Uncovering new global risks from commercial chemicals in air

Commercial chemicals are used extensively across global urban centers, posing a potential exposure risk to 4.2 billion people, which accounts for 55% of the global population. Harmful chemicals are often assessed and regulated based on their environmental persistence, bioaccumulation, and toxic properties, under international and national initiatives such as the Stockholm Convention. However, current regulatory frameworks largely rely upon knowledge of the properties of the parent chemicals, with minimal consideration given to their atmospheric transformation products. This is mainly due to a signicant lack of experimental data, as identifying transformation products in complex mixtures of airborne chemicals is an immense analytical challenge,hence making a comprehensive and reliable risk assessment for harmful chemicals currently unachievable. Here, we develop a novel framework, combining laboratory and eld experiments, non-target analysis techniques, and in-silico modelling, to identify and assess the hazards of airborne chemicals, which takes into account atmospheric chemical reactions. By applying this framework to organophosphate ame retardants, as representative chemicals of emerging concern, we nd that their transformation products are globally distributed across 18 megacities, representing a previously unrecognized exposure risk for the world’s urban populations. Furthermore, the transformation products can be up to an order of magnitude more persistent, environmentally mobile, and toxic than the parent chemicals. The results indicate that the overall human and environmental risks associated with ame retardants could be signicantly underestimated, while highlighting a strong need to include atmospheric transformations in the development of regulations for all harmful chemicals moving forward. mass spectrometer (Waters Xevo G2XS). Sample extracts of 1 μL were injected into a desorption liner placed inside of a Gerstel thermal desorption and cooled injection system (model CIS-4) using a Gerstel multi-purpose autosampler. Upon injection, the temperature of the thermal desorption unit was rapidly ramped from 30 to 320 °C at 12 °C min –1 to evaporate analytes which were trapped on the CIS at 30 °C. The temperature of CIS was then ramped to 320 °C at 12 °C min –1 to load analytes to a DB-5HT column (J&W Scientic, 15 m length, 250 μm i.d, 0.10 μm lm thickness). Chemicals loaded onto the column were separated using 3 mL min –1 of helium as the carrier gas. The GC temperature program ramped from 90 °C to 115, 150, 210, 280, 310 and 330 °C at rates of 95, 65, 45, 35, 30 and 25 °C min –1 respectively, followed by isothermal heating at 330 °C for 3.67 min. The temperature of the transfer line between the GC oven and the ionization source was set at 340 °C. An atmospheric pressure chemical ionization (APCI) source (positive ion mode, 150 °C, corona current: 3.0 µA, cone gas: 180 L h –1 ) was employed to ionize chemicals eluted from the GC column. As a soft ionization technique, APCI preserves analyte molecules exchange between N M or proton exchange +


Main Text
In response to the environmental and health risks posed by commercial chemicals, substantial efforts have been made to prioritize and regulate chemicals of emerging concern (CECs). 9 These efforts usually occur in the context of international 6 and national regulatory initiatives such as the US Toxic Substances Control Act, 10 and the European Union's Registration, Evaluation, Authorization and Restriction of Chemicals program. 11 Most regulatory measures begin with an assessment of a chemical's potential for causing adverse effects in organisms and the environment, mainly via their ability to resist abiotic or biotic degradation (Persistence; P), accumulate in the tissues of organisms or humans (Bioaccumulation; B), and exert adverse health outcomes (Toxicity; T) (i.e., PBT). However, the current PBT assessment approach for CECs is primarily based upon available physicochemical and toxicological data for parent chemicals, with minimal consideration given to the products associated with their atmospheric transformations.
During their residence time in the atmosphere, CECs will be transformed into other species through photooxidation. 12 Consequently, humans and ecosystems are exposed to both the parent CECs and a diverse suite of transformation products. However, CEC transformation products formed in air are not measured in current international and national atmospheric monitoring networks despite their intent of tracking global and regional distributions of CECs. [13][14][15] As a result, only a fraction of the CEC-related chemicals which exist have been monitored. This is of particular concern should the unknown transformation products be more hazardous than their precursor species, thus making an accurate risk assessment of CECs unattainable.
