage and processing time; particularly valuable for large-scale environmental samples [7]
HRMS Instrumentation (Q-TOF, Orbitrap)	High-resolution mass spectral data acquisition	Enables precise mass measurement; resolves isotopic patterns and fragmentation signatures [7] [9]
Chromatographic Systems (LC/GC)	Compound separation prior to mass spectrometry	Reduces matrix effects; complements HRMS detection [7]
Certified Reference Materials (CRMs)	Analytical confidence verification and compound identity confirmation	Essential for method validation and quality assurance [7]
Quality Control (QC) Samples	Batch-specific quality assurance throughout analysis	Monitors instrument performance; ensures data integrity across samples [7]

Data Processing and Statistical Analysis Framework

The transition from raw HRMS data to quantitative concentrations involves sequential computational steps with distinct statistical approaches for qualitative versus quantitative data types. The following table compares the analytical approaches for these two data paradigms:

Table 2: Analytical Approaches for Qualitative vs. Quantitative Data in NTA

Analytical Aspect	Qualitative Data Approach	Quantitative Data Approach
Primary Goal	Chemical identification and classification	Concentration estimation and quantification
Data Characteristics	Descriptions, types, and names; mutually exclusive categories	Numerical measurements; continuous or discrete values
Common Statistical Analyses	Chi-square tests, Proportion tests, Frequency analysis	T-tests, ANOVA, Correlation analysis, Regression
Visualization Methods	Bar charts, Pie charts	Histograms, Scatterplots, Frequency polygons
Key Outputs	Chemical identities, Classification categories	Concentration estimates, Uncertainty measures, Dose-response relationships

For quantitative data analysis, initial preprocessing addresses data quality through noise filtering, missing value imputation (e.g., k-nearest neighbors), and normalization (e.g., Total Ion Current normalization) to mitigate batch effects. Exploratory analysis then 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) simplify high-dimensional data, while clustering methods (hierarchical cluster analysis, k-means) group samples by chemical similarity [7].

Experimental Protocols: Tiered Validation Strategy for ML-Assisted QNTA

Comprehensive Workflow for QNTA Implementation

Diagram 1: Comprehensive ML-Assisted qNTA Workflow

Protocol: Sample Treatment and Extraction

Objective: Maximize compound recovery while minimizing matrix interference to ensure comprehensive contaminant detection.

Materials:

• Solid Phase Extraction (SPE) apparatus with multi-sorbent cartridges (e.g., Oasis HLB, ISOLUTE ENV+, Strata WAX, Strata WCX)
• Pressurized Liquid Extraction (PLE) system or Microwave-Assisted Extraction (MAE) equipment
• QuEChERS extraction kits for rapid sample preparation
• Appropriate solvents (methanol, acetonitrile, water) with optimised purity grades

Procedure:

• Sample Preparation: Homogenize environmental samples (water, soil, sediment, biota) using appropriate methods. For water samples, filter through 0.45μm glass fiber filters to remove particulate matter.
• Multi-sorbent SPE Extraction: a. Condition SPE cartridges with 5mL methanol followed by 5mL reagent water. b. Load samples at controlled flow rates (3-5mL/min). c. Dry cartridges under vacuum for 20-30 minutes. d. Elute with sequential solvent gradients (e.g., 5mL methanol followed by 5mL dichloromethane). e. Concentrate eluents under gentle nitrogen stream to near dryness. f. Reconstitute in initial mobile phase composition for analysis.
• Alternative Extraction Methods: For solid samples, employ PLE (100°C, 1500psi, 5min static time) or MAE (100°C, 20min) with methanol:water mixtures.
• Purification: Apply additional cleanup using gel permeation chromatography (GPC) for lipid-rich samples.

Quality Control: Include procedural blanks, matrix spikes, and internal standards throughout the extraction process to monitor contamination and recovery efficiencies.

Protocol: Data Generation and Acquisition via HRMS

Objective: Generate high-quality spectral data with sufficient resolution for compound identification and quantification.

