Sources of particulate-matter air pollution and its oxidative potential in Europe

Particulate matter is a component of ambient air pollution that has been linked to millions of annual premature deaths globally1–3. Assessments of the chronic and acute effects of particulate matter on human health tend to be based on mass concentration, with particle size and composition also thought to play a part4. Oxidative potential has been suggested to be one of the many possible drivers of the acute health effects of particulate matter, but the link remains uncertain5–8. Studies investigating the particulate-matter components that manifest an oxidative activity have yielded conflicting results7. In consequence, there is still much to be learned about the sources of particulate matter that may control the oxidative potential concentration7. Here we use field observations and air-quality modelling to quantify the major primary and secondary sources of particulate matter and of oxidative potential in Europe. We find that secondary inorganic components, crustal material and secondary biogenic organic aerosols control the mass concentration of particulate matter. By contrast, oxidative potential concentration is associated mostly with anthropogenic sources, in particular with fine-mode secondary organic aerosols largely from residential biomass burning and coarse-mode metals from vehicular non-exhaust emissions. Our results suggest that mitigation strategies aimed at reducing the mass concentrations of particulate matter alone may not reduce the oxidative potential concentration. If the oxidative potential can be linked to major health impacts, it may be more effective to control specific sources of particulate matter rather than overall particulate mass. Observations and air-quality modelling reveal that the sources of particulate matter and oxidative potential in Europe are different, implying that reducing mass concentrations of particulate matter alone may not reduce oxidative potential.

Particulate matter is a component of ambient air pollution that has been linked to millions of annual premature deaths globally [1][2][3] . Assessments of the chronic and acute effects of particulate matter on human health tend to be based on mass concentration, with particle size and composition also thought to play a part 4 . Oxidative potential has been suggested to be one of the many possible drivers of the acute health effects of particulate matter, but the link remains uncertain [5][6][7][8] . Studies investigating the particulate-matter components that manifest an oxidative activity have yielded conflicting results 7 . In consequence, there is still much to be learned about the sources of particulate matter that may control the oxidative potential concentration 7 . Here we use field observations and air-quality modelling to quantify the major primary and secondary sources of particulate matter and of oxidative potential in Europe. We find that secondary inorganic components, crustal material and secondary biogenic organic aerosols control the mass concentration of particulate matter. By contrast, oxidative potential concentration is associated mostly with anthropogenic sources, in particular with fine-mode secondary organic aerosols largely from residential biomass burning and coarse-mode metals from vehicular non-exhaust emissions. Our results suggest that mitigation strategies aimed at reducing the mass concentrations of particulate matter alone may not reduce the oxidative potential concentration. If the oxidative potential can be linked to major health impacts, it may be more effective to control specific sources of particulate matter rather than overall particulate mass.
Poor air quality associated with high levels of particulate matter (PM) is one of the five greatest health risks worldwide, alongside high blood pressure, smoking, diabetes and obesity 3,9 . Multiple factors are involved in the pathogenesis of non-communicable diseases resulting from exposure to PM. This makes causal links between health effects caused by PM and PM exposure and properties (composition, size) challenging to establish, especially for chronic effects of long-term exposures to low PM concentrations. One of the many pathways by which PM induces acute health effects is by causing an oxidant/anti-oxidant imbalance in the human respiratory system, which induces a cascade of inflammatory processes that increase the risks of cardiovascular and pulmonary diseases 10,11 . Accordingly, some epidemiological research has suggested that the oxidative potential (OP) of PM, which can be measured using acellular assays, would be a relevant metric for specific acute (not chronic) PM health effects 7,8 . As a result, considerable field and laboratory measurements have been devoted to the identification of PM components manifesting an oxidative activity, and the OP of these components is often related to the PM mass with which they are associated, OP m (where OP m is the OP activity per mass of the aerosol component, in units of nmol min −1 µg −1 ) (refs. 7,12 ). Focusing on different PM fractions, these studies have shown conflicting results 7 , some pointing towards transition metals, for example, from primary vehicular emissions, as important drivers and others towards secondary organic aerosols (SOA). Often transition metals are not related to specific sources, nor are SOA precursors identified. Consequently, which emission sources of primary PM and SOA precursors control the OP concentrations (that is, OP v , which is the OP activity per volume of air of the aerosol component, in units of nmol min −1 m −3 ) and the exposure to OP of the population on a large scale remain unknown at present.
Here, we explore the sources of PM and OP across Europe, by combining field observations and air-quality modelling, and find that they are controlled by different sources. We use field observations to derive the OP m parameters for the major European PM sources. For this, we have determined the chemical composition, sources and OP content of all PM components (elements, and primary and secondary organic aerosol), using an extensive number of samples at sites where the major aerosol sources in Europe are present on a yearly basis (see Methods). Samples cover two size fractions: PM 2.5 and PM 10 , referring to particles smaller than 2.5 µm and 10 µm, respectively. The OP m parameters obtained were combined with a modified air-quality model to determine the main sources contributing to PM and OP exposure in Europe, where exposure refers to the amount of PM or OP v in inhaled ambient air accumulated over a full year and integrated over the population.
