While the prevalence of cigarette smoking is declining in many developed countries (Global Burden of Diseases, 2017), the actual number of adult smokers is still significant. For instance, there are over 34 million current smokers in the United States (US) (NHIS, 2018; U.S. Centers for Disease Controland Prevention, 2019) and cigarette smoking remains the leading cause of major preventable diseases, morbidity, and mortality. While public health strategies for reducing smoking-related harm have focused on preventing smoking initiation and promoting smoking cessation (U.S. Surgeon General, 2010; U.S. Surgeon General, 2014; Royal College of Physicians, 2016), many smokers are either unwilling or unable to quit and the long-term cessation success rate remains very low with most quit attempts often failing within the first year (Hughes et al., 2004; U.S. Surgeon General, 2010; Medicine, 2012; Royal College of Physicians, 2016). As former US Food and Drug Administration (FDA) Commissioner, Scott Gottlieb, pointed out, “we must recognize the potential for innovative, less harmful products that can efficiently deliver satisfying levels of nicotine to adults who want them (U.S. Food and Drug Administration, 2018). In the United Kingdom, the use of reduced-risk tobacco products (RRPs), in particular e-vapor products, is viewed as a promising tobacco harm reduction (THR) strategy. An expert review published by the Public Health England (PHE) stated that “vaping poses only a small fraction of the risks of smoking and switching completely from smoking to vaping conveys substantial health benefits (McNeill et al., 2018). At the same time, the long-term health risks of e-vapor and flavor ingredient exposures are unknown and further investigations in long-term adverse effects are necessary (National Academies of Sciences, 2018). Currently a variety of potential RRPs are available around the world, including smokeless oral tobacco products such as snus, “tobacco-free” oral nicotine products, heated tobacco products (e.g., IQOS®, Phillip Morris International), and electronic nicotine delivery systems (ENDS) or e-vapor products. While promising, the long-term success of complete switching from cigarettes to RRPs is unknown (Polosa et al., 2013; Farsalinos et al., 2018). If smokers do not accept these products as sustainable alternatives to cigarettes regardless of how low their toxicity profile is, they will not switch, and cigarette smoking cessation will not occur. Among RRP candidates, e-vapor products deliver nicotine-containing aerosols via inhalation, potentially enabling rapid nicotine uptake and a sensorial experience overall similar to smoking (Voos et al., 2020). As postulated in Abrams et al., e-vapor products are in a unique position to achieve THR for adult smokers in terms of (a) (reduced) harmfulness, (b) appeal, and (c) satisfaction (Abrams et al., 2018). The Institute of Medicine similarly describes the importance of product appeal to a smoker seeking RRPs to switch: “Ideally, a reduced risk tobacco product would be sufficiently reinforcing so as to attract smokers away from conventional cigarettes but not encourage the widespread dependent use of the product by individuals who were previously nonusers or who would have quit smoking.” (Medicine, 2012). The “sweet spot” on the THR model is where RRPs are substantially lower than cigarettes in toxicity/harmfulness, relatively high in appeal to current tobacco users but not adversely impacting initiation by non-smokers including youths (Smith et al., 2016; Abrams et al., 2018; Warner, 2019).

Among product- -related factors that would affect the liking and acceptability of RRPs, flavor ingredients as well as nicotine contribute to taste and sensory experiences that may enhance satisfaction and switching away from cigarettes. The US Population Assessment of Tobacco and Health (PATH) survey data suggest the potential role of flavor ingredients to the e-vapor product user experience that adult e-vapor users engage with a wide variety of e-liquid flavor ingredients, with preferences for non-tobacco flavor ingredients (e.g., fruit, mint, or menthol flavors; Figure 1). Several studies show that e-vapor product users frequently report first using or buying a flavored e-vapor product (Harrell et al., 2017; Russell et al., 2018; Landry et al., 2019; Villanti et al., 2019; Vu et al., 2019) and having diverse flavor options was critical to aid continued e-vapor product (only) use over smoking (Chen et al., 2018; Friedman and Xu, 2020).