The lack of transformation product identi cation and information associated with their PBT is in part, due to the current target-analysis approach to identifying harmful airborne CECs. 13,14 It is also a result of the immense analytical challenge associated with tracking complex mixtures of airborne chemicals. 7 In particular, atmospheric transformation products, for which no analytical standards exist, are very di cult to distinguish from other species present in complex mixtures. With recent developments in high-resolution mass spectrometry, limited studies have attempted to utilize non-target analysis (NTA) techniques to identify unknown pollutants within complex mixtures without speci cally targeting individual parent molecules. 16,17 Although NTA techniques can help to infer chemical formulae for unknown pollutants, they cannot provide the exact chemical structures of these pollutants, or determine the mechanistic pathways through which they are formed and hence their chemical origin. 18 Such missing information hinders a comprehensive risk assessment of CECs that serves as the scienti c basis for regulations.
Organophosphate ame retardants (OPFRs) are a group of CECs that has garnered signi cant international attention. 8 They are widely used in consumer and industrial products to reduce the risk of re, and have been mass-produced due to the regulation of traditionally used polybrominated diphenyl ether (PBDE) ame retardants in recent years. 19 While OPFRs represent a type of CEC in the early stages of risk assessment, there is serious concern that they may be regrettable alternatives to PBDEs, 19 as suggested by their ubiquity in the global atmosphere, 2 long-range transport potential, 20 and demonstrated toxicity. 8 However, like many other CECs, the atmospheric chemistryhas not been considered in their overall risk assessment, despite the certainty of atmospheric chemical reactions. 12 This underscores the need for new approaches for conducting CEC risk assessment, rooted in a framework that takes into account the formation of transformation products, their presence in the ambient air, and their associated PBT properties. Below we describe the development and application of such a novel framework that considers the atmospheric chemistry of CECs.

Hazard assessment framework
The framework consists of three components: laboratory oxidation experiments, a non-target analysis, and an in-silico hazard assessment ( Fig. 1). The rst step is to simulate atmospheric photooxidation of CECs using an oxidation ow reactor, 21 combined with online transformation product detection. 22 The laboratory oxidation experiments provide comprehensive information on the atmospheric photooxidation of CECs, including the rate of photooxidation reactions (i.e., reaction kinetics), how the parent chemicals form products (i.e., atmospheric transformation mechanisms), and what products are formed (i.e., a product list). Consequently, this step forms the basis of molecular identi cation of transformation products in eld samples, as discussed below.
Non-target analysis is then applied to both laboratory samples (derived from photooxidation experiments) and eld samples (ambient air samples). NTA of eld samples often yields extremely complex mass spectra as a result of thousands of unknown species. 17 However, through the application of NTA to laboratory samples (step a in Fig. 1), we can reduce this complexity by obtaining ngerprint-like features associated with transformation products (e.g., retention times, elemental formulae, chemical structures, and isotopic patterns; see Methods) arising from any given CEC. These ngerprint-like features serve as "chemical standards", which can be used to guide the analysis of transformation products in complex eld samples. This approach enables unambiguous identi cation of transformation products in eld samples, and the ability to determine the parent chemicals from which they originated (an essential requirement for future chemical regulations).
Finally, fed with the molecular structures of the transformation products identi ed in laboratory oxidation experiments, well-established insilico models are used to predict their behavior and fate, and thus assess their PBT in a holistic multimedia environment (step b in Fig. 1). As an advantage to in-vitro and in-vivo experiments, an in-silico assessment can assess thousands of chemicals in a high-throughput, costeffective manner. Notably, the PBT hazards are quanti ed as numerical values re ecting a transformation product's impacts across multiple environmental media and multiple organs, which facilitates screening and prioritization for decision-making, and risk communication.