Materials:

• High-Resolution Mass Spectrometer (Q-TOF or Orbitrap system)
• Liquid or Gas Chromatography system for compound separation
• Data acquisition and processing software
• Quality control standards and calibration solutions

Procedure:

• Chromatographic Separation: a. Implement reverse-phase LC separation using C18 columns (100 × 2.1mm, 1.7-1.8μm) b. Utilize binary mobile phase gradient: (A) water with 0.1% formic acid and (B) methanol or acetonitrile with 0.1% formic acid c. Apply linear gradient from 5% B to 95% B over 15-30 minutes d. Maintain column temperature at 40°C with flow rate of 0.3mL/min
• Mass Spectrometric Detection: a. Operate in both positive and negative electrospray ionization modes b. Set source temperature to 150°C with desolvation temperature at 350°C c. Apply capillary voltage of 1.0kV and cone voltage of 20-40V d. Use data-independent acquisition (DIA) with alternating low and high collision energies e. Set mass resolution to >25,000 FWHM for Orbitrap or >30,000 FWHM for Q-TOF f. Implement mass range of m/z 50-1200 with scan time of 0.1-0.3s
• Data Processing: a. Perform centroiding of raw spectral data b. Conduct peak detection, alignment, and componentization c. Generate feature-intensity matrix with samples as rows and chemical features as columns

Quality Control: Inject quality control samples (pooled quality control samples) regularly throughout sequence to monitor instrument stability. Include internal standards for retention time alignment and mass accuracy calibration.

Protocol: ML-Oriented Data Processing and Analysis

Objective: Transform raw HRMS data into interpretable patterns and quantifiable concentrations through machine learning approaches.

Materials:

• Computational environment (R, Python)
• ML libraries (scikit-learn, XGBoost, TensorFlow)
• Statistical analysis software
• High-performance computing resources for large datasets

Procedure:

• Data Preprocessing: a. Apply retention time correction and m/z recalibration for data alignment b. Perform missing value imputation using k-nearest neighbors (k=5) c. Normalize data using Total Ion Current (TIC) or quantile normalization d. Conduct noise filtering with relative standard deviation thresholds (≥30% in QC samples)
• Dimensionality Reduction: a. Implement Principal Component Analysis (PCA) to identify major sources of variance b. Apply t-distributed Stochastic Neighbor Embedding (t-SNE) for nonlinear pattern recognition c. Use Uniform Manifold Approximation and Projection (UMAP) for enhanced visualization
• Clustering Analysis: a. Perform hierarchical cluster analysis (HCA) with Ward's linkage and Euclidean distance b. Conduct k-means clustering to identify sample groupings c. Validate cluster stability with silhouette analysis
• Supervised Classification: a. Split dataset into training (70%) and testing (30%) sets b. Train multiple classifiers (Random Forest, Support Vector Machine, Logistic Regression) c. Optimize hyperparameters through grid search with cross-validation d. Evaluate model performance using accuracy, precision, recall, and F1-score
• Quantitative Model Development: a. Select surrogates using structure-based chemical space embedding b. Develop predictive models for concentration estimation c. Calculate Leveraged Averaged Representative Distance (LARD) to assess surrogate coverage

Validation: Implement k-fold cross-validation (k=5 or 10) to assess model robustness and prevent overfitting.

Protocol: Tiered Validation Strategy

Objective: Ensure reliability and environmental relevance of ML-assisted qNTA outputs through multi-faceted validation.

Diagram 2: Tiered Validation Strategy Framework

Materials:

• Certified Reference Materials (CRMs)
• External validation datasets
• Geographical Information System (GIS) software
• Statistical analysis tools

Procedure:

• Tier 1: Analytical Confidence Verification a. Confirm compound identities using Certified Reference Materials (CRMs) b. Match spectra with authentic standards in spectral libraries c. Assign confidence levels (Level 1-5) following Schymanski et al. framework d. Verify quantitative accuracy through spike-recovery experiments

• Tier 2: Model Generalizability Assessment a. Validate classifiers on independent external datasets b. Perform 10-fold cross-validation to evaluate overfitting risks c. Assess model performance across different sample matrices d. Calculate variable importance metrics for model interpretability

• Tier 3: Environmental Plausibility Check a. Correlate model predictions with geospatial proximity to emission sources b. Verify presence of known source-specific chemical markers c. Compare temporal trends with known usage patterns d. Assess consistency with complementary environmental data

Documentation: Maintain comprehensive records of all validation procedures, including acceptance criteria, performance metrics, and any deviations from protocols.

