Qian Cui1,

1 College of Electrical and Automation Engineering, East China Jiaotong University, 808 East Shuanggang Street, Nanchang 330013, Jiangxi, China

Article type: Review

Running title: XAI and Railway Particle Toxicology

Subject categories: A3 Metals; A8 Environmental pollutants; A11 Nanomaterials; B7 Respiratory system; D11 Toxicity evaluation methods.

*Corresponding author: Qian Cui, College of Electrical and Automation Engineering, East China Jiaotong University, 808 East Shuanggang Street, Nanchang 330013, Jiangxi, China.

E-mail: 2023021005000325@ecjtu.edu.cn; alternate e-mail: cuiqian041016@163.com

Phone: +86-791-87046184

 

Abstract

Railway and underground railway environments contain metal-enriched particulate matter generated by brake wear, wheel-rail contact and electrical-contact abrasion. These particles are often dominated by iron oxides and may contain copper, manganese, zinc, barium, chromium and nickel, yet their toxicological significance remains incompletely defined. This review synthesizes recent evidence published mainly from 2020 to 2026 on railway and brake-wear particles, focusing on oxidative stress, inflammatory signaling, mitochondrial dysfunction and explainable deep learning for particle-toxicity assessment. Exposure studies show that subway and railway microenvironments can reach high fine-particle concentrations, while test-bench and toxicological studies indicate that composition, size fraction, surface reactivity and soluble metals influence biological responses. In lung epithelial cells, macrophages and animal models, metal-rich brake-wear particles induce reactive oxygen species, antioxidant depletion, cytokine release, metabolic reprogramming, mitochondrial perturbation and DNA damage. Evidence remains fragmented because particle generation, sampling, dosing, cell models and endpoint selection vary widely. We propose an integrated framework coupling standardized particle characterization, airway-relevant assays and explainable artificial intelligence to identify toxicity-driving features such as iron, copper, oxidative potential and mitochondrial endpoints. This framework may support mechanism-based hazard ranking, safer railway material design and exposure-mitigation strategies.

Keywords: Railway brake-wear particles; metal-enriched particulate matter; oxidative stress; inflammation; mitochondrial dysfunction; explainable deep learning

Abbreviations: AI, artificial intelligence; DCFH-DA, 2,7-dichlorodihydrofluorescein diacetate; DTT, dithiothreitol; EBC, exhaled breath condensate; ICP-MS, inductively coupled plasma mass spectrometry; LDH, lactate dehydrogenase; ML, machine learning; MMP, mitochondrial membrane potential; PM, particulate matter; ROS, reactive oxygen species; SHAP, SHapley Additive exPlanations; XAI, explainable artificial intelligence.

 

Introduction

Transport-derived particulate matter has entered a transitional period. Exhaust emissions have been progressively reduced by fuel, engine and after-treatment regulations, whereas non-exhaust emissions from brakes, tyres, rails, wheels, power-collection systems and resuspended dust remain difficult to control. In road transport, brake and tyre wear particles are now recognized as major urban non-exhaust contributors (Fussell et al., 2022; Beddows et al., 2023; Christou et al., 2025). In rail transport, the same general problem is intensified by repeated metal-on-metal contact, enclosed station architecture and high passenger throughput. The toxicological question is therefore no longer whether railway environments contain particles, but which particle characteristics drive hazard and how the evidence can be integrated into interpretable risk assessment.

Railway and subway particles are chemically distinctive. Underground-railway reviews and monitoring studies show that subway PM2.5 is frequently enriched in iron-rich particles derived from wheel, track, brake and rail-system wear (Loxham and Nieuwenhuijsen, 2019; Smith et al., 2020; Chang et al., 2021; Luglio et al., 2021; Sheikh et al., 2022; Azad et al., 2023). These findings suggest that railway-derived particles should not be assessed simply as generic ambient PM2.5; instead, their metal composition, surface reactivity, solubility and particle-size distribution require toxicological interpretation.

In recent years, brake-wear toxicology has advanced quickly. Systematic reviews have concluded that mammalian evidence remains limited but consistently points toward oxidative stress, pro-inflammatory responses and DNA damage as recurring outcomes (Forest and Pourchez, 2023; Christou et al., 2025). New in vitro studies have strengthened the causal link between metal composition and cellular responses, especially for copper-enriched non-asbestos organic and ceramic brake-wear particles. Earlier cell studies established that subway and brake-wear particles can induce epithelial toxicity, oxidative stress and genotoxicity (Karlsson et al., 2005; Gasser et al., 2009). More recent controlled particle-generation and in vivo studies have characterized non-asbestos organic and low-metallic brake-pad emissions and linked brake-pad PM2.5 with pulmonary, metabolic and microbiota alterations (Woo et al., 2022; Zhou et al., 2025).

The railway-specific evidence base is narrower than the automotive brake-wear literature, but the overlap is mechanistically relevant. Railway brake systems, wheel–rail interfaces and pantograph–catenary contacts generate particles enriched in Fe, Cu and other metals. These metals can catalyze reactive oxygen species generation, interfere with cellular metal homeostasis, activate inflammatory signaling and impair mitochondrial energy metabolism. For toxicology journals, the conceptual novelty lies in translating the recent brake-wear and subway exposure literature into a mechanism-oriented framework tailored to railway environments.