The new framework bridges the knowledge gap between chemical formulae (typically obtained via NTA) and chemical structures of CEC transformation products. This is particularly important given that chemical structures form the basis of the environmental and human health risk assessment of CECs, 23 and since chemical identi cation in current NTA approaches can be ambiguous (one empirical formula can correspond to multiple chemical structures). 18 Furthermore, the framework provides a new avenue for including transformation products in routine national and international monitoring programs and prioritizing transformation products of high concern for further scrutiny (step c and d in Fig. 1). In summary, the framework presents an experimentally con rmed, structure-based strategy for assessing potential hazardous outcomes associated with atmospherically transformed CECs, which has not been considered previously.

Application to organophosphate ame retardants
In this illustrative case, our framework was applied to 9 OPFRs (Extended Data Fig. 1) which represented the most frequently detected OPFRs in ambient air. 13 Upon photooxidation in laboratory experiments, 186 transformation products were detected, representing early (formed through less than three reaction steps) to later generation transformation products (Supplementary Table 1). Knowledge of the product formulae and principles of organic reactions were used to infer the most likely transformation mechanisms and to elucidate the molecular structures of transformation products. Based upon individual mechanisms (Supplementary Figs. 1 and 2), a general reaction scheme that includes three main channels can be formulated for OPFRs (Extended Data Fig. 2).
Use of NTA on the laboratory samples derived from 5 reacted OPFRs (the ve most abundant OPFRs in the ambient air) 13 resulted in chromatographic and mass spectral information for 27 products (Supplementary Table 2). This information was then used to screen NTA data from eld samples collected in 18 megacities under the Global Atmospheric Passive Sampling-Megacities program (see Methods). In the eld samples, 20 transformation products (derived from 5 OPFRs) were characterized. Of these transformation products, 11 can be assigned to known molecular structures (i.e., identi ed products; Extended Data Figs. 3 and 4) and 9 were derived from chlorinated OPFRs but without known structures (i.e., unidenti ed products; Supplementary Table 3). The 11 identi ed products, with detection frequencies of 94-100% (Supplementary Table 4), can be further divided into two groups based upon their structures: 7 chlorinated products (identi ed based upon their retention times, mass peaks, and isotopic patterns) and 4 non-chlorinated products. These transformation products are primarily early-generation products, consistent with the expectation of short oxidation timescales associated with urban regions. 24 These 20 compounds are the rst reported measurements of airborne OPFR transformation products. However, we expect that they represent the tip of the iceberg, as more transformation products are likely to be found in the atmosphere, e.g., in rural and remote regions where more highly aged air masses favor the formation of additional products. 25 The global distribution of the 11 identi ed OPFR transformation products is shown in Fig. 2, using volume normalized signal intensity (counts m -3 ) as a proxy for air concentration (in the absence of authentic standards). The geographic distribution in Fig. 2 re ects the ubiquity of transformation products in the global atmosphere arising from the widespread use of OPFRs. Also, high levels of parent OPFRs in a given city result in correspondingly high levels of transformation products (Fig. 3a). For example, the total normalized signal intensities of 11 identi ed products are highest in London, UK, and New York, USA (2-35 times higher than other cities), consistent with the highest concentrations of parent OPFRs (sum=14,000 pg m -3 ). 13 Further analysis also indicates that there is a seasonality associated with transformation product formation. For example, among the urban samples from Toronto, Canada, 26 the ratio of product to parent OPFR signal intensity (R signal ; a parameter that re ects the extent of photooxidation) is up to 6 times higher in summer than in winter (Fig. 3b).
Given that the increased light levels in summer favor atmospheric photooxidation, this observed seasonal contrast veri es that the identi ed species are a result of photochemical reactions.