The structured integration of quantitative non-targeted analysis with machine learning frameworks represents a paradigm shift in environmental analytical chemistry. The tiered validation strategy ensures that concentration estimates derived from qNTA are both analytically rigorous and environmentally relevant, providing a solid foundation for risk-based decision making. By implementing the detailed protocols outlined in this document—from optimized sample preparation through ML-oriented data processing to comprehensive validation—researchers can reliably translate complex HRMS data into actionable environmental insights. As these methodologies continue to evolve, they will increasingly support regulatory applications and public health protection through more accurate characterization of contaminant exposure and potential risk.

Conclusion

The successful implementation of a tiered validation strategy is paramount for transforming ML-assisted Non-Target Analysis from an exploratory tool into a reliable source for critical decision-making in biomedical and environmental science. This synthesis of core intents demonstrates that foundational knowledge, a meticulous methodological workflow, proactive troubleshooting, and, most importantly, a multi-faceted validation approach are inseparable components. By adhering to this framework, researchers can effectively mitigate the risks associated with 'black-box' models and complex datasets, thereby generating findings that are not only computationally sound but also chemically accurate and contextually relevant. Future directions will involve the deeper integration of these validated ML-NTA workflows into regulatory frameworks, the advancement of fully quantitative NTA for robust risk assessment, and the development of more sophisticated, inherently interpretable AI models. This progression will undoubtedly accelerate drug discovery, enhance environmental monitoring, and ultimately strengthen the bridge between high-throughput analytical science and tangible public health outcomes.