While RRPs, such as e-vapor products, are designed to be substantially less toxic than cigarettes (e.g., lack of tobacco combustion), no tobacco product (including RRPs)is risk-free. For e-vapor products, the long-term health risks of their aerosols, which contain humectants and flavor ingredients, have not yet been assessed and it will take decades before any potential long term health impacts associated with chronic e-vapor use are fully understood. The National Academies of Sciences, Engineering, and Medicine confirmed this notion, suggesting “that e-cigarettes are not without physiological activity in humans” and that further investigations are needed (National Academies of Sciences, 2018). Many non-nicotine ingredients used in e-vapor products [e.g., flavor ingredients and humectants such as propylene glycol [PG] and vegetable glycerol (VG)] are GRAS for oral consumption in food, but in general have insufficient toxicity data available when exposure occurs via inhalation (Committee on Toxicity of Chemicals in Food C.P.a.t.E.C, 2016). While safety profiles via oral exposure provide some assurance in terms of systemic toxicity, they are not applicable in evaluating local aerosol toxicity in the respiratory tract. An often-quoted example is that of diacetyl, a buttery flavor that is commonly used in food products and GRAS for oral consumption (Hallagan, 2017), which is extremely toxic via chronic inhalation exposure causing bronchiolitis obliterans–a lung specific and irreversible respiratory disease (Kreiss et al., 2002; Wallace, 2017). An increasing number of studies suggest that e-cigarettes or e-liquids may cause respiratory tissue damage via inflammation and immunogenic effects and that flavor ingredients may contribute to their toxicity (Higham et al., 2016; Martin et al., 2016; Kaur et al., 2018; Hua et al., 2019). In the United States, all tobacco products marketed after 15 February 2007 must seek market authorization from the FDA. Given the relatively new presence of ENDS, FDA recommends submission of Pre-Market Tobacco Applications (PMTAs) to show whether permitting such new tobacco product to be marketed is appropriate for the protection of the public health, which includes potential health risk to the individual and the population as a whole.

Ideally, e-vapor toxicity assessments should address whether flavor ingredients individually, or as mixtures, cause adverse effects to the respiratory tract upon repeated exposures. The classical approach to evaluate the toxicity of flavor ingredients requires a series of in vitro and in vivo studies to fill data gaps for individual flavor ingredients and to determine acceptable use levels. However, considering the large number of flavor ingredients available and the even greater number of possible combinations of mixtures in e-liquids (Zhu et al., 2014), the standard approach requires years of testing and thousands of laboratory animals. To add complexity to investigating health effects of e-vapor products, there are no harmonized, preclinical testing methodologies, including e-vapor aerosol generation and characterization, making comparison of findings from different studies difficult (Orr, 2014; Royal College of Physicians, 2016; Breland et al., 2017; McNeill et al., 2018; National Academies of Sciences, 2018; Noël et al., 2018).

In this paper, we propose an alternative approach to evaluate to toxicity of flavors via inhalation, utilizing the “read-across” and a “flavor toolbox” concepts. Instead of testing individual flavor ingredients, we assign each flavor ingredient to a structurally related group (“read-across”), rank and select the “Flavor Group Representatives (FGRs)” based on defined criteria and create mixtures of FGRs (to represent the “flavor toolbox”) for subsequent confirmatory in vitro or in vivo toxicity testing. This concept was presented at scientific meetings (Sciuscio et al. (2021); Marescotti et al., 2021; Supplementary Resource S1).

The first step of the flavor toolbox approach is to select a list of candidate flavor ingredients and evaluate their toxicity profiles for potential ranking. In this case study, 245 commonly used individual flavor ingredients were selected and their available toxicological data were collected using both the (European Chemicals Agency, 2002)) and (Toxplanet, 2021) databases. If there were multiple values for the same toxicological endpoint (e.g., LD50), the ECHA approach was applied and the most relevant data set was selected as follows: in the context of REACH (Registration Evaluation Authorization and Restriction of Chemicals), evaluation of data quality includes an assessment of the adequacy of the information for hazard/risk assessment as well as for classification and labelling (C&L) purposes and elements of relevance and reliability. The ECHA data categorization for the selection of data from reliable “key studies” only (Klimisch et al., 1997) was used whenever possible. A study was considered a reliable “key study” if it was generally expected to be the most adequate, reliable, and relevant for a specific element/endpoint study section. If properly reported, a key study may fulfil a REACH information requirement on its own. At the same time, we confirmed that all 245 flavor ingredients were not classified as respiratory sensitizers or as group 1, 2A, or 2B carcinogens in the IARC classification or CMR by the FDA.