The ubiquity of OPFR transformation products (Fig. 2) is especially concerning given their in-silico derived PBT properties illustrated in Fig.   4. Fig. 4a shows the fold changes in the PBT of 186 transformation products relative to the 9 investigated parent OPFRs; a fold change higher than 1 indicates an increased environmental concern associated with a product relative to its precursor. Fig. 4a indicates that the overall persistence, i.e., the overall ability to resist various degradation processes across all environmental media, for the transformation products of chlorinated OPFRs are on average 2.5 times higher than their originating parent molecules (90% of the red data points in Fig. 4a lie to the left of the dashed line; for chemical-speci c results, see Extended Data Fig. 5). Conversely, transformation products of nonchlorinated OPFRs (blue data points in Fig. 4a) are on average 25% less persistent. This contrasting observation arises from differences in the multimedia mass distributions (atmosphere vs. surface media including surface water, soil, and sediment) of the transformation products derived from chlorinated and non-chlorinated OPFRs. Speci cally, transformation products of chlorinated OPFRs, have a higher a nity for surface media due to an increase in octanol-air partition coe cients (K OA ) from atmospheric photooxidation (Extended Data Fig.   6). Since almost all the product species investigated here have an order-of-magnitude slower rate of degradation in surface media than the rate of atmospheric photooxidation ( Supplementary Fig. 3), such a preference for surface media preserves more of the total mass within the environment, resulting in a higher overall persistence. Conversely, atmospheric photooxidation of non-chlorinated OPFRs forms products with a higher fraction present in the atmosphere, leading to lower overall persistence. Nevertheless, 7 of the 9 OPFRs investigated form transformation products of which at least 5% (and up to 93%) are more persistent than the parents (Fig. 4b).
Transformation products of all OPFRs are orders of magnitude less bioaccumulative in aquatic biota than their corresponding parents ( Fig.   4a and Extended Data Fig. 7). Atmospheric oxidation adds oxygen-containing functional groups to OPFRs and generates more hydrophilic transformation products, as demonstrated by their lower (relative to parent molecules) octanol-water partition coe cients (K OW ; see Extended Data Fig. 6). Hence, these products are more easily eliminated through gill respiration in aquatic animals and are thus associated with reduced bioaccumulation potential. 27 However, the increase in hydrophilicity does not always imply a reduced environmental concern, as a chemical will exhibit increased environmental mobility if its K OW is lower than 10 4 , 28 as is the case for 91% of the transformation products investigated here. This makes many of these products more accessible to aquatic biota and more di cult to remove from water through traditional water treatment procedures, 28 implying an additional risk to the health of aquatic ecosystems and to the safety of drinking water.
The systemic toxicity, i.e., the toxicity to the entire animal, differs greatly between parent OPFRs and transformation products (Fig. 4a). While some chlorinated OPFRs can form less toxic products, the majority of non-chlorinated OPFRs form products up to an order of magnitude more toxic than their corresponding parents (Fig. 4a). This is consistent with photooxidation, which can toxify non-chlorinated OPFRs through the addition of more toxicologically active functionality, such as carbonyl groups. 29 Conversely, photooxidation can also detoxify chlorinated OPFRs by substituting more toxicologically active structures (e.g., chlorine) with less active moieties such as hydroxyl groups. It should be noted that recent experimental evidence indicates certain transformation products may also cause adverse toxicological effects. 30 For example, product TCPP-21 and TDCPP-14 (8.0-fold and 3.7-fold increase in relative LD50 ratio respectively; Extended Data Fig.   8) have been con rmed to be able to disrupt endocrine receptors in animals. 30 Moreover, all the transformation products investigated here are predicted to form a signi cant fraction (24-89%) of more toxic transformation products (Fig. 4b). We also note that all but three of the transformation products identi ed in ambient megacity samples are more toxic than their parent compounds (Extended Data Fig. 8).
The current results have profound implications for the ongoing assessment of hazardous outcomes of CECs. The global atmospheric ubiquity of OPFR transformation products shown here, combined with their tendency to be more persistent, environmentally mobile, and toxic, than their parent chemicals indicates that the human and environmental risks of OPFRs in current assessments which focus on parent compounds alone, may be underestimated. Including atmospheric transformations in the assessment of OPFRs is hence a requirement for determining the actual hazard associated with their use as PBDE replacements. Furthermore, the new framework described herein is transferable to any CEC, as the vast majority of commercial chemicals released to air will be subjected to atmospheric transformations. It also provides a holistic means to include atmospheric transformation products of CECs in future monitoring networks, exposome studies, 31 and chemical management practices.