A second novelty is computational. Deep learning and machine learning are increasingly used in predictive toxicology (Ajisafe et al., 2025), but their application to environmental particulate mixtures is still emerging. Explainable artificial intelligence (XAI), including SHAP-based interpretation, can help identify which particle characteristics and biological endpoints most strongly influence model predictions. For railway particle toxicology, XAI is attractive because the exposure is a complex mixture problem: multiple metals, different size fractions, varying operational conditions, and multiple correlated biological endpoints. However, AI should not be presented as a black-box replacement for toxicology; it should operate as a transparent integrator of physicochemical and mechanistic evidence.

This review therefore aims to synthesize recent publications on railway/subway particles, brake-wear toxicology and explainable toxicity modeling; identify convergent mechanisms involving oxidative stress, inflammation and mitochondrial dysfunction; and propose a direct, testable framework for explainable deep learning-assisted toxicological assessment of metal-enriched railway brake-wear particles.

Materials and Methods

This review was designed as a narrative and mechanism-oriented synthesis rather than a quantitative meta-analysis. Literature was identified using PubMed, Web of Science, Scopus, Google Scholar and journal websites. The primary search window was January 2020 to July 2026, because the purpose was to capture recent evidence. Older landmark studies were retained where they provided essential mechanistic or exposure context, such as early in vitro subway particle and brake-wear toxicity studies. AI-tool use was limited to language editing, reference-format checking and manuscript formatting; no AI tool was used to generate research data, determine eligibility of evidence or replace authorial interpretation.

Search terms included combinations of “railway particles”, “subway PM2.5”, “underground railway particulate matter”, “brake wear particles”, “rail disc brake emissions”, “wheel-rail contact particles”, “metal-rich particulate matter”, “iron oxide nanoparticles”, “copper brake particles”, “oxidative stress”, “inflammation”, “mitochondrial dysfunction”, “DNA damage”, “airway epithelial cells”, “macrophages”, “machine learning toxicity”, “deep learning toxicology”, “oxidative potential particulate matter” and “SHAP”.

Articles were prioritized when they met one or more of the following criteria: (1) measured railway or subway particle concentrations, composition or source characteristics; (2) generated or characterized brake-wear or rail-wear particles under controlled test-bench conditions; (3) evaluated mammalian in vitro, in vivo or occupational biomarkers relevant to inhalation toxicity; (4) reported oxidative stress, inflammatory, mitochondrial, genotoxic or epithelial-barrier endpoints; or (5) used machine learning or XAI for toxicity, oxidative potential or particulate-matter hazard prediction. Articles focusing exclusively on tyre microplastics, non-mammalian ecotoxicology or unrelated chemical toxicity were used only where they clarified broader non-exhaust-emission trends.

Evidence was grouped into four domains: source and exposure, particle physicochemistry, biological mechanisms, and computational interpretation. The resulting synthesis is intended to support a review manuscript for Journal of Toxicological Sciences and to generate a study framework that can be tested in future original experiments.

Results and Discussion

Railway and subway particles as a distinct metal-rich exposure problem

Recent exposure studies show that railway and subway microenvironments can be high-concentration PM environments. In New York City, Azad et al. (2024) reported mean underground platform PM2.5 of 139 ± 25 μg/m3 and underground train PM2.5 of 99 ± 21 μg/m3, and used a network model to estimate commuter-level exposure. The study also highlighted exposure inequality, with higher exposure among workers from economically disadvantaged and minority communities. These findings are important for toxicology because they show that railway particle exposure is not only an environmental chemistry problem but also an inhalation exposure and public-health problem.

London Underground studies provide complementary evidence. Smith et al. (2020) reported that sampled London Underground PM2.5 was composed largely of iron oxide, with additional elemental carbon, organic carbon and metallic/mineral oxides. Sheikh et al. (2022) further showed that London Underground PM is dominated by 5–500 nm maghemite particles occurring as aggregated clusters. Hedges et al. (2026) extended this exposure context by measuring particle number size distributions in London subway train carriages. Together, these results indicate that underground railway PM is chemically and morphologically distinct from many urban background particles and may require different toxicological assumptions.

The source profile of railway particles is complex. Fruhwirt et al. (2023) measured PM emissions from wheel-rail contact and reported PM10 emission factors from 0.33 to 5.57 g/km, with high proportions of PM2.5 and PM1 within PM10. A subsequent rail disc-brake test-bench study identified contact strips, wires, wheels, rails and brakes as major non-exhaust railway sources and found PM10 emission factors from axle-mounted rail disc brakes in the range of 0.11 to 4.21 g/km (Fruhwirt et al., 2024). These test-bench studies are important because they move beyond station-level concentration measurements and provide source-specific information relevant to intervention.