We examined the contributions of different constituents to PM and their regional variability. PM mass concentration is dominated by regional secondary inorganic aerosol (SIA = NH 4 + +NO 3 − +SO 4 2− , for PM 10 46 ± 13% and for PM 2.5 56 ± 6%) (Fig. 1k), coming from the agriculture, transport and energy sectors. We quantified the sources of organic aerosol (OA) using a combination of offline aerosol mass spectrometry and positive matrix factorization 13,14 . In summer, OA is dominated by fine-mode secondary OA from biogenic precursors (bioSOA). The variation in bioSOA concentrations is consistent with (1) the exponential      15 . Unlike SOA, primary OA (POA) is more local (Supplementary Fig. 7) and is dominated by emissions from biomass burning during winter (BBOA), vehicles (HOA) and cooking (COA). Sulphur-containing OA (SCOA) probably derives from non-exhaust vehicular emissions, for example, tyre wear, because it correlates strongly with vehicular emissions (for example, NO x ), is composed of fossil carbon, and contributes mainly to the coarse mode of OA 14,16 . Finally, vehicular wear and crustal emissions dominate the coarse-mode metals, whereas in the fine mode, metals originate from residential heating (Supplementary Figs. 9,10).
Since the particles that dominate PM mass concentration, with diameters of 0.1-5 µm, have a relatively constant probability of being deposited in the tracheobronchial region over this size range (Supplementary Fig. 28), both coarse PM and PM 2.5 should be considered when estimating the health effects of PM 17 , as also suggested by the REVIHAAP project 18 . Therefore, we measured the OP v related to PM 2.5 and PM 10 using three different acellular assays (dithiothreitol (DTT), ascorbic acid (AA) and dichlorofluorescin (DCFH); see Methods section 'OP of PM'). Despite their variable sensitivity towards different aerosol constituents 19-21 , DTT v and AA v correlate well with each other (correlation coefficient R 2 = 0.68) and to a lesser extent with DCFH v (R 2 (DTT v , DCFH v ) = 0.35, R 2 (AA v , DCFH v ) = 0.45). We confirm for our case that exposing re-differentiated human bronchial epithelia to PM extracts with increasing DTT activity (16 biological experiments) enhances the release of the pro-inflammatory cytokine interleukin-6 (IL-6, Extended Data Fig. 2), triggering inflammation. Although the relevance of OP v for the acute health risks of PM is still debatable 7,8 , our results are in agreement with studies that suggest such links. Therefore, in the following we use OP v as a representation of the acute health impacts of PM via pathological oxidative stress.
We quantified the OP m (DTT m , AA m , DCFH m ) of the various OA and metal components based on a multiple linear regression model. We assessed the uncertainty of OP m via Monte Carlo simulations (Fig. 1i  response ( Supplementary Fig. 12), may facilitate metal dissolution, and thus inducing an indirect influence on OP 22 .
We find that the OP based on the three different acellular assays responds to the same set of components ( Fig. 1 3,4). We demonstrate that anthropogenic emissions have a higher oxidative potential per mass unit of the aerosol component (OP m ) than biogenic emissions and crustal material. In the coarse mode, higher OP m is found for non-exhaust vehicular emissions than for crustal material, which is in agreement with previous studies showing that transition metals have a large oxidative capacity 7 . The OP m of aSOA, mainly from ageing of residential biomass burning emissions 15,16 , is at least three times higher than the OP m of bioSOA. This is in general agreement with laboratory experiments showing that SOA from single aromatic precursors have higher OP m than biogenic terpene and isoprene SOA 7 . We note that potent reactive oxygen species, for example, quinones from the ageing of complex mixtures of polyaromatic hydrocarbons present in biomass smoke (Extended Data Fig. 1, Supplementary Fig. 6) 23 , are enhanced during winter when aSOA dominates. Such species are often not accessible by ageing single precursors.
This can explain the lower OP m of laboratory SOA compared to ambient aerosol 7 (challenges in comparing OP m between studies are described in Methods section 'Discussion of OP m results'). Finally, we find different types of fine anthropogenic combustion POA (HOA, BBOA) to have a substantial OP m , which agrees with earlier work 7 . Independent of the assay used, OP v in the fine mode is largely explained by aSOA, with minor contributions from combustion derived POA and bioSOA ( Fig. 1, Extended Data Fig. 3, 4). By contrast, OP v in the coarse mode is mostly attributed to non-exhaust vehicular emissions (vehicular wear and SCOA), especially dominant at urban sites. These findings are in contrast with the source contributions to PM mass concentrations ( Fig. 1), strongly affected by SIA. They also imply that reducing specific anthropogenic emissions will result in a considerable reduction of OP v .