Finally, it is widely recognized that PM 2.5 (particulate matter with a diameter of smaller than 2.5 micrometers) is a major global human health concern. 32 Since many CECs (including OPFRs) are components of PM 2.5 , global efforts in reducing PM 2.5 levels might also be expected to reduce exposure to CECs and their transformation products. However, as shown in Extended Data Fig. 9, the global distribution of OPFR transformation products in air observed here is clearly skewed to high-income nations (and lowest in low-income nations). 33 Conversely, global PM 2.5 mass concentrations are highest in lower-middle-income nations and lowest in high-income nations. While there are clearly other factors contributing to this pattern, these opposing trends suggest that the impact of commercial chemicals and their transformation products in the atmosphere, will be independent of having successfully reduced PM 2.5 levels. It may also act as an early warning to both high and low-income nations, highlighting the need to tackle both PM 2.5 reductions and chemical exposures simultaneously.

Methods
Global Atmospheric Passive Sampling-Megacities (GAPS-MC) Sampling details. The GAPS-MC network consists of 18 megacities across the globe with a population ranging from 2 million to 22 million. 13 A representative site for each megacity was chosen such that they were well away from potential sources of targeted CECs and persistent organic pollutants (POPs) and with an unobstructed air ow and preferably centrally located within the city. Sampling dates are shown in Extended Data Table 1. Further details regarding the coordinates and site characteristics have been described previously. 13 Polyurethane foam disk passive air samplers (PUF-PAS; Tisch Environmental, 14 cm diameter, 1.35 cm thickness, 370 cm 2 surface area) were used for sampling during the GAPS-MC project in 2018-2019. This double-dome con guration of PAS has been previously used for GAPS network sampling of a range of gas-and particle-phase CECs and POPs. 2 Extraction procedure. PUF-PAS samples were extracted using an ASE instrument with the combination of petroleum ether and acetone solvents (5:1, v/v for PUF-PAS samples; 3:1, v/v for ow reactor samples). The PUF-PAS extracts were then reduced to 0.5 mL using rotary evaporation and nitrogen blowdown. Mirex was added to PUF-PAS sample extracts as an internal standard before splitting them into two equal halves by volume. The rst half of the sample was reconstituted in isooctane along with the addition of respective standards, and was used for the non-target analysis of OPFR transformation products. The other half of the sample was eluted through a silica column followed by addition of OPFR internal standards and re-constituting the nal volume in methanol, which was used for the target-analysis of parent OPFRs (see Supplementary Information). The oxidation ow reactor samples were not split into two halves and the entire volume was reduced to 0.5 mL using acetonitrile as a solvent without addition of any standards.
Sampling volume. The equivalent sampling air volume (Veq, m 3 ) was estimated using the GAPS template. 34 This template takes into consideration the PUF-air partition coe cient for each target analyte and average temperature of the sampling period for correction of Veq, and has been used previously under the GAPS network and other regional studies. 2,13 A generic sampling rate of 4m 3 day -1 was used for calculations. 13 Oxidation ow reactor (OFR) experiments Experiments investigating the heterogeneous OH oxidation of 9 different types of OPFR particles were performed using an OFR. 35 Seed particles of (NH 4 ) 2 SO 4 were generated via atomization (TSI, model 3706), dried through a diffusion drier (TSI, model 3062), and passed through the headspace of a temperature-controlled Pyrex tube (343 K) containing a pure OPFR, to produce coated particles with a calculated coating thickness of ~15 nm based upon the shift in the peak of the particle size distribution which was measured using a TSI scanning mobility particle sizer. For TCEP and TCPP, pure organic particles were generated by sending an aqueous solution of each organic through a TSI atomizer and drawn through a diffusion drier. Those OPFR particles were introduced into the OFR and were exposed to O 3 (0−2 ppm) under 254 nm UV light irradiation at 298 K in each experiment. OH radicals were generated by the photolysis of O 3 at 254 nm followed by reaction with water vapor (35% RH). In o ine calibrations, the OH exposure was calculated through the loss of CO due to its reaction with OH, 36 and was in the range of 1.2×10 11 -1.6×10 12 molecules cm -3 s. Further details regarding the OFR experiments are provided in the Supplementary Information.