Occupational biomonitoring adds another layer. In Paris subway-worker studies, Sauvain et al. (2024, 2025) measured metals in exhaled breath condensate and urine and linked elevated metal biomarkers with altered bronchopulmonary metabolites and systemic oxidative-stress markers. Although such observational data cannot prove causality, it supports the biological plausibility that metal-rich underground railway exposures are reflected in human biomarkers. A robust review must therefore connect environmental measurements, source characterization, and biological effects rather than treat them as separate literatures.

Mechanistic toxicology of brake-wear and railway-relevant metal particles

The strongest mechanistic theme across recent brake-wear studies is oxidative stress. Hyman et al. (2026) comprehensively characterized airborne size-separated brake-wear particles from two brake-pad types and found iron, primarily as iron oxide, to be the most abundant element, accounting for 33–50% of particle mass. Their acellular and cellular assays showed concentration-dependent reactive oxygen species production, antioxidant depletion and oxidative stress-mediated DNA damage. Importantly, the non-asbestos organic pad with substantially higher copper content produced stronger DNA damage and acellular antioxidant depletion, again pointing toward metals as toxicity modifiers.

Parkin et al. (2025) provided one of the most informative recent cell-based datasets. Using alveolar type-II epithelial cells, they found that non-asbestos organic and ceramic brake-wear PM induced the strongest oxidative stress, with increased ROS generation and upregulation of HMOX1 and GCLM. RNA-seq analysis showed thousands of differentially expressed genes after exposure to specific brake-wear PM types, while gene-set analysis identified TNF-alpha/NF-kappaB, oxidative-stress and hypoxia pathways. The same study found strong metal-homeostasis responses and identified copper enrichment as a major determinant of cellular perturbation.

Inflammation is the second recurring pathway. Selley et al. (2020) reported that brake abrasion dust induced inflammatory cytokine release in macrophages and transiently compromised phagocytosis, with biological effects attenuated by metal chelation. Parkin et al. (2025) observed IL-6 and IL-8 secretion in alveolar epithelial cells after exposure to specific brake-wear particles. These findings suggest that metal-rich brake particles can activate both epithelial and innate immune pathways, which is relevant to airway inflammation and occupational exposure.

Mitochondrial effects are increasingly recognized but less standardized across studies. Brake-wear particles can produce oxidative stress near or within mitochondria, modify ATP production, shift metabolism toward glycolysis, reduce mitochondrial membrane potential, and alter mitophagy or cell-death signaling. Parkin et al. (2025) observed metabolic reprogramming toward glycolytic ATP production in response to non-asbestos organic brake-wear PM, whereas Selley et al. (2020) reported mitochondrial depolarization in macrophages. These findings support including mitochondrial membrane potential, mitochondrial ROS, ATP production and metabolic flux in future railway particle assays.

Epithelial-barrier and mucosal responses may explain why overt cytotoxicity is not always observed. Puisney-Dakhli et al. (2023) examined brake-wear nanoparticles in Calu-3 airway epithelial barriers and found limited overt cytotoxicity or pro-inflammatory response, but small translocation across the barrier, increased MUC5AC expression and altered repair-associated proliferation. This is important because a narrow cytotoxicity-only screen may miss airway-relevant effects such as mucus overproduction, barrier interaction and abnormal repair signaling.

Animal evidence is still limited but has become more informative. Kim et al. (2026) generated brake-wear particles from non-asbestos organic and low-metallic brake pads and compared their physicochemical properties, intrinsic oxidative potential and pulmonary inflammatory effects. The study found that surface-area normalization changed toxicity ranking, highlighting a key methodological issue: dose metrics based only on mass may obscure the importance of surface area, solubility and particle reactivity. Such results should inform railway particle testing, where ultrafine fractions and aggregated iron oxides are common.

Why railway brake-wear particles cannot be reduced to automotive brake-wear particles

The automotive brake-wear literature provides valuable mechanistic analogies, but railway particles are not identical. Railway systems involve different braking materials, higher axle loads, steel wheel–rail contact, enclosed platforms, piston-effect ventilation, and power-collection wear. The same particle mass may therefore represent a different mixture of iron oxides, copper-containing fragments, manganese, barium, zinc, chromium, nickel and carbonaceous components. In addition, underground systems repeatedly resuspend settled particles, increasing the residence time of metal-rich dust.

The railway environment also changes exposure dynamics. Commuters experience intermittent short-duration exposure, whereas workers may experience repeated occupational exposure. Platform staff, train operators, maintenance workers and tunnel workers may have distinct dose profiles. Occupational biomonitoring studies therefore complement commuter-exposure studies and should be integrated into a unified risk framework. For direct toxicological testing, this means exposure scenarios should be stratified by particle source, dose rate, duration and susceptible population rather than by particle mass alone.

A railway-specific review must also consider the engineering side. Test-bench data suggest that operational factors such as brake-pad material, temperature, braking intensity and wheel–rail contact conditions change emission rates and particle-size distribution. These variables are rarely incorporated into toxicological design. The next stage of the field should connect engineering variables with biological endpoints: for example, linking brake-pad composition and disc temperature to Fe/Cu-rich ultrafine particle output, oxidative potential and epithelial stress responses.