We combined the OP m of the different sources with a modified air-quality model to predict the major sources of OP v and PM mass concentrations in Europe (Fig. 2, Extended Data Fig. 5). The model reliably reproduces non-exhaust vehicular emissions, SIA, POA and SOA concentrations (Methods section 'Air-quality model') 24 . Modelled OP v shows a good agreement with measured OP v also for sites and   Article periods that were not included in the training dataset (Extended Data Fig. 6, Supplementary Fig. 26). This signifies that the model accurately represents the PM mass concentrations from various sources and supports the OP m parameters derived from ambient observations. In Europe, OP v and PM 10 concentrations have hotspots in urban environments (for example, Paris and the Po valley), although PM 10 concentrations are more regional than OP v (Fig. 2a, b). Despite this, OP v is dominated by different sources than is PM mass. PM mass is governed by regionally formed SIA in northern Europe and long-range transported crustal material in southern Europe (Fig. 2c, Extended Data Fig. 7). In contrast, DTT PM10 v is dominated by anthropogenic SOA (largely from residential biomass burning) in large areas in Europe and vehicular emissions (non-exhaust and exhaust) in densely populated urban areas (Fig. 2d, Extended Data Fig. 7). As a result, PM mass concentrations are not only higher in urban areas (by a factor of about 7 compared to the area with the lowest population density), but urban PM has a threefold-higher OP per unit of PM mass concentration than rural PM (Fig. 3b, for PM 2.5 see Extended Data Fig. 8).
By scaling OP v and PM mass concentrations with the population density, we estimate the contributions of different sources to OP and PM exposures. We find that modelled PM exposure (Fig. 3a, Extended Data Fig. 8) is governed by SIA (31%), crustal material (27%) and bioSOA (6%). The high contribution of SIA to PM exposure is consistent with the hypothesis that agricultural emissions drive PM-related premature mortality, as formulated by Lelieveld et al. 2 . By contrast, modelled OP exposure is overwhelmingly driven by anthropogenic sources (Fig. 3a), from non-exhaust vehicular emissions (15%-49%) in the coarse mode and aSOA (24%-51%) in the fine mode. It is worthwhile to note that aSOA could gain importance compared to PM from vehicular wear if aSOA, because of its size, is deposited in more vulnerable parts of the respiratory system. Observed differences between the sources of PM and OP exposures are much larger than the differences between the type of OP assay used or the uncertainties related to positive matrix factorization and the multiple linear regression model analyses (Fig. 3a, Supplementary Table 2). In densely populated regions, non-exhaust vehicular emissions dominate OP exposure, consistent with the increase in cardiovascular disease risk associated with highway proximity 25 . However, a large fraction of the population resides in less populated areas where the largest fraction of OP v is ascribed to aSOA (Fig. 3b). As PM and OP exposures are governed by different sources, our finding could imply that acting on the dominant PM sources might not result in the anticipated reduction of the acute health effects of PM.
PM mass concentrations have substantially decreased in Europe 26 over the past decades ( Supplementary Fig. 1), owing to efficient mitigation strategies that continuously reduced SIA precursors (SO 2 , NO x ) and vehicular exhaust emissions 27 (Fig. 3c). However, the successful reduction of PM mass concentrations does not necessarily reflect a commensurate reduction in OP v . In fact, the main sources of OP v , including residential biomass burning (the main emitter of aSOA precursors) and non-exhaust vehicular emissions, have roughly remained constant over the past decades. With the increasing future demand for renewable energy for domestic heating, the eco-design directive (for example, the National Emission Ceilings (NEC) Directive 2016/2284 28 , see Methods section 'Historical and projected emissions') has been recently implemented. Such implementation is expected to halve the residential emissions by 2030 (Fig. 3c, Supplementary Fig. 30), but only if cleaner combustion technologies are adopted. Otherwise, these emissions will remain stable for the next decade 29 . Although considerable efforts are still devoted to further reducing direct exhaust emissions, non-exhaust vehicular emissions are projected to slightly increase ( Fig. 3c) with a future rise in vehicle numbers (for example, passenger car traffic volume increased by 27% in the 28 European Union member countries from 2010 to 2015 30,31 ). With the phase-out of older vehicles with high PM exhaust emissions, non-exhaust emissions are expected to constitute 80%-90% of the direct road transport PM 10 after 2020 ( Supplementary   Fig. 29) 27 , unless additional measures are taken 17 (for example, reducing the copper content in brake pads, or a better sealing of the brakes). This could imply that targeting specific PM sources rather than overall PM mass might become more important for public health.
The aerosol sources examined here are widespread and can affect human health in other parts of the world. Further research is required to address other PM sources such as industrial metal emissions, heavy fuel oil combustion or coal emissions. The latter is a major PM source in highly populated urban environments in Asia, where most of the global PM-induced premature mortality is expected to occur. Our results show that different sources drive OP v and PM mass concentrations, suggesting that knowing the sources of PM alone and taking action to reduce them is not sufficient to reduce the OP v effectively. Although some studies have related OP v to acute health effects such as acute cardiovascular and respiratory diseases, the main properties of PM that drive its chronic health effects, a leading cause of premature deaths, are yet to be determined and the role of OP v in these effects remains undefined. If OP is found to be related to major health impacts this could imply that controlling its specific sources might be more effective. The framework developed here can be used to identify the sources responsible for the chronic health impacts of PM (for example, using long-term time series of PM components from different sources).