The particulate OPFR products formed from heterogeneous OH reactions were measured using a recently developed extractive electrospray ionization time-of-ight mass spectrometer (EESI-TOFMS). 22,35 Brie y, the particle-containing ow exiting the ow tube reactor passed through a multichannel extruded carbon denuder at the inlet of the EESI-TOFMS which removed most of the gaseous species with high e ciency. 22 The oxidized OPFR particles then collide with electrospray droplets generated at the end of an electrospray capillary (New Objective; #TT360-50-50-N-5) at a ow rate of 1 μL min -1 . Soluble components of the particles were extracted, ionized through a Coulomb explosion of the charged droplets, and detected by a Tofwerk APi-TOFMS. The electrospray working solution was a wateracetonitrile mixture (1:1 by volume) with 110 ppm of sodium iodide as a charge carrier. The mass spectra were recorded in positive ion mode (i.e., Na + adducts).
Non-target analysis GC-APCI-TOFMS measurement. Using an NTA method described previously, 37 the o ine OFR sample extracts and ambient air extracts from the GAPS-MC network were analyzed using a gas chromatography (GC, Agilent 7890) system coupled with a quadrupole time-of-ight mass spectrometer (Waters Xevo G2XS). Sample extracts of 1 μL were injected into a desorption liner placed inside of a Gerstel thermal desorption and cooled injection system (model CIS-4) using a Gerstel multi-purpose autosampler. Upon injection, the temperature of the thermal desorption unit was rapidly ramped from 30 to 320 °C at 12 °C min -1 to evaporate analytes which were trapped on the CIS at 30 °C.  Table 1) obtained from online OFR measurements (step a in Fig. 1). In this step, we observed 27 products in o ine OFR samples collected from 5 oxidized OPFRs (TCEP, TCPP,  TDCPP, TBEP, and TPhP), with detailed information (retention time, m/z peak, isotopic patterns, chemical formula, and chemical structure) given in Supplementary Table 2. The number of detected transformation products in o ine OFR samples and in ambient samples using this method were less than those detected in OFR online experiments (128 products for the 5 OPFRs mentioned above). This is expected, as some products may be below the method detection limits for the transformation products. Also, it is known that some oxidized products (e.g., peroxides) are likely to undergo decomposition during sample collection and extraction due to their high reactivity (i.e., having a short lifetime). 38 Finally, the number of transformation products identi ed in ambient samples here, are likely to represent only a fraction of the products which exist, due to the potential for decomposition of oxidized species at high temperatures 39 often associated with GC systems.
Unidenti ed OPFR products. In addition to the 11 identi ed OPFR products, we observed 9 unidenti ed products in the GAPS-MC eld samples (Supplementary Table 3). Their chemical formulae are tentatively proposed according to their mass peaks and isotopic patterns ( Supplementary Fig. 4). They are considered "unidenti ed products" for two reasons: (1) some of the products in Supplementary Table 3, which were included in our product list obtained from OFR experiments (Supplementary Table 1), had mass uncertainty greater than 20 ppm (e.g., U 2 and U 4 in Supplementary Table 3). (2) several products were not observed in the OFR experiments (i.e., not included in our product list), and their chemical structures are unknown. While the chemical structures of these products are unclear, we may infer their source (parent compounds) from the proposed chemical formulae. Given that all of these unidenti ed products exhibit clear 37 Cl isotopic patterns ( Supplementary Fig. 4), they are originated from chlorinated OPFRs such as TCEP, TCPP, and TDCPP. The volume normalized signal intensities of 9 unidenti ed products in 18 megacities are shown in Supplementary Fig. 5. It is interesting to note that London, UK and New York, USA possess the highest overall signal intensities (36.6 counts m -3 for London and 5.1 counts m -3 for New York). This is consistent with the observed high levels of parent chlorinated OPFRs (>10000 pg m -3 ) 13 and high levels of identi ed chlorinated transformation products in these two cities (Fig. 3a). The consistency of the levels of chlorinated OPFR parent compounds (TCEP, TCPP, and TDCPP), identi ed chlorinated products (e.g., TCEP-1, TCPP-21, and TDCPP-14) and the sum of the unidenti ed chlorinated products (U 1 -U 9 ) provides additional con rmation of our product identi cation.