Explainable deep learning as an integrative tool rather than a replacement for toxicology

The biological evidence described above is highly multidimensional. A typical railway particle dataset can contain source variables, particle-size fractions, morphology, specific surface area, zeta potential, soluble and total metal concentrations, acellular oxidative potential, dose metrics, cell type, exposure duration and multiple biological endpoints. Traditional regression can identify individual associations but may underperform when interactions and nonlinear dose-response patterns dominate. Machine learning and deep learning can improve prediction, but black-box prediction is insufficient for toxicology unless the model output is interpretable and biologically plausible.

The relevance of XAI is demonstrated by recent particulate-matter and toxicology modeling studies. Esu et al. (2024) developed FLIT-SHAP to extract dose-response relationships from multi-pollutant oxidative potential data and showed that explainable ML can reveal interaction effects in particle mixtures. Kim and Choi (2025) reviewed 93 peer-reviewed ToxCast-based AI toxicity models and emphasized the need for explainability, data-quality control and future integration into next-generation risk assessment. These studies do not directly solve railway particle toxicity, but they provide a methodological foundation for interpreting mixture-driven endpoints.

For railway brake-wear particles, a useful XAI model should predict a biologically interpretable target such as low, moderate or high toxicity severity; ROS-dominant versus inflammation-dominant response class; or a continuous composite toxicity score. Input features should be separated into physicochemical features and biological-response features. Physicochemical features may include Fe, Cu, Mn, Zn, Ba, Cr and Ni concentrations, soluble metal fraction, hydrodynamic size, surface area and oxidative potential. Biological features may include viability, LDH, ROS, GSH, MDA, HMOX1, NQO1, IL-6, IL-8, TNF-alpha, mitochondrial membrane potential, ATP production, mitochondrial ROS, MUC5AC and DNA damage.

The most important requirement is interpretability. SHAP, partial-dependence plots, accumulated local effects and interaction values can show whether Fe, Cu, oxidative potential, ROS or mitochondrial markers are driving model predictions. Such explanations must then be checked against known toxicological mechanisms. For example, a model that ranks copper, ROS and IL-8 highly is plausible in light of recent copper-enriched brake-wear studies, whereas a model that ranks sample ID or batch as the strongest determinant would indicate confounding. XAI therefore acts as a quality-control tool as well as a hypothesis generator.

Model validation is equally important. Particle toxicology datasets are often small, imbalanced and batch-dependent. A direct submission-quality study would require nested cross-validation, external validation using particles from another railway system or test bench, sensitivity analysis for dose metrics, and transparent reporting of preprocessing decisions. For reviews, the recommendation is clear: future railway particle toxicology should collect data in a format that supports modeling from the beginning rather than add AI after the experiment is completed.

Proposed toxicity-assessment framework for future railway particle studies

On the basis of recent evidence, a direct framework can be proposed. First, particle collection should separate sources when possible: brake-pad/disc emissions, wheel–rail particles, station dust, train-cabin particles and power-collection wear. Second, particle characterization should combine SEM or TEM, EDS, ICP-MS, hydrodynamic size, zeta potential, surface area and acellular oxidative-potential assays. Third, biological testing should include airway epithelial cells, macrophages and ideally air–liquid interface or co-culture systems to capture barrier and immune interactions.

The endpoint panel should be mechanistically tiered. Tier 1 should include viability, LDH and basic dose-response screening. Tier 2 should evaluate oxidative stress using ROS, GSH, MDA, HMOX1, NQO1 and DTT/ascorbate depletion. Tier 3 should evaluate inflammation using IL-6, IL-8, TNF-alpha, NF-kappaB and macrophage activation. Tier 4 should evaluate mitochondria and metabolism using mitochondrial membrane potential, ATP production, mitochondrial ROS and metabolic flux. Tier 5 should include airway-barrier and genotoxic endpoints, such as MUC5AC, permeability, wound repair, comet assay and 8-OHdG. This tiered approach helps avoid both under-testing and unfocused endpoint accumulation.

Dose translation is a critical unresolved issue. Many in vitro studies use μg/cm2 mass doses, whereas real-world exposure is measured as μg/m3 over time. Future studies should report estimated deposited dose, exposure duration, particle density, aerodynamic diameter and whether mass, surface area or particle number best represents biological potency. The Kim et al. (2026) surface-area-normalized results show that mass-based ranking alone may be misleading. For railway particles dominated by ultrafine aggregates, this issue is especially important.

Finally, AI should be embedded in the framework from the beginning. The minimum reporting set should include raw feature matrices, feature definitions, units, missing-value handling, batch identifiers, model architecture, hyperparameters, validation strategy, performance metrics and explanation plots. The model should be framed as a decision-support and mechanistic-prioritization tool, not as proof of causality. Causality still depends on toxicological perturbation experiments, such as metal chelation, antioxidant pretreatment, mitochondrial inhibitors and source-material substitution.

Research gaps and future directions

First, railway-specific toxicology is still far less developed than automotive brake-wear toxicology. More original studies are needed using particles collected from real railway braking systems, wheel–rail contacts and station environments. These studies should not simply import automotive conclusions but should test whether railway particles have distinct toxicity profiles due to their iron oxide dominance, source mixture and enclosed exposure setting.