Online content
Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41586-020-2902-8. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Measurement sites
We collected 24-h integrated PM 10

Analyses
Aerosol composition and gas-phase pollutants. We determined the bulk composition of OA with an aerosol mass spectrometer using filter samples following the procedures described in ref. 13 . This approach is composed of water-extracting the PM (milli-Q water, 18.2 MΩ cm, 30 °C, 20 min), followed by vortexing (60 s) and subsequent filtration using nylon membrane filters (pore size 0.45 µm, 8813Y-N-4, Infochroma AG). The resulting aqueous solutions were nebulized (custom designed APEX nebulizer, Elemental Scientific), dried (Nafion dryer) and measured by an Aerodyne high-resolution time-of-flight aerosol mass spectrometer, where sample analyses were alternating with analyses of milli-Q water and pure NH 4 NO 3 solutions. The detailed working principles of the high-resolution time-of-flight aerosol mass spectrometer are summarized in Canagaratna et al. 32 . Data corrections were applied as outlined in Allan et al. 33 and Daellenbach et al. regarding the offline application of the aerosol mass spectrometer 13,14 . Field blanks were handled in the same way as exposed samples and showed a negligible signal. We analysed 15 trace elements (Al, Fe, Ti, As, Cd, Cu, Mn, Mo, Ni, Pb, Rb, Sb, Se, V, Zn) by inductively coupled plasma mass spectrometry after acid digestion. Briefly, each sub-sample was acid-digested using HNO 3 , and H 2 O 2 in a microwave oven. Analysis took place with an inductively coupled plasma mass spectrometer (ELAN 6100 DRC II, Perkin Elmer) 34 .
We Additionally, eight oxygenated-polyaromatic hydrocarbons (quinones) were measured: 1,4-naphthoquinone, 1,2-naphthoquinone, acenaphthenequinone, 9,10-anthraquinone, 1,4-anthraquinone, 2-methylanthraquinone, aceanthrenequinone and benz[a] anthracene-7,12-dione. The analyses were conducted using ultra-performance liquid chromatography/fluorescence (Thermo-Scientific, Dionex Ultimate 3000) and gas chromatography coupled to mass spectrometry with negative-ion chemical ionization (Agilent 7890A gas chromatograph coupled to 5975C mass spectrometer) 35-37 following a quick, easy, cheap, effective, rugged and safe (QuEChERS)-like extraction procedure of filter punches with acetonitrile as solvent 35,36 . Quality was assessed following the CEN (European Committee for Standardization) standard procedures EN 15549 and TS 16645 (2008 and 2014). In addition, the institute where the analyses were conducted (Institut National de l'Environnement Industriel et des Risques; Ineris) participates, every two years, in national and European polyaromatic hydrocarbon analytical inter-comparison exercises. The last exercise showed results in good agreement with reference values including the QuEChERS extraction 38 . Quinone extraction efficiencies were checked using NIST standard reference material (SRM 1649b, urban dust; https://www-s.nist.gov/srmors/view_cert. cfm?srm=1649B). Results obtained were in good agreement with NIST reference and indicative concentration values and with those previously reported in the literature 36 .
We quantified anhydrous sugars (levoglucosan, galactosan and mannosan), sugar alcohols (xylitol, sorbitol, arabitol and mannitol) and glucose using high-performance liquid chromatography with the pulsed amperometric detector method 34,39 . Briefly, the water extract of a fraction of each sample was analysed on an ICS 5000+ high-performance liquid chromatograph with pulsed amperometric detection (Thermo-Fisher, equipped with a 4-mm-diameter Metrosep Carb 2 × 150-mm column and a 50-mm pre-column). The analytical run was isocratic with 15% of an eluent of sodium hydroxide (200 mM) and sodium acetate (4 mM) and 85% water, at 1 ml min −1 .
We quantified organic acids (glutaric, phthalic, methylglutaric, malonic, malic, pinic acids, and 3-3-methyl-1,2,3-butanetricarboxylic acid (MBTCA)) by liquid chromatography with mass spectrometry equipped with an electrospray ionization source (LCQ-Fleet, Thermo Fisher), with chromatographic separation performed on a Synergi 4-µm Fusion RP 80A column (Phenomenex), with a gradient of wateracetonitrile-formic acid eluent. The same water extract as for sugars and polyols was used for this analysis. Pinic acid and MBTCA are known oxidation products of biogenic terpenes 40 . Other organic acids are derived from anthropogenically dominated volatile organic compound emissions (for example, phthalic acid from naphthalene) or influenced by both biogenic and anthropogenic sources 40 .
The gas-phase pollutants NO x (chemiluminescence, Horiba APNA 360), SO 2 (ultraviolet fluorescence, CEN Norm EN 14212), CO (infrared absorption, CEN Norm 14212) and O 3 (ultraviolet absorption, Thermo Environmental Instruments 49C, Thermo Electron Corp.) and PM 10 (gravimetric) were additionally monitored. The long-term variability of PM 10 is displayed in Supplementary Fig. 1. Filters were extracted in a simulated lung fluid (here Gamble solution + dipalmitoylphosphateidylcholine) at 37 °C simulating the epithelial lung lining (extracellular environment) in the gas exchange region 19 during 1.25 h to take bioaccessibility with preferential solubility into account, and at iso-concentration of 25 µg ml −1 to get comparable results among different sites. PM extracts were not filtered after the extraction, and were submitted to the three OP assays in parallel. While DTT and AA are sensitive to redox-active PM constituents, DCFH reacts directly with oxidants in PM 7,47,48 . OP has been related to acute health outcomes (mostly from the DTT assay) and several studies have found a stronger association between OP and acute health effects than with PM 7 .