In-silico modelling RAIDAR model. Overall persistence and bioaccumulation were modelled using the Risk Assessment, Identi cation, And Ranking (RAIDAR) model. RAIDAR is a steady-state multimedia fate and exposure model in support of high-throughput hazard and risk assessments for chemical substances. 40,41 RAIDAR has been successfully applied to screen and prioritize more than 10 thousand neutral and ionogenic organic chemicals within the markets of OECD (Organization for Economic Co-operation and Development) countries, the US, and Canada. [42][43][44][45][46] RAIDAR is a mechanistic model ( Supplementary Fig. 6), it builds upon mechanistic descriptions of a series of physical, chemical, biological, and physiological processes governing chemical fate and exposure, which re ect the state-of-the-art understanding of chemical fate and toxicokinetics. It describes the transport and transformation of chemicals in a regional environment (a total area of 10 11 m 2 ) comprising four bulk compartments: air (with a volume of 1.0×10 14 m 3 ; gaseous and particle phases), surface water (2.0×10 11 m 3 ; water and suspended solid phases), soil (1.8×10 10 m 3 ; gaseous, pore water, and solid phases), and sediment (5.0×10 8 m 3 ; water and solid phases).
Here, the surface water, soil, and sediment can be collectively referred to as surface media. Chemicals undergo advection, diffusion, and transformation within and between these compartments. Notably, ionogenic organic chemicals can dissociate in the water phase of surface media (soil has a pH of 6 whereas other surface compartments share a pH of 7). The modeled region has an average temperature of 10 °C Persistence. The "overall persistence (P OV )" is de ned as the average time that a chemical resides in the multiple environmental compartments. 47,48 In RAIDAR, P OV is mathematically expressed as the ratio of the total amount of chemical in the modeled region to the total rate of loss from the modeled region given by: 41 where i = air, surface water, soil, and sediment. Equation 1 indicates that P OV is determined mainly by two factors: (1) rates of chemical losses in individual environmental compartments, including the rates of reaction with OH radicals (for the air compartment) and biodegradation (for surface media), and (2) multi-compartmental distribution, which is governed by the phase partition behavior of the chemical of interest. Since P OV is a hypothetical, integrative measure of a chemical's overall persistence in the multimedia environment, it is not observable in real-world studies. The higher the P OV , the longer a chemical is present in the environment, and its contamination, once it occurs, is less reversible. P OV is an integrative measure because it allows for the interplay between the multimedia distribution and degradation in individual media; that is, we can expect a high overall persistence if the property of a chemical favors its stay in an environmental medium where it undergoes negligible degradation. An advantage of using P OV in this work is that we do not need to measure the yields of transformation products, given that P OV is an intensive property, which is independent of the actual mass formed or present in the environment.
Bioaccumulation. The RAIDAR model also describes absorption, biotransformation, and excretion of chemicals in living organisms in terrestrial and aquatic food webs. Speci cally, it contains a sh bioaccumulation module, which evaluates the accumulation of chemicals in the sh tissue from water by quantifying toxicokinetic processes including gill respiration, dietary ingestion, metabolic transformation, fecal egestion, and growth dilution. In doing so, RAIDAR estimates chemical concentrations (C B ) in the sh tissue. Mechanistically, the net accumulation of a chemical in the sh tissue is the difference between chemical ows entering and leaving the sh body; as such, a chemical would have a limited bioaccumulation potential if it easily departs from the sh body, e.g., due to a high rate of gill mass exchange (for highly soluble chemicals) or metabolic transformation (for less persistent chemicals).