During the review of available data, we found that most individual flavor ingredients lacked experimental toxicity data. Therefore, additional in vitro testing was performed on all 245 individual flavor ingredients for cytotoxicity screening by real-time cellular analysis (RTCA). Figure 3 summarizes the Tox Score (ratio of the EC50 of the base solution to the EC50 of the flavored solution) versus the p values for all individual flavor ingredients tested. As described previously (Marescotti et al., 2020), we applied a statistically significant Tox Score value of 1.5 to identify the most cytotoxic flavor ingredients in the list and selected a total of 34 flavor ingredients for subsequent in vitro characterization of their mode of action using HCS analysis.

In the HCS assessment (Table 1), Normal Human Bronchial Epithelial (NHBE) cells were exposed in a concentration-dependent manner to 34 individual flavor ingredients and to non-flavored (base) matrix for either 4 h or 24 h (30 min for the NF-kB translocation endpoint.) (Gonzalez-Suarez et al., 2016; Marescotti et al., 2016) Results are shown in Figure 4 in three heatmaps which grouped these flavor ingredients based on minimal effective concentration (MEC) profile similarities and the Tox or Phenotypic Score (ratio of MEC base solution/MEC flavored solution) is shown in Table 2.

Flavor ingredients in heatmap “A” had the lowest impact for most of the HCS endpoints tested, which also fits with their lower Tox Score as well as with their lower phenotypic score (Table 2). Flavor ingredients in heatmap “B” showed the highest and most significant impact for reactive oxygen species (ROS) and glutathione (GSH) content endpoints after 4 h exposure and for DNA damage and GSH content endpoints after 24 h exposure (Figure 4). Also, for this group, no ratio was computed for the cytochrome c release endpoint at 4 h (except alpha damascone) or the cell membrane permeability endpoint at 24 h. For heatmap “B,” the range of phenotypic scores was between 5.88 and 13.36 (Table 2). Flavor ingredients in heatmap “C” had a phenotypic score ranging between 0.71 and 5.24 (Table 2).

In addition to the experimental data, TOPKAT^® (Toxicity Prediction by Komputer Assisted Technology, 2017) prediction software was used to acquire predicted toxicological information for all 245 flavor ingredients. TOPKAT contains models covering a range of different toxicity endpoints including rat chronic LOAEL, irritancy, developmental toxicity, rodent carcinogenicity, carcinogenicity potency TD50, and other endpoints. Many of the models provide a binary summary prediction (yes or no), whereas other models, such as those for acute oral toxicity, estimate a LD50 value. Additionally, Cramer classes were assigned to each flavor ingredient using profilers available in the QSAR toolbox software (OECD QSAR, 2019). Cramer classes have been proposed to support the Threshold of Toxicological Concern (TTC) for inhalation toxicity and chemicals assigned to Cramer class I are deemed less toxic than those assigned to Cramer classes II or III both at local (respiratory) and systemic levels (Carthew et al., 2009; Costigan and Meredith, 2015). For example, the application of the Cramer classification tree to a data set of 92 subacute or subchronic rat inhalation studies resulted in TTC values of 1400 (class 1) and 470 µg/person/day (class III) for local effects (based on human lung weight of 650 g) and of 980 (class 1) and 170 µg/person/day (class III) for systemic effects (Carthew et al., 2009). Although these TTC values have been proposed to aid in toxicological risk assessment of flavor ingredients in e-liquids (Costigan and Meredith, 2015), in this case study we did not use TTCs as the limit for AULs. Instead, we used the TTC as well as in silico data to rank flavor ingredients for their toxicity potential for the subsequent FGR selection. Supplementary Resources S2, S3 show the summary of the obtained literature/experimental (Group 1–2B) toxicity data sources and the in silico predicted toxicological tools and endpoints.