References

• 1. Advances and challenges in non-targeted analysis [https://www.sciencedirect.com/science/article/abs/pii/S0021967324008331]
• 2. High-resolution mass spectrometry: more than exact mass [https://www.bioanalysis-zone.com/high-resolution-mass-spectrometry-bioanalysis/]
• 3. NORMAN guidance on suspect and non-target screening in ... [https://enveurope.springeropen.com/articles/10.1186/s12302-023-00779-4]
• 4. Quantitative non-targeted analysis: Bridging the gap ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8979303/]
• 5. Non-targeted analysis (NTA) and suspect screening ... [https://www.nature.com/articles/s41370-023-00574-6]
• 6. Working Principle of HRMS [https://resolvemass.ca/working-principle-of-hrms/]
• 7. Integrating non-target analysis and machine learning [https://www.nature.com/articles/s41545-025-00504-z]
• 8. Machine Learning-based Classification for the Prioritization ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11924234/]
• 9. Advancing non-target analysis of emerging environmental ... [https://www.sciencedirect.com/science/article/pii/S0160412025001552]
• 10. Top 5 Validation Methods In Machine Learning [https://medium.com/@mehmetalitor/top-5-model-validation-methods-in-machine-learning-77eed8d08937]
• 11. An information bottleneck approach for feature selection [https://www.sciencedirect.com/science/article/abs/pii/S0031320325002249]
• 12. Performance Analysis and Bottleneck Identification in AI ... [https://multicorewareinc.com/performance-analysis-and-bottleneck-identification-in-ai-workflows/]
• 13. Datasets of Visualization for Machine Learning [https://arxiv.org/html/2407.16351v1]
• 14. Data Quality Issues Affecting the Pharmaceutical Industry [https://firsteigen.com/blog/data-quality-issues-affecting-the-pharmaceutical-industry-finding-a-solution/]
• 15. How To Build Reliable AI Models That Don't Fail [https://www.montecarlodata.com/blog-ai-reliability/]
• 16. Validation of Machine Learning Models: Challenges and ... [https://www.protiviti.com/gl-en/blogs/validation-machine-learning-models-challenges-and-alternatives]
• 17. An automated toxicity based prioritization framework for ... [https://www.sciencedirect.com/science/article/abs/pii/S0304389423001759]
• 18. A Step-by-Step Guide to Effective AI Model Validation [https://galileo.ai/blog/ai-model-validation]
• 19. AI Model Validation: Best Practices for Accuracy & Reliability [https://galileo.ai/blog/best-practices-for-ai-model-validation-in-machine-learning]
• 20. AI Model Testing: How to Build Trust in Intelligent Systems [https://testgrid.io/blog/ai-model-testing/]
• 21. Techniques for extraction and isolation of natural products [https://cmjournal.biomedcentral.com/articles/10.1186/s13020-018-0177-x]
• 22. Quantification Approaches in Non-Target LC/ESI/HRMS ... [https://www.nist.gov/publications/quantification-approaches-non-target-lcesihrms-analysis-interlaboratory-comparison]
• 23. Scalable Analysis of Untargeted LC-HRMS Data by Means ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10018448/]
• 24. Data Preprocessing in Machine Learning: Steps & Best ... [https://lakefs.io/blog/data-preprocessing-in-machine-learning/]
• 25. A Beginner's Guide to Data Preprocessing in Machine ... [https://www.dataentryoutsourced.com/blog/data-preprocessing-for-machine-learning-clean-data-smarter-models/]
• 26. Dimensionality Reduction in Machine Learning for ... [https://www.guvi.in/blog/dimensionality-reduction-in-machine-learning/]
• 27. Introduction to Dimensionality Reduction [https://www.geeksforgeeks.org/machine-learning/dimensionality-reduction/]
• 28. Clustering in Machine Learning [https://www.geeksforgeeks.org/machine-learning/clustering-in-machine-learning/]
• 29. Clustering algorithms | Machine Learning [https://developers.google.com/machine-learning/clustering/clustering-algorithms]
• 30. 三个有效的特征选择策略 [https://bigquant.com/wiki/doc/fKA4SNFjlo]
• 31. 7 Strategies to Choose the Best Features for Your Product [https://www.productplan.com/learn/strategies-prioritize-product-features/]
• 32. Tabular特征选择基准原创 [https://blog.csdn.net/qq_40943760/article/details/134477728]
• 33. What is feature prioritization? Five methods and examples [https://www.optimizely.com/optimization-glossary/feature-prioritization/]
• 34. 基于m/z和RT的机器学习框架：非靶向代谢组学中脂质分类的 ... [https://www.ebiotrade.com/newsf/2025-10/20251019204802775.htm]
• 35. 基于谱图特征解析与机器学习模型融合的非靶向污染物识别 ... [https://blog.csdn.net/max500600/article/details/154026731]
• 36. Navigating Through the Noise: A Roadmap for Combining ... [https://link.springer.com/article/10.1007/s11692-025-09655-w]
• 37. Batch Effect Correction [https://www.10xgenomics.com/analysis-guides/introduction-batch-effect-correction]
• 38. Comprehensive noise reduction in single-cell data with the ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12570323/]
• 39. Comparison of missing value imputation tools for machine ... [https://www.sciencedirect.com/science/article/pii/S0023643825002695]
• 40. A comparative study of imputation techniques for missing ... [https://link.springer.com/article/10.1007/s41060-025-00825-9]