The DTT assay is based on the ability of PM to transfer electrons from DTT to oxygen, generating a superoxide anion. We chose to use it here because it is the most frequently used approach owing to its relatively balanced response towards inorganic and organic atmospheric aerosol constituents 19,20,49 . The measurements were conducted according to the protocols in Calas et al. 45 , adapted from Cho et al. 50 .
The AA assay is based on the depletion of this natural lung antioxidant when in contact with PM. This assay can be used to quantify the transition-metal-based redox activities but has also shown to be sensitive to (organic) biomass burning tracers 21, 45 . The AA assay is modified from Mudway et al. 51 and is fully described in Calas et al. 45 The DCFH assay relies on a non-fluorescent probe, which is oxidized to a fluorescent product in the presence of reactive oxygen species and horseradish peroxidase. This assay is usually used in biology to detect reactive oxygen species at cellular level and has been adapted for acellular assays 46 . This probe is known to be non-specific to oxidant species 51,52 , but also to reactive nitrogen species. The procedure relies on Foucaud et al. 46 . Analysis of biological response. The inflammatory response of normal human bronchial epithelial cells to exposure of PM from the urban roadside (Bern-Bollwerk) and the rural valley site (Magadino-Cadenazzo) was assessed in different seasons. For this purpose, the PM filter samples (exposed diameter of 14.7 cm) were pooled to form seasonal composites (each sample represented by a circular punch of 10 mm diameter, 0.8 cm 2 ), that is, January to March (Bern winter 2013 PM 10 10 ). Immediately before the cellular exposure, the composites were extracted in 7 ml milli-Q water (milli-Q water, 18.2 MΩ cm, 30 °C), homogenized by vortexing (1 min) and filtered through a 0.45-µm nylon membrane syringe filter (Infochroma AG, Zug, Switzerland). Field blanks were extracted with the same procedure using an equivalent filter surface area as for the samples.
Human bronchial epithelial cells were isolated from human donor lungs. Two normal lung samples deemed not suitable for transplantation were obtained from the Life Alliance Organ Recovery Agency of the University of Miami, USA. Institutional-Review-Board-approved consent for research with these tissues was obtained by the Life Alliance Organ Recovery Agency and conformed to the Declaration of Helsinki. Air-liquid interface cultures of re-differentiated human bronchial epithelial cells were generated as previously described 53 . Cell cultures were exposed to PM extracts for 4 h (7.4-22.2 µg per cm 2 of cell culture area and 1:10 diluted resulting in 0.7-2.2 µg per cm 2 of cell culture). Control cell cultures were either exposed to extracts from field blanks or left untreated in the incubator. Repeatability was assessed by performing 3-9 experiments for each seasonal composite. The release of the inflammatory mediator IL-6 was measured in the basolateral compartment collected at 24 h after exposure to filter extracts, using the Bio-Plex multiplex bead-based suspension array system and the appropriate detection kit (Bio-Rad Laboratories AG) according to the manufacturer's protocol.

Source apportionment of OA and trace elements
Positive matrix factorization. We performed source apportionment of the OA and of trace metals using positive matrix factorization, which is a statistical unmixing model 54 . Positive matrix factorization explains the variability in a chosen dataset (x i,j ) with linear combinations of constant factor profiles (f k,j ) representing the chemical composition of the sources/components) and their contribution varying in time and/or space (g i,k ). Since positive matrix factorization allows only for positive results (g i,k , f k,j ), it is ideal for environmental purposes. The residual is termed e i,j . The index i represents a measurement (time and space), j an observed variable (trace metal, ion in aerosol mass spectrometer), and k is a factor.
This equation is solved by Multilinear Engine 2 (ref. 55 ) and controlled by the software Source Finder (SoFi 56 ). Sources of OA. Multilinear Engine 2 allows using a priori information by constraining known factor profiles g i,k within a certain range defined by the scalar a value ( a 0 ≤ ≤1).
For the source apportionment of the OA in PM 10 at the nine sites in Switzerland and Liechtenstein during 2013 (see Daellenbach et al. 14 ), we used reference spectra representing HOA and COA (Crippa et al. 57 ). The sources' contributions to water-soluble OA were subsequently corrected to contributions to OA using source-specific recoveries derived in Daellenbach et al. 13 and confirmed in other studies 16, 58 . The evaluation, sensitivity tests (a-value, extreme bootstrap scenarios, source-specific recoveries) and uncertainty estimates of the organic source apportionment results were presented in Daellenbach et al. 14 .
For Magadino (2013 and 2014), an analogous separate size-resolved (PM 2.5 and PM 10 ) source apportionment analysis (HOA constrained from Crippa et al. 57 ) was performed and presented in Vlachou et al. 16 . The time series of the resolved OA source contributions are compared to other chemical analyses in Supplementary Fig. 4. aSOA time series are compared to phthalic acid and quinone concentrations (Extended Data Fig. 1, Supplementary Fig. 5) and bioSOA to MBTCA concentrations (Extended Data Fig. 1, Supplementary Fig. 5). In Extended Data Fig. 1 (Supplementary Fig. 6), the relative contribution of quinones to OA is presented. Sources of trace metals. We performed unconstrained (no a priori knowledge) positive matrix factorization using trace metal data (Al, As, Cd, Cu, Fe, Mn, Mo, Ni, Pb, Rb, Sb, Se, Ti, V and Zn). We used PM 10  The data matrix x i,j consists of the concentrations determined for the mentioned trace metals (if the concentration was smaller than the half the quantification limit, QL/2). The error matrix s i,j was computed using the quantification limit and a term depending on the concentration. . Q/Q exp was small (around 1) when employing only one or two factors, which could indicate that the uncertainty s i,j was overestimated ( Supplementary Fig. 8).