Second, human relevance needs strengthening. Most in vitro studies use submerged monocultures, yet airway exposure occurs at the air–liquid interface with mucus, surfactant, immune cells and repeated low-dose exposure. Primary bronchial epithelial cells, induced pluripotent stem cell-derived airway models and epithelial–macrophage co-cultures would provide more mechanistic relevance. Occupational biomonitoring should be linked with particle measurements and biological endpoints to connect exposure to early effect markers.

Third, mitochondrial dysfunction should become a standard endpoint. Many particle studies stop at ROS and cytokines, but mitochondria link redox imbalance, metabolism, inflammation and cell death. Measurement of mitochondrial membrane potential, mitochondrial ROS, ATP source, oxygen-consumption rate and mitophagy-related signaling would clarify whether metal-rich railway particles disrupt energy metabolism directly or indirectly through oxidative stress.

Fourth, evidence harmonization is needed. Differences in particle generation, sampling medium, storage, sterilization, endotoxin control, dispersion method, sonication, protein corona and exposure dose make cross-study comparison difficult. Standardized reporting could dramatically increase the value of future studies and make data suitable for machine-learning integration.

Fifth, intervention studies remain sparse. Toxicological evidence should be connected to railway engineering controls such as brake material substitution, regenerative braking, ventilation, platform screen doors, tunnel cleaning, filtration and operational changes. The most useful toxicity framework would not only identify hazard but also rank mitigation strategies by reducing emission quantity, particle reactivity or deposited dose.

Conclusion

Recent evidence indicates that metal-enriched railway and brake-wear particles represent an inhalation-toxicology problem that cannot be fully understood by mass concentration alone. Railway and subway environments can contain high concentrations of Fe-rich fine particles, while brake-wear studies show that metal composition, especially iron and copper, surface reactivity and size fraction strongly influence oxidative stress, inflammatory signaling, mitochondrial dysfunction and DNA damage. The field is moving from simple cytotoxicity screening toward integrated mechanistic assessment, but railway-specific data remain limited. Explainable deep learning offers a practical way to integrate heterogeneous physicochemical and biological endpoints, identify toxicity-driving features and guide future experiments. A direct next step is to develop standardized railway particle datasets that combine source-resolved characterization, airway-relevant toxicology and transparent XAI interpretation. Such a framework could support safer railway materials, targeted exposure mitigation and mechanism-based toxicity evaluation.

Acknowledgments

The author thanks researchers whose recent work on non-exhaust emissions, railway particulate matter, brake-wear toxicology and explainable toxicity modeling provided the basis for this review.

Funding

This review received no specific grant from any funding agency in the public, commercial or not-for-profit sectors.

Conflict of Interest Statement

The author declares no competing interests.

Data Availability Statement

No new experimental datasets were generated or analyzed in this review. The literature discussed is available in the cited publications.

Author Contribution Statement

Conceptualization: Qian Cui. Literature curation: Qian Cui. Visualization: Qian Cui. Writing – original draft: Qian Cui. Writing – review & editing: Qian Cui.

Ethics approval and consent to participate

Not applicable. This review did not involve human participants, human materials or animal experiments.

Patient consent for publication

Not applicable.

References

Ajisafe, O.M., Adekunle, Y.A., Egbon, E., Ogbonna, C.E. and Olawade, D.B. (2025): The role of machine learning in predictive toxicology: A review of current trends and future perspectives. Life Sci., 378, 123821.

Azad, S., Ferrer-Cid, P. and Ghandehari, M. (2024): Exposure to fine particulate matter in the New York City subway system during home-work commute. PLoS One, 19, e0307096.

Azad, S., Ghandehari, M., Gordon, T., Luglio, D.G. and Thurston, G. (2023): Particulate matter concentration and composition in the New York City subway system. Atmos. Pollut. Res., 14, 101767.

Beddows, D.C.S., Harrison, R.M., Gonet, T., Maher, B.A. and Odling, N. (2023): Measurement of road traffic brake and tyre dust emissions using both particle composition and size distribution data. Environ. Pollut., 331, 121830.

Chang, L., Chong, W.T., Wang, X., Pei, F., Zhang, X. and Wang, T. (2021): Recent progress in research on PM2.5 in subways. Environ. Sci. Process. Impacts, 23, 642-663.

Christou, A., Louka, P., Gkatzelis, G.I., Katsoulis, B.D. and Pandis, S.N. (2025): Review of health effects of automotive brake and tyre wear particles. Toxics, 13, 301.

Esu, C.O., Pyo, J. and Cho, K. (2024): Machine learning-derived dose-response relationships considering interactions in mixtures: applications to the oxidative potential of particulate matter. J. Hazard. Mater., 475, 134864.

Forest, V. and Pourchez, J. (2023): Biological effects of brake wear particles in mammalian models: a systematic review. Sci. Total Environ., 905, 167266.

Fruhwirt, D., Sturm, P.J., Nöst, T., Leonhardt, P., Bode, G., Michael, S. and Rodler, J. (2023): PM emissions from railways – results of tests on a wheel-rail test bench. Transp. Res. Part D Transp. Environ., 122, 103858.