• 41. Protein-level batch-effect correction enhances robustness ... [https://www.nature.com/articles/s41467-025-64718-y]
• 42. A comparison of various imputation algorithms for missing data [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0319784]
• 43. Top 12 Dimensionality Reduction Techniques for Machine ... [https://encord.com/blog/dimentionality-reduction-techniques-machine-learning/]
• 44. Evaluation of a decided sample size in machine learning ... [https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-023-05156-9]
• 45. A Survey on Small Sample Imbalance Problem: Metrics, ... [https://arxiv.org/html/2504.14800v1]
• 46. Feature Selection in Healthcare Datasets: Towards a ... [https://www.sciencedirect.com/science/article/pii/S0010482525011631]
• 47. Accuracy, robustness and scalability of dimensionality ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6902413/]
• 48. AI-driven discovery of synergistic drug combinations ... [https://www.nature.com/articles/s41467-025-56818-6]
• 49. Developing a robust machine learning framework for ... [https://www.sciencedirect.com/science/article/pii/S2352710225004401]
• 50. Integration of machine learning and experimental ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12373879/]
• 51. The trade-off between explainability and model complexity ... [https://milvus.io/ai-quick-reference/what-is-a-tradeoff-between-explainability-and-model-complexity]
• 52. 2 Interpretability – Interpretable Machine Learning [https://christophm.github.io/interpretable-ml-book/interpretability.html]
• 53. The Explainability-Interpretability Trade-Off In AI [https://www.wissen.com/blog/the-explainability-interpretability-trade-off-in-ai]
• 54. Integration of machine learning and experimental validation ... [http://www.chinaphar.com/article/view/11353]
• 55. The Accuracy vs Interpretability Trade-off Is a Lie [https://medium.com/data-science/the-accuracy-vs-interpretability-trade-off-is-a-lie-0bf89c84f5b3]
• 56. A Review of Feature Selection Methods for Machine Learning ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9580915/]
• 57. Hyperparameter Tuning [https://www.geeksforgeeks.org/machine-learning/hyperparameter-tuning/]
• 58. Model validation in machine learning: How to do it [https://www.leewayhertz.com/model-validation-in-machine-learning/]
• 59. A Comprehensive Guide to Hyperparameter Tuning in ... [https://medium.com/@aditib259/a-comprehensive-guide-to-hyperparameter-tuning-in-machine-learning-dd9bb8072d02]
• 60. How to Do Model Visualization in Machine Learning? [https://neptune.ai/blog/visualization-in-machine-learning]
• 61. Categorizing Concentration Confidence: A Framework for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12068507/]
• 62. Using Validation Dataset and Cross-validation Techniques [https://www.deepchecks.com/evaluating-model-performance-using-validation-dataset-splits-and-cross-validation-techniques/]
• 63. Generalizability of Machine Learning Models: Quantitative ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9885377/]
• 64. A Guide to Cross-Validation for Artificial Intelligence in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10388213/]
• 65. Cross-Validation in Machine Learning: How to Do It Right [https://neptune.ai/blog/cross-validation-in-machine-learning-how-to-do-it-right]
• 66. Evaluating generalizability of artificial intelligence models ... [https://www.nature.com/articles/s42256-024-00931-6]
• 67. Cross-validation (statistics) [https://en.wikipedia.org/wiki/Cross-validation_(statistics)]
• 68. How to choose a predictive model after k-fold cross-validation? [https://stats.stackexchange.com/questions/52274/how-to-choose-a-predictive-model-after-k-fold-cross-validation]
• 69. Assessing the Generalizability of a Clinical Machine ... [https://www.sciencedirect.com/science/article/pii/S2542454822000145]
• 70. Defining Biological and Clinical Plausibility: The DICSA ... [https://link.springer.com/article/10.1007/s40273-025-01485-0]
• 71. An Evaluation of Interpretable Machine Learning Models [https://link.springer.com/article/10.1007/s12599-024-00922-2]
• 72. Interpretability and accuracy of machine learning ... [https://www.sciencedirect.com/science/article/pii/S1746809425006640]
• 73. Real-World Data and Causal Machine Learning to Enhance ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12579681/]
• 74. Demystifying the Accuracy-Interpretability Trade-Off [https://arxiv.org/html/2503.07914v1]
• 75. Comparative analysis of deep learning and tree-based ... [https://pubmed.ncbi.nlm.nih.gov/40630870/]
• 76. Framework for Benchmarking Machine Learning Models ... [https://aisel.aisnet.org/amcis2025/data_science/sig_dsa/7/]
• 77. [2504.13151] MIB: A Mechanistic Interpretability Benchmark [https://arxiv.org/abs/2504.13151]
• 78. Examining structure-based surrogate selection for quantitative... [https://pubmed.ncbi.nlm.nih.gov/40488856/]
• 79. Qualitative vs Quantitative Data Differences [https://statisticsbyjim.com/basics/qualitative-vs-quantitative/]