A clear seasonality in the uncertainty-weighted residuals e i,j /s i,j was observed for the two-factor solution (higher in winter than summer; Supplementary Fig. 8). When employing three factors, there was no longer any apparent seasonality in e i,j /s i,j and Pb and Zn were better explained by the factorization (Supplementary Fig. 8). The three-factor separation is displayed in Supplementary Figs. 9 and 10, yielding: (1) a component perceived mostly in the coarse mode of PM explaining Al, Se and Ti contents and thus identified as crustal material.
(2) another component also perceived mostly in the coarse mode of PM, but explaining Cu, Fe, Mo and Sb contents, which-in contrast to crustal material-were also strongly enhanced in a road traffic tunnel. This component was therefore related to vehicular wear 59 .
(3) a third component explaining Cd, Pb, Rb and Zn contents. This component was perceived mostly in PM 2.5 and showed a temporal behaviour similar to combustion-related components (enhanced in winter compared to summer). Thus, this component was identified as emissions from residential heating.
We performed 100 bootstrap runs (including/excluding samples and initializing model information randomly to assess the rotational ambiguity) to obtain robust source contributions and to estimate the uncertainty related to fingerprints (chemical composition) and time series. The chemical fingerprints (and uncertainties) of the separated sources are presented in Supplementary Fig. 9 and the concentration time series in Supplementary Fig. 10 (uncertainty in Supplementary Fig. 11). While vehicular wear and crustal material had a relatively small uncertainty, residential heating was found to be more uncertain ( Supplementary  Fig. 11). The time series of the resolved metal source contributions is compared to the other chemical analyses in Supplementary Fig. 4. Size distribution of source concentrations. Five of six OA sources were mostly present in PM 2.5 , whereas SCOA represented coarse fossil organic material related to traffic activity (NO x ) 14

Sources of OP
Multiple linear regression model. To explain the spatial and temporal variability of OP v , both land-use regression and source impact regression models were used in previous studies (here we cite only a selection) 7,12,[60][61][62][63][64] . In this study, we used a source impact regression model to derive OP m and source contributions to OP v . OP v was apportioned to its sources using a multiple linear regression model as in equation (4). We used 109 samples: 24 from an urban roadside site (Bern, Bollwerk), 36 from an urban background site (Zurich, Kaserne), 24 from a rural background site (Payerne, MeteoSuisse), 24 from a rural background site in an alpine valley affected by wood burning (Magadino-Cadenazzo), and one from a wintertime wood-burning pollution episode in an alpine valley (San Vittore). Possible candidate predictors considered were trace-metal sources (crustal material, vehicular wear, residential heating) and OA sources (HOA, COA, BBOA, SCOA, aSOA, bioSOA). Whereas the majority of OA sources were mostly present in PM 2.5 , SCOA represented coarse fossil organic material (see Methods section 'Size distribution') related to traffic activity (NO x ) 14,16 . We used the known size distribution (PM 2.5 versus PM 10 ) for estimating the OA sources in PM 2.5 based on the source concentrations in PM 10 . For the metal sources, we used the directly measured distribution from the source apportionment analysis. NH 4 NO 3 and (NH 4 ) 2 SO 4 were excluded, since-in agreement with literature 45,47 -these components did not induce OP ( Supplementary Fig. 12). Elemental carbon was excluded because it is influenced by several combustion sources (traffic, biomass burning).
For DTT, a model with five predictors had the lowest AIC (Supplementary Fig. 13). Whereas models with six predictors had a comparable AIC, the Bayesian information criterion (a stronger penalty term for including more predictors than AIC) suggested that the five-predictor models were clearly favourable. Of the five-predictor models, there were two of similar quality (vehicular wear, HOA, BBOA, aSOA and either bioSOA or COA). Since PM 2.5 DTT v as well as PM 10 DTT v showed an increase in summer, suggesting the importance of bioSOA, we chose the model with bioSOA. Using AIC as the decisive metric, we also chose for DCFH a five-predictor model with vehicular wear, SCOA, BBOA, aSOA, and bioSOA ( Supplementary Fig. 14). For AA, a 3-predictor model using vehicular wear, BBOA, and aSOA was chosen, since including further predictors did not lead to further improvement using the AIC as metric (Supplementary Fig. 15).
