2025 Volume 50 Issue 2 Pages 57-68
A physiologically based kinetic (PBK) model is used for predicting chemical concentrations of toxicological concern in target tissues. Such models are important for understanding toxicokinetics. However, it is challenging to obtain chemical-specific empirical parameter values used for PBK modeling. Thus, developing methods predicting these values is necessary. Herein, we researched PBK models of inhalation exposure to industrial chemicals and developed a database of parameters of approximately 200 chemicals in humans and rodents. Next, the chemicals in the database were classified into three categories (I, IIA, and IIB) based on the intermolecular interactions for humans and rats. Quantitative relationships between blood/air and tissue/blood partition coefficients and physicochemical parameters were derived for the chemicals in each category. Regression analyses of blood/air and fat/blood partition coefficients against Henry’s law constant and log D at pH 7.4 for chemicals in category IIA for humans, in which van der Waals and dipole–dipole interactions were involved, yielded 0.88 and 0.54 coefficients of determination, respectively. Moreover, these methods worked for other categories and species. The metabolic parameters maximal velocity (Vmax) and Michaelis–Menten constant (Km) of the chemicals that are primarily metabolized by cytochrome P450 were calculated for humans and rats. Multiple regression analyses of logs Vmax and Km against the occurrence frequency of molecular fragments showed good correlations, respectively. The aforementioned models predicted values close to the reported values for test chemicals within the applicability domains. Our approach could also be applied to other chemicals within the domains that are not included in the database.
Various chemicals cause toxic effects after entering the body; the seriousness of these effects depends on the amount absorbed by target tissues and the intrinsic potency of the interacting chemicals. The aim of the field of toxicokinetics (TK) is to investigate the processes of absorption, metabolization, and distribution of chemicals, their metabolites in and by tissues, and the excretion of these chemicals and metabolites from the body. Thus, TK plays an important role in chemical risk assessment (Fiserova-Bergerova, 1985). However, as obtaining the in vivo empirical data can be difficult in many chemicals, those must be complemented by data predicted through various methods.
A mathematical model for predicting TK is known as a physiologically based kinetic (PBK) model, synonymous with a physiologically based pharmacokinetic (PBPK) model or physiologically based toxicokinetics model (Lautz et al., 2019). The PBK model represents the body as a series of interconnected compartments linked via blood flow and simulates the concentration–time curve of a target organ or its alternatives (such as blood). The absorption, distribution, metabolism, and excretion (ADME) processes of chemicals between these interconnected compartments are modeled as differential equations (Paini et al., 2019). Sager et al. (2015) noted an increasing number of articles related to this topic in PubMed, indicating a growing interest in the topic. PBK models are conventionally used for extrapolating animal results to humans, one route of exposure to another, and high to low doses (Cuello et al., 2012). More recently, they have been used for extrapolating in vitro results to in vivo data. The chemical-specific parameters necessary for PBK modeling have been obtained through in vitro assays with high-throughput screening and through in silico models using in vitro data (USEPA, 2006; Kamiya et al., 2022). The high-throughput TK approaches adopted by the U.S. Environmental Protection Agency (U.S. EPA) rapidly predict the ADME parameters for TK using chemical-specific values obtained from in vitro and in silico methods (Breen et al., 2021).
Although PBK models are commonly applied to oral and intravenous exposure routes, their applications to inhalation exposure are limited. As highlighted by Borghardt et al. (2015), inhalation is a complex process involving pulmonary deposition, dissolution, and absorption. Inhalation is also an important exposure route in work environments. For instance, workers in the rubber, plastic, paint, and automobile manufacturing industries are frequently exposed to volatile organic compounds (VOCs) (Ghobakhloo et al., 2023). In the printing industry, VOCs volatilized during various processes compromise the health of workers (Leung et al., 2005). VOCs are low-boiling-point compounds represented by aromatic hydrocarbons such as benzene, toluene, xylene, and ethylbenzene, along with halogenated hydrocarbons such as chloroethylene and trichloroethylene (Xie et al., 2022). Exposure to certain concentrations of VOCs can induce headaches, nausea, and more severe symptoms such as convulsions and comas. VOCs also contain many carcinogens that damage the liver, kidneys, brain, and nervous system (Wang et al., 2020).