The second step of the flavor toolbox approach is to assign each individual flavor ingredient into the appropriate flavor group based on their structural, toxicological, and metabolic properties. The 245 flavor ingredients selected in this case study were initially allocated to one of the 34 structural groups defined in the European Commission (EC) Regulation No 1565/2000 (European Commission, 2000). Some groups (e.g., EC groups 1 and 2) contain many flavor ingredients (primary aliphatic alcohols/aldehydes/acids, acetals, and esters) and these broader, heterogeneous, structural groups were further subdivided to better represent the range of structural differences. Following this approach, a total of 38 groups were defined, encompassing 27 of the original 34 EC groups (Table 3).

The third step of the flavor toolbox approach is to select the flavor ingredients that best represent each flavor group. This step can be achieved by ranking each of the flavor ingredients within each flavor group on the basis of the available experimental and predicted toxicological data and subsequently applying a numerical scoring and a computational procedure to select the FGR, the flavor ingredient with the predicted highest toxicity potential within its flavor group.

A numerical score (code) for toxicological attributes (if available) was assigned as follows:

• pCramer: Cramer class coded 0, 1 or 2 (for Cramer class I, II, and III, respectively).

In order to select FGRs, flavor ingredients within each group were ranked based on pLD50, pDevToxicity, PredictedToxPi, pChronicLOAEL and pIrritancy scores. For each flavor ingredient, the average rank was then computed and used to generate the final ranking. The resulting FGRs are listed in Table 3 and Supplementary Resource S3 details the ranking process using one example (group 1–2B).

The fourth step of the flavor toolbox approach is to create a mixture as a proxy to the full “flavor toolbox” by combining all FGRs and assess its toxicity. This RFM should be tested in vitro and in vivo in order to characterize both the hazards associated with the RFM and to define the in vivo benchmark dose or NOAEL. Below is the recommended plan for these preclinical studies. Some of preliminary findings for the case study mixtures were presented at scientific meetings (Supplementary Resource S1), as briefly summarized.

The fifth step of the flavor toolbox approach is to confirm AULs for the remaining flavor ingredients included in the initial list. This step is applicable after the completion of the chronic in vivo studies of the RFMs and the confirmation of NOAELs based on evaluation of chronic toxicity outcomes. We hypothesize the chronic in vivo inhalation data would allow the determination of the benchmark dose or NOAEL (or LOAEL) based on systemic and respiratory toxicity and carcinogenicity outcomes of RFMs. Then, AULs for the rest of flavor ingredients within a given chemical group could be matched to the AUL of its respective FGR based on the ‘read-cross’ principle that flavor ingredients within a group have an equal/lower biological activity than the respective FGR. As discussed earlier, this read-cross step should be accompanied by independent mechanism-based HCS in vitro studies for all flavor ingredients against the selected FGRs. The outcomes of these in vitro studies should confirm that the biological activity across all flavor ingredients within a chemical group are qualitatively comparable and that the selected FGRs are indeed the most biologically active (toxic) ingredient in the respective chemical groups. Following these in vitro and in vivo confirmatory studies, the NOAELs of RFMs could be applied to the remaining flavors in the toolbox list.

After the above steps are executed according to the proposed rationale and assumptions, the sixth and the last step of the flavor toolbox approach is to explore the approach applied to new flavor mixtures. In this step, a hypothetical new flavor mixture (nFM) can be generated using any flavors in the toolbox list and the AULs defined in the above sections. This nFM should be tested for the battery of in vitro OECD and mechanistic assays suggested above, and the results compared to the corresponding in vitro testing of the specific RFM tested in chronic in vivo studies at the NOAEL or LOAEL level. The in vitro testing would allow an indirect screening of unexpected, synergistic, or adverse responses from the nFM mixtures. If the comparison of in vitro battery results shows that the nFM has toxicity lower than/equal to that of the RFM, then the nFM could be considered for subsequent product development. Over time, the cumulative results of this approach would help expand the toolbox database, inform potential interactions among flavor ingredients, enhance predictability of the flavor “read-across” approach, and limit the need for in vivo studies to investigate in vitro results.