The multiple linear regression model uncertainty and the sensitivity of OP m (DTT m , AA m , DCFH m ) to errors in OP v measurements and PM source apportionment were assessed via Monte Carlo simulations (50,000 runs). In this process, we combined different randomly selected OA and metal source apportionment solutions, and additionally performed bootstrap analysis on the time points (cumulative density functions of OP m (Fig. 1, Extended Data Figs. 3, 4, Supplementary Table 2). Discussion of OP m results. Transition metals have been shown to have a substantial impact on the OP activity of PM 20,21, [65][66][67][68][69] . In this study, we find that vehicular wear induces a strong impact on OP. Overall, this is consistent with findings from London suggesting that coarse traffic emissions have a higher OP m than PM from other sources 64 .
Aqueous solutions of NH 4 NO 3 (0.08-40.02 µg ml −1 ) and (NH 4 ) 2 SO 4 (0.13-66.07 µg ml −1 ) were analysed with the DTT and AA assays and did not exhibit reactivity ( Supplementary Fig. 12, analyses of ambient PM conducted at 25 µg ml −1 ). This is in agreement with literature 45 and previously reported results for DCFH 47 . We note that we did not introduce any interaction term into the multiple linear regression model, although studies show that SIA, while not directly inducing an OP response, can facilitate the metal dissolution, having an indirect influence on OP 22 . We observe such nonlinear matrix effects when using OP m parameters derived from the multiple linear regression model in a road tunnel where vehicular wear is dominant: OP v from vehicular wear alone overestimated the measured OP v (DTT v : 3×, AA v : 2×, DCFH v : 3×) as SIA is minimal and thus metals are not soluble. However, the multiple linear regression model explains the OP v measurements well (Fig. 1,  Extended Data Figs. 3, 4) and thus such effects apparently have a minor role under ambient conditions.
Earlier studies showed that both POA and SOA have a substantial impact on OP v (refs. 7,62,[69][70][71][72][73][74][75][76][77]. We find that aSOA has a higher OP m than bioSOA. This is in agreement with literature showing that the SOA precursor type has a strong impact on OP m . SOA from anthropogenic aromatic precursors (for example, naphthalene and phenanthrene) has a higher OP m than SOA from biogenic precursors (for example, α-pinene, isoprene and limonene) 72,73,78,79 . This is consistent with previous research pointing out the important role of quinones formed from aromatic compounds on OP v (refs. 50,80,81 ). Further, also different types of fine anthropogenic combustion POA (HOA and BBOA) were found to have a considerable OP m , which is in agreement with earlier work 70,71,82 .
Quantitative comparisons of OP m,PM (here reported in units of nmol min −1 per µg of PM 2.5 ; but elsewhere in this paper OP m is consistently reported in units of nmol min -1 per µg of source mass) between measurement techniques remain challenging because of differences in measurement techniques as stated by Bates et al. 7 . Nevertheless, we compare the measured OP m,PM value that we find for PM 2.5 (as is usually reported in the literature) to the results in this study. OP m,PM2.5 values in this study (an annual average for measurement sites between 60 and 140 nmol min −1 per µg of PM 2.5 ) are at the upper end of ambient OP m,PM values previously reported (between 10 and 80 nmol min −1 per µg of PM 2.5 ) 12 .

Air-quality model
Modelling PM contribution in Europe. We used the regional chemical transport model Comprehensive Air Quality Model with extensions (CAMx) version 6.3 (http://www.camx.com) to simulate air quality in Europe in 2011 ( Supplementary Fig. 16), with a spatial resolution of 0.25° × 0.125° and 14 terrain-following vertical layers, with the first layer being about 20 m thick. The meteorological parameters were calculated using the Weather Research and Forecasting Model WRF-ARW version 3.7.1 (ref. 104 ) The Carbon Bond 6 Revision 2 (CB6r2) mechanism 83 was used for gas-phase chemistry. Partitioning of inorganic aerosol components (sulphate, nitrate, ammonium, sodium and chloride) between the gas and particle phases was calculated by the ISORROPIA thermodynamic model 84 . OA formation was estimated by the 1.5-dimensional volatility basis set OA chemistry/partitioning module 85 . The volatility basis set module of the model was modified to separate the contributions of specific OA sources including biomass burning, gasoline and diesel vehicles, other anthropogenic activities and biogenic sources 86,87 . The parametrization of biomass burning and vehicles was updated based on Ciarelli et al. 86 and Platt et al. 88 , respectively. The source-specific anthropogenic emissions were obtained from the high-resolution European emission inventory TNO-MACC-III (The Netherlands Organization for Applied Scientific Research-Monitoring Atmospheric Composition and Climate) 89 . Biogenic emissions were prepared using the PSI model 87 Fig. 16). The species related to non-tailpipe emissions of traffic such as metals from vehicular wear and SCOA were not explicitly represented in the model. In addition, the emission information on these components is highly uncertain and emission factors vary largely from country to country. As the emissions of vehicular wear and SCOA originate from the road traffic and show high correlation with the traffic-NO x emissions, we estimated the levels of vehicular wear and SCOA as functions of NO x from vehicles ( Supplementary  Fig. 17). The size distributions of vehicular wear and SCOA were computed based on a parametrization derived from the measurements ( Supplementary Fig. 18). As the model simulates total NO x from all sources, we calculated the NO x from road traffic by multiplying the total NO x with the grid-scale ratio of NO x emissions from road traffic to total NO x emissions ( Supplementary Fig. 19). The estimated copper concentration from vehicular wear (copper being 4.5% of total vehicular wear as found by metal positive matrix factorization) agreed well (absolute concentrations and spatial distribution) with PM 10 copper concentrations from brake wear modelled independently (analogously to Hulskotte et al. 93 ) (Supplementary Fig. 20). Additionally, total measured copper concentrations from different European sites agreed with the estimated copper concentrations from vehicular wear (data from EBAS; http://ebas.nilu.no/) quite well (Supplementary Figs. 20-22) 94,95 . However, at some sites local point sources (not vehicular wear) result in total measured copper concentrations much higher than the modelled vehicular wear copper concentrations. Modelling OP v contribution in Europe. OP v (DTT v , DCFH v , AA v ) were modelled using the parameterization from the field data. To estimate the spatial distributions of OP v (DTT v , DCFH v and AA v ) in Europe, the median values of OP m of all sources were multiplied with the modelled concentrations of the relevant OA components (HOA, BBOA, bioSOA, aSOA and SCOA) and vehicle wear metal. The modelled DTT v , AA v and DCFH v from the relevant sources are presented in Extended Data Fig. 5 and the most important sources for DTT v , AA v , DCFH v and PM concentrations are displayed in Extended Data Fig. 7.