Fruhwirt, D., Nöst, T., Leonhardt, P., Leitner, M., Brunnhofer, P., Bode, G., Michael, S., Sturm, P.J. and Rodler, J. (2024): Characterization of emissions from axle-mounted rail disc brakes. Transp. Res. Part D Transp. Environ., 130, 104181.

Fussell, J.C., Franklin, M., Green, D.C., Gustafsson, M., Harrison, R.M., Hicks, W., Kelly, F.J., Kishta, F., Miller, M.R., Mudway, I.S. and Oroumiyeh, F. (2022): A review of road traffic-derived non-exhaust particles: emissions, physicochemical characteristics, health risks, and mitigation measures. Environ. Sci. Technol., 56, 6813-6835.

Gasser, M., Riediker, M., Mueller, L., Perrenoud, A., Blank, F., Gehr, P. and Rothen-Rutishauser, B. (2009): Toxic effects of brake wear particles on epithelial lung cells in vitro. Part. Fibre Toxicol., 6, 30.

Hedges, M., Priestman, M., Chadeau-Hyam, M., Sinharay, R., Kent, A.E., Kelly, F.J. and Green, D.C. (2026): Exposure assessment and particle number size distribution in train carriages in the London subway. Atmos. Environ., 367, 121745.

Hyman, S., Saladin, S., Tsybrii, Y., Nosko, O., Williams, M., Zherebker, A., Risby, K., Topping, D., Boies, A., Giorio, C., Roursgaard, M. and Møller, P. (2026): Physicochemical and toxicological characterization of airborne brake wear particles reveals oxidative stress-mediated DNA damage. Environ. Sci. Technol., 60, 4955-4965.

Karlsson, H.L., Nilsson, L. and Möller, L. (2005): Subway particles are more genotoxic than street particles and induce oxidative stress in cultured human lung cells. Chem. Res. Toxicol., 18, 19-23.

Kim, D. and Choi, J. (2025): AI-based toxicity prediction models using ToxCast data: current status and future directions for explainable models. Toxicology, 517, 154230.

Kim, G., Jeon, S., Shin, T.H., Park, Y., Kim, K., Woo, S.-H., Huh, Y.S., Lee, J.-G., Lee, S. and Cho, W.-S. (2026): Pulmonary toxicity of brake wear particles: effects of composition and surface reactivity. Part. Fibre Toxicol., 23, 27.

Loxham, M. and Nieuwenhuijsen, M.J. (2019): Health effects of particulate matter air pollution in underground railway systems: a critical review of the evidence. Part. Fibre Toxicol., 16, 12.

Luglio, D.G., Katsigeorgis, M., Hess, J., Kim, R., Adragna, J., Raja, A., Gordon, C., Fine, J., Thurston, G., Gordon, T. and Vilcassim, M.R. (2021): PM2.5 concentration and composition in subway systems in the Northeastern United States. Environ. Health Perspect., 129, 027001.

Parkin, J.G.H., Dean, L.S.N., Bell, J.A., Easton, N.H.C., Edgeway, L.J., Cooper, M.J., Ridley, R., Conforti, F., Wang, S., Yao, L. and Loxham, M. (2025): Copper-enriched automotive brake wear particles perturb human alveolar cellular homeostasis. Part. Fibre Toxicol., 22, 4.

Puisney-Dakhli, C., Oikonomou, E.K., Tharaud, M., Sivry, Y., Berret, J.F. and Baeza-Squiban, A. (2023): Effects of brake wear nanoparticles on the protection and repair functions of the airway epithelium. Environ. Pollut., 327, 121554.

Sauvain, J.J., Hemmendinger, M., Charreau, T., Jouannique, V., Debatisse, A., Suárez, G., Hopf, N.B. and Guseva Canu, I. (2024): Metal and oxidative potential exposure through particle inhalation and oxidative stress biomarkers: a 2-week pilot prospective study among Parisian subway workers. Int. Arch. Occup. Environ. Health, 97, 387-400.

Sauvain, J.J., Wild, P., Charreau, T., Jouannique, V., Sakthithasan, K., Debatisse, A., Suárez, G., Hopf, N.B. and Guseva Canu, I. (2025): Are metals in exhaled breath condensate and urine associated with oxidative/nitrosative stress and metabolism-related biomarkers? Results from 303 randomly selected Parisian subway workers. Environ. Int., 196, 109325.

Selley, L., Schuster, L., Marbach, H., Forster, R., Forbes, B., Gant, T.W., Sandström, T., Mudway, I.S. and Kumar, A. (2020): Brake dust exposure exacerbates inflammation and transiently compromises phagocytosis in macrophages. Metallomics, 12, 371-386.

Sheikh, H.A., Tung, P.Y., Ringe, E., Harrison, R.J. and Maher, B.A. (2022): Magnetic and microscopic investigation of airborne iron oxide nanoparticles in the London Underground. Sci. Rep., 12, 20298.

Smith, J.D., Barratt, B.M., Fuller, G.W., Kelly, F.J., Loxham, M., Nicolosi, E., Priestman, M., Tremper, A.H. and Green, D.C. (2020): PM2.5 on the London Underground. Environ. Int., 134, 105188.