For some chemicals, past empirical data for human and animal experiments are available and are used to develop PBK models, but inhalation experiments on humans and animals are notoriously difficult. For instance, inhalation exposure tests of rodents are beset by technical difficulties such as special equipment requirements. Moreover, the results depend on the selected exposure method (nose-only or whole-body) (Cheng et al., 2010). In silico methods that provide direct results on humans are strongly expected to resolve these difficulties.
In our previous study, we successfully predicted the maximum blood concentration (Cmax) and area under the blood concentration–time curve (AUC) of human inhalation exposure to one chemical, ethyl tert-butyl ether (ETBE), using a PBK model (Watanabe-Matsumoto et al., 2022). In the present study, the scope of model predictions was expanded by examining the literature on PBK models of inhalation exposure and collecting the chemical-specific parameter values of more than 200 chemicals required for developing PBK models of humans and rodents (rats and mice) into a database. We then developed prediction methods of the chemical-specific parameter values using our database. This database is being published here for the first time as an open-source resource, presented in the Supplementary material of this article.
In 2006, the U.S. EPA released the report “Approaches for the Application of Physiologically Based Pharmacokinetic (PBPK) Models and Supporting Data in Risk Assessment (final report),” including an overview of PBPK modeling, the data needed to develop these models, and the considerations before applying the models (USEPA, 2006). Accompanying the report was a “List of publications relevant to PBPK modeling of environmental chemicals and its use” containing information relating to approximately 1,000 articles.
In this study, articles were selected from this list based on titles and abstracts. The physicochemical and biochemical parameters used in PBK models of approximately 200 industrial chemicals were extracted from approximately 700 different articles.
Physicochemical (including blood/air, tissue/air, and tissue/blood partition coefficients) and biochemical (metabolic parameters Vmax and Km) parameters of target chemicals were included in the database along with chemical structural information and target species (humans, rats, and mice), including citations to trace the original literature.
Prediction of partition coefficientsThe quantitative relationships between partition coefficients and physicochemical descriptors were analyzed for chemicals in the database. Chemicals were classified into four categories (I, IIA, IIB, and III) based on the type of their dominant intermolecular interactions (van der Waals force, dipole–dipole interaction, hydrogen bonding interaction, and ionic interaction, respectively) between the chemicals and biomembrane molecules during the membrane permeation (National Institute of Technology and Evaluation, 2009, 2010, 2011, 2012). In general, from weakest to strongest, intermolecular interactions can be ordered as follows: van der Waals force < dipole–dipole interaction < hydrogen bonding (Israelachvili, 2011).
In practice, the chemicals were categorized based on the number of hydrogen bond donors and acceptors, as well as log D at pH 7.4 and log P. Category I has neither hydrogen bond donors nor acceptors. Category IIA has no hydrogen bond donors and one or more acceptors. Category IIB has one or more hydrogen bond donors and the absolute value of the difference between log D at pH 7.4 and log P is less than 2. Under the latter condition, this category includes only partially or non-ionized chemicals. Category III has one or more hydrogen bond donors and the difference between log D at pH 7.4 and log P is 2 or more. Therefore, this category includes only completely ionized chemicals in water. The numbers of hydrogen bond donors and acceptors were obtained from the “Chemical and Physical Properties” section in PubChem (https://pubchem.ncbi.nlm.nih.gov/). The log D at pH 7.4 and log P values were calculated by Integrated Chemical Environment version 2.0 of the National Toxicology Program. The fluorine atom, unlike other halogens, forms a hydrogen bond with the hydrogen atom due to its very large electronegativity (like oxygen and nitrogen atoms). Therefore, it is regarded as a hydrogen bond acceptor in the “Chemical and Physical Properties” in PubChem and was treated as such in this study.
The major category members are as follows: Category I, aliphatic and non-fluorine halogenated hydrocarbons, etc.; category IIA, ethers and ketones, etc.; category IIB, alcohols and amides, etc.; and category III, carboxylic and sulfonic acids, etc. (National Institute of Technology and Evaluation, 2009, 2010, 2011, 2012).