A Monte Carlo approach with 300 runs for each OP assay was conducted to assess the uncertainties of the estimated OP v . Three sources of uncertainty were included: (1) OP m of each assay; (2) parametrization of SCOA and vehicle wear as a function of road traffic NO x ; and (3) parametrization of the size distributions of SCOA and vehicular wear (PM 2.5 versus PM 10 ). We note that (1) includes both uncertainties related to the positive matrix factorization analysis and the multiple linear regression model parameterization. For each run, a set of OP m combinations was randomly selected among OP m combinations generated by the multiple linear regression model developed in this study. A comparison between all and selected OP m combinations shown in Supplementary Fig. 23 indicates that the distribution of OP m was well represented by the randomly selected values. The factors (2) and (3) were sampled from the normal distributions with the mean value and standard error shown in Supplementary Figs. 17 and 18. Model validation. An accurate representation of PM OP in Europe based on our measurements entails two major requirements: (1) PM sources at our sites must be representative of sources encountered in Europe; and (2) OP m determined from our measurements for the different sources are consistent across Europe and not highly dependent on the emission conditions. Below, we demonstrate that these two requirements are fulfilled in our analysis:(1) The sources contributing to PM found at our measurement sites are identified as the major and most widespread sources throughout Europe, and the PM chemical composition is consistent overall with other European sites 96,97 . The OA components extracted by positive matrix factorization (HOA, BBOA, aSOA and bioSOA) are very frequently reported at other sites and the modelled population-weighted concentrations in Europe compare well with the measurements at the Swiss sites investigated here ( Supplementary Fig. 24). We are aware that we cannot represent coal combustion (though included as a part of biomass burning), which is important in Poland, or local industrial emissions, as pointed out in the main text. Assessing the impact of local sources requires city-level modelling, and coal combustion should be targeted in future research, especially in the context of China. (2) The OP m for a specific source is a property of the chemical composition of the aerosol emitted by this source. The chemical fingerprints of the sources found in this study are similar to those reported at other European sites 96,98 . We compare OP m estimated here for vehicular wear and primary biomass burning with estimates made at two sites in France dominated by these sources (Alpine Valley site for biomass burning and traffic site for vehicular wear, Supplementary Fig. 25) 45 . Results from France are largely in agreement with our results from the multiple linear regression (MLR) at least for these two sources (for DTT, the median values are: OP m (BBOA(MLR)) = 0.08 nmol min −1 per µg of source mass versus OP m (BBOA(France)) = 0.22 nmol min −1 per µg of source mass; OP m (vehicular wear(MLR)) = 3.51 nmol min −1 per µg of source mass versus OP m (vehicular wear(France)) = 1.40 nmol min −1 per µg of source mass). We also compare the modelled OP v to measurements from 2011 that were not part of the training dataset. We use samples from Zurich (Switzerland, 19 samples) and from an additional site in Lens 34 , France (95 samples, Extended Data Fig. 6). All the predicted OP assays show remarkable agreement with the measurements ( Supplementary  Fig. 26

Inhalation exposure of PM and OP
We compute the PM and OP exposure as the population integrated amount in inhaled ambient air accumulated over a full year (Supplementary Fig. 27). In essence, exposure is proportional to the product of OP or PM concentration and population density. The exposure to PM and OP for one individual i in age group j was estimated based on equation (7):  Table 3). PM and OP exposure on the regional scale were then calculated by summing up the exposure of all individuals i in all age groups j.

Multiple-Path Particle Dosimetry Model
We used the Multiple-Path Particle Dosimetry Model (MPPD) 101 to estimate the number of particles deposited in the human body in comparison to the number of inhaled particles, termed deposited fraction (pulmonary and tracheobronchial region) ( Supplementary Fig. 28).  102 . These results are presented in Fig. 3c, and in addition, a more detailed comparison of the historical exhaust and non-exhaust road traffic emissions is presented in Supplementary  Fig. 29.

Data availability
The full dataset shown in the figures and tables is publicly available at https://doi.org/10.5281/zenodo.4048589. Source data are provided with this paper.