Woo, S.-H., Jang, H., Na, M.Y., Chang, H.J. and Lee, S. (2022): Characterization of brake particles emitted from non-asbestos organic and low-metallic brake pads under normal and harsh braking conditions. Atmos. Environ., 278, 119089.

Zhou, L., Song, C., Lei, Y., Zhao, L., Han, Y., Xu, Y., Li, B. and Guo, J. (2025): Health impacts of PM2.5 emissions from brake pad wear: a comprehensive study on pulmonary, metabolic, and microbiota alterations. Toxicology, 511, 154055.

Figure legends

Fig. 1. Literature selection and evidence synthesis workflow. The review was designed as a narrative evidence synthesis covering recent exposure, source-characterization, toxicological and explainable-AI literature. It is not a quantitative meta-analysis.

Fig. 2. Source-to-effect framework for railway-derived metal-rich particles. Railway brake pads/discs, wheel–rail contact and electrical-contact systems produce Fe-, Cu- and other metal-enriched particle mixtures that interact with airway cells and may induce oxidative stress, inflammation, mitochondrial dysfunction, barrier effects and genotoxicity.

Fig. 3. Explainable deep learning framework for particle toxicity assessment. Physicochemical and biological features can be integrated into transparent toxicity prediction models, while XAI methods such as SHAP help identify the main toxicity-driving variables and support mechanistic interpretation.

Fig. 4. Author-curated evidence map of cited recent literature. Counts indicate the number of cited studies or review/guidance articles used in each evidence category in this narrative review; the figure is descriptive and not intended as a bibliometric analysis.

 

 

Table 1. Representative recent studies supporting a railway brake-wear particle toxicology review.

Study	Particle/source	Approach	Key finding	Relevance to this review
Smith et al. (2020)	London Underground PM2.5	Environmental monitoring and chemical composition	PM2.5 contained a large iron oxide fraction with carbonaceous and metallic/mineral components.	Shows railway PM is compositionally distinct from urban background particles.
Sheikh et al. (2022)	London Underground iron-rich particles	Magnetic and microscopic characterization	Dominance of nanoscale maghemite particles in aggregated clusters.	Supports nanoscale Fe-oxide mechanism and surface-reactivity concerns.
Fruhwirt et al. (2023)	Wheel–rail contact particles	Wheel–rail test bench	PM10 emission factors varied by test condition; PM2.5 and PM1 were major fractions.	Connects engineering variables with respirable railway PM.
Fruhwirt et al. (2024)	Axle-mounted rail disc brakes	Rail brake test bench	Organic pads produced higher emissions than sintered materials; PM10 factors up to 4.21 g/km.	Identifies brake material and temperature as emission determinants.
Azad et al. (2024)	NYC subway PM2.5	Network exposure modeling	Underground platform mean 139 μg/m3; train mean 99 μg/m3; exposure inequality observed.	Links railway PM to commuter exposure and public-health relevance.
Sauvain et al. (2025)	Paris subway workers	EBC/urine metals and biomarkers	Metals associated with bronchopulmonary metabolites and systemic oxidative-stress markers.	Adds occupational biomarker evidence.
Selley et al. (2020)	Brake abrasion dust	Macrophage in vitro study	Inflammation, mitochondrial depolarization and impaired phagocytosis; metal chelation attenuated effects.	Supports metal-dependent immune and mitochondrial mechanisms.
Puisney-Dakhli et al. (2023)	Brake-wear nanoparticles	Calu-3 epithelial barrier model	Small translocation, increased MUC5AC and altered repair response without overt cytotoxicity.	Shows barrier/mucus endpoints should be included.
Parkin et al. (2025)	Copper-enriched brake-wear PM	ATII cells, RNA-seq, cytokines and metal analysis	NAO/ceramic particles induced ROS, IL-6/IL-8, glycolytic shift and metal-homeostasis responses; copper was a key driver.	Strong evidence for composition-specific toxicity.
Hyman et al. (2026)	Airborne brake-wear particles	Physicochemical and A549 toxicology assays	Fe oxide was abundant; particles caused ROS, antioxidant depletion and DNA damage; Cu-rich NAO particles were more damaging.	Supports integrated characterization plus toxicology.
Kim et al. (2026)	Brake-wear PM2.5	Mouse pulmonary toxicity and oxidative potential	Toxicity ranking depended on mass versus surface-area normalization.	Highlights importance of dose metric and surface reactivity.
Esu et al. (2024)	PM oxidative potential mixtures	FLIT-SHAP explainable ML	XAI extracted nonlinear dose-response and interaction patterns.	Methodological basis for explainable particle toxicity modeling.

Abbreviations: ATII, alveolar type-II epithelial cell; EBC, exhaled breath condensate; NAO, non-asbestos organic; PM, particulate matter; XAI, explainable artificial intelligence.

 

 

Table 2. Recommended mechanistic endpoint panel for railway brake-wear particle assessment.