Here, we calculated a "bio-concentration factor" (BCF; in L kg −1 ) as the ratio of wet weight-based concentration of chemical in sh tissue (C B , g chem /kg wet weight ) to the total concentration of chemical in the surface water compartment (C W , g chem /L water ), assuming an absence of dietary intake of the chemical by the modeled sh (i.e., gill absorption is the sole source of chemical to sh): This manner of calculation is consistent with the common practice in OPFR bioaccumulation experimental studies. 49 RAIDAR requires properties of chemicals as basic inputs for the computation, including: (1) partition coe cients (K OW , K AW , and K OA ) (referred to as "distribution ratios", D OW , D AW , and D OA , for ionogenic organic compounds) between the air, water, and n-octanol phases, (2) dissociation constants (pKa), (3) rate constants of environmental reactions with OH radicals (for calculating degradation in the air) and biodegradation (for calculating degradation in surface media), and (4) rate constants of biotransformation in sh. In this work, we use a quantitative structure-activity relationship (QSAR) model called "OPEn-QSAR-App (OPERA)" to predict partition coe cients for neutral transformation products. 53 For ionogenic transformation products, the partition coe cients of the neutral fractions are predicted using OPERA, whereas the partition coe cients of charged fractions are calculated from those of the neutral fractions based on an empirical relationship described previously. 54 The distribution ratios, which combine partition coe cients of the neutral and charged fractions, are then calculated using the Henderson-Hasselbalch equation. Dissociation constants (pKa) and rate constants of environmental reactions (2 &3 above) are also predicted using OPERA. 53,55 Rate constants of biotransformation in sh are predicted using a QSAR model named Iterative Fragment Selection (IFS). 56 Of the total of 195 investigated OPFR-related species (9 parent OPFRs + 186 transformation products), 190, 134, 72, and 179 have partition coe cient predictions, pKa predictions, environmental reaction rate constants, and biotransformation rate constants, within the domains of applicability of the QSAR models. Chemicals falling in the domain of applicability are of the highest con dence; however, chemicals outside the domains of applicability are also used in this work due to the lack of other available QSAR models at this point.
Toxicity. Toxicity is quanti ed using an oral 50% lethal dose (LD 50 ) for rats, which is de ned as the dose of chemical (normalized by bodyweight) that can cause 50% of rats in a population to die after oral ingestion. A lower LD50 corresponds to higher toxicity. In this work, the toxicities of the OPFR transformation products were estimated using the Toxicity Estimation Software Tool (TEST). 57 TEST is a QSAR model in support of the prediction of various toxicological endpoints from a chemical's structure. Speci cally, TEST rst calculates a series of chemical-speci c molecular descriptors based on the entire or partial features of a molecule, selects appropriate molecular descriptors based on pre-de ned statistical algorithms, and then uses mathematical combinations of these molecular descriptors to predict toxicological endpoints. TEST's prediction algorithm for oral rat LD50 is based on a training set of more than 7,400 chemical substances. 57 For each compound, TEST employs three methods, i.e., the hierarchical, FDA, and nearest neighbor methods, to predict the oral rat LD50 and uses their average as a consensus estimate for output. Earlier statistical external validation indicates that the consensus model explains 62.6% of total variation observed among chemicals in the training set, with a root mean square error (RMSE) of 0.54 log units. Oral rat LD50 has previously been experimentally determined by toxicological studies for select OPFRs. Supplementary Table 5 compares TEST predictions for OPFRs with observed oral rat LD50 used as the TEST training set, which indicates satisfactory agreement for most OPFRs investigated here.
Note that our analysis (OPFR product identi cation and NTA) does not distinguish between positional isomers, i.e., isomers sharing the same functional groups but with different positions (such as ortho-, meta-, and para-aromatic substitutions). However, Supplementary Fig. 7 indicates that PBT differs only marginally between positional isomers. As such, it is acceptable to ignore positional isomers of transformation products in such high-throughput screening efforts.