Domain	Suggested methods	Main output	Interpretive value
Particle identity	SEM/TEM, EDS, ICP-MS, XRD, surface area, zeta potential	Fe, Cu, Mn, Zn, Ba, Cr, Ni; morphology; agglomeration; soluble fraction	Defines whether toxicity should be interpreted as metal, particle-size, surface or mixture effect.
Acellular oxidative potential	DTT, ascorbate, DCFH-DA cell-free assays	Electron-transfer capacity and antioxidant depletion	Early predictor of redox-active particle behavior.
Cytotoxicity	CCK-8/MTT, LDH, impedance, microscopy	Viability and membrane injury	Necessary but insufficient; low cytotoxicity does not exclude airway-relevant effects.
Oxidative stress	Cellular ROS, GSH/GSSG, MDA, 8-OHdG, HMOX1, NQO1, GCLC	Redox imbalance and antioxidant response	Central mechanism for metal-rich brake and railway PM.
Inflammation	ELISA/qPCR for IL-6, IL-8, TNF-alpha, IL-1beta; NF-kappaB assays	Epithelial and macrophage activation	Links particle exposure to airway inflammatory risk.
Mitochondrial dysfunction	JC-1/TMRE, MitoSOX, ATP, Seahorse, mitophagy markers	MMP loss, mitochondrial ROS, ATP shift and metabolic reprogramming	Connects oxidative stress with energy failure and cell fate.
Airway barrier and mucus	TEER, permeability, MUC5AC, wound-healing assay, ALI cultures	Barrier integrity, translocation, mucus response and repair	Improves relevance beyond submerged monoculture cytotoxicity.
Genotoxicity	Comet assay, micronucleus, gamma-H2AX, 8-OHdG	DNA strand breaks, oxidative DNA damage and repair stress	Important for hazard ranking and long-term risk hypotheses.
Model-ready metadata	Dose, time, particle number, deposited dose, batch, source, dispersion protocol	Harmonized feature matrix	Required for reproducible ML/XAI analysis.

ALI, air-liquid interface; CCK-8, cell counting kit-8; DTT, dithiothreitol; EDS, energy dispersive X-ray spectroscopy; ICP-MS, inductively coupled plasma mass spectrometry; MDA, malondialdehyde; MMP, mitochondrial membrane potential; SEM, scanning electron microscopy; TEM, transmission electron microscopy; TEER, transepithelial electrical resistance; XRD, X-ray diffraction.

 

 

Table 3. Reporting checklist for explainable deep learning in particle toxicology.

Item	Minimum requirement	Reason
Problem definition	State whether the model predicts toxicity class, composite score, oxidative potential or endpoint-specific response.	Prevents vague “AI-assisted” claims.
Feature provenance	Report source, unit, assay method, transformation and missing-value rule for every input feature.	Allows toxicological and computational audit.
Data splitting	Use nested cross-validation and, where possible, external validation using a different particle source.	Reduces overfitting in small particle datasets.
Dose metrics	Compare mass, surface area, particle number and estimated deposited dose.	Avoids misleading toxicity rankings.
Baseline models	Compare deep learning with logistic regression, random forest, XGBoost or support vector models.	Shows whether deep learning is justified.
Performance metrics	Report accuracy, balanced accuracy, precision, recall, F1, AUROC and calibration.	Prevents cherry-picking a single metric.
Explainability	Use SHAP, partial dependence, accumulated local effects and interaction analyses.	Links predictions to metals, oxidative potential and biological mechanisms.
Biological plausibility	Check whether top-ranked features match known mechanisms or experimental perturbation results.	Distinguishes mechanistic insight from batch artifacts.
Uncertainty	Report confidence intervals, bootstrapping or Bayesian/ensemble uncertainty.	Supports transparent risk prioritization.
Reproducibility	Share code, metadata and non-sensitive data or state access conditions.	Improves credibility for toxicology journals.

AUROC, area under the receiver operating characteristic curve; SHAP, SHapley Additive exPlanations.

Table 4. Main evidence gaps and practical future directions.

Evidence gap	Recommended action	Expected contribution
Railway-specific particle toxicology remains limited.	Use source-resolved railway brake, wheel–rail and station particles in standardized assays.	Clarifies whether railway particles behave differently from automotive brake particles.
Mass-based dose may be insufficient.	Report mass, surface area, particle number, soluble metals and estimated deposited dose.	Improves cross-study comparability and hazard ranking.
Mitochondrial outcomes are underreported.	Add MMP, ATP, mitochondrial ROS and metabolic-flux endpoints.	Connects redox stress to energy metabolism and cell fate.
Human relevance needs strengthening.	Use ALI cultures, primary airway cells and epithelial–macrophage co-cultures.	Better represents inhalation exposure and airway barrier biology.
AI models lack external validation.	Validate models on particles from independent railway systems or test benches.	Prevents site-specific overfitting.
Mitigation is rarely linked with toxicology.	Compare particle toxicity before and after material substitution, ventilation or filtration interventions.	Converts mechanistic evidence into practical railway-health strategy.

 

 

Fig. 1. Literature selection and evidence synthesis workflow.

Fig. 2. Source-to-effect framework for railway-derived metal-rich particles.

Fig. 3. Explainable deep learning framework for particle toxicity assessment.

Fig. 4. Author-curated evidence map of cited recent literature.