The correlations between blood/air and tissue/blood partition coefficients and physicochemical descriptors (Henry’s law constant or log D at pH 7.4) were analyzed for chemicals in each category. If good correlation was not observed, one of two approaches was then applied. The first approach involved subcategorizing the chemicals based on their structural similarities. A linear regression analysis was then performed on the partition coefficients and physicochemical descriptors. The second approach was applied when the partition coefficients were within too narrow a range for this analysis to be performed—in such cases, the average value for the category members was used as the predicted value. Microsoft Excel 2019 and R version 4.3.1 (R Core Team, 2023) were used for linear regression analysis, and leave-one-out cross-validation (LOOCV) was used for an internal validation. If multiple values of a coefficient were listed for a chemical in the database, the modal value was used. For Henry’s law constant, the value predicted by EPI Suite version 4.11 (U.S. EPA, 2012) was used; for log D at pH 7.4, the value calculated by Integrated Chemical Environment version 2.0 of the National Toxicology Program was applied. The values of other molecular descriptors were calculated using ADMET Predictor version 10.2 (Simulations Plus, Lancaster, CA, USA). Case studies were performed in which the values of human or rat partition coefficients for test chemicals were predicted based on the results of the regression analysis. The applicability domains were defined as the ranges of descriptors that showed good correlation between the descriptors and the partition coefficients. Substances whose descriptors had values around the middle of the applicability domains were selected as test chemicals.
Prediction of Vmax and KmThe metabolic parameters (Vmaxc (Vmax normalized to body weight) and Km) in humans and rats for chemicals primarily metabolized by cytochrome P450 (P450) during phase I metabolism were predicted using the Free-Wilson approach that predicts Vmax and Km values based on the contribution to each parameter, made by a particular fragment of a chemical molecule of interest, and the number of these fragments present in the molecule. The Free-Wilson approach was used by Price and Krishnan (2011) to predict Vmax and Km for aliphatic, aromatic, and halogenated aliphatic hydrocarbons in rats. It was also applied by Watanabe-Matsumoto et al. (2022) to predict the same parameters for humans.
Vmaxc or Km values were calculated using
log M = ∑iCi∙fi ,
where M represents Vmaxc or Km, Ci is the contribution of structural fragment i to Vmaxc or Km, and fi is the number of fragments of this type present in the molecule. Similar to the experiments described in Watanabe-Matsumoto et al. (2022) study, the following fragments were selected for investigation: benzene ring (BzR), benzene ring hydrogen (H_BzR), carbon–carbon double bond (C=C), hydrogen of carbon–carbon double bond (H–C=C), methyl group (CH3), methylene group (CH2), methine group (CH), quaternary carbon (C), ether (–O–), chlorine (Cl), bromine (Br), and fluorine (F). To determine the contribution made by each fragment, chemicals consisting exclusively of one or a combination of these 12 structural fragments and undergoing initial oxidation primarily through the action of P450 were selected for further study. A multiple regression analysis was performed using log Vmaxc or Km value as the objective variable and the number of occurrences of 12 structural fragments as the explanatory variable (Watanabe-Matsumoto et al., 2022), performed using Microsoft Excel 2019 and R version 4.3.1 (R Core Team, 2023). Herein, to verify our results, we constructed several models for test chemicals. One model was created for one test chemical by calculating the contributions of fragments using training set chemicals excluding the test chemical (the number of training set chemicals: 28 and 29 chemicals for Vmaxc and Km in humans, and 30 and 33 chemicals for Vmaxc and Km in rats, respectively). Test chemicals for humans or rats were chosen from three main chemical classes in the training set: halogenated aliphatic hydrocarbons (1,1,2-trichloroethylene or dibromo(chloro)methane), aromatic hydrocarbons (styrene or toluene) and aliphatic hydrocarbons (hexane or buta-1,3-diene). The number of fragments and molecular weight of these chemicals were in the range of those of the training set chemicals.
The database consisted of chemicals, parameter values for use in PBK models concerning exposure by inhalation, and details of literature references. For each chemical, the chemical name (International Union of Pure and Applied Chemistry name), Chemical Abstracts Service (CAS) number, simplified molecular-input line-entry system notation for the structure, and molecular weight were obtained from PubChem and added to the database. The parameter values included blood/air, tissue/blood, and tissue/air partition coefficients, as well as Vmax and Km. The reference information consisted of author, title, year, journal, and PMID. The full database is available in the Supplementary material (Table S1).
The database included data for 203 chemicals; for each parameter, the number of collected chemicals and the parameter values are shown in Table 1. Among the three species, the total amount of collected data was highest for rats; however, the amount of collected human data was similar. For the physicochemical parameters, the number of chemicals for which human data were collected was 95% of that for rats, and for the biochemical parameters, the corresponding figure was 70%. Classification of 203 chemicals in the database based on their structure concluded the inclusion of 50 chlorides, 43 fluorides, 12 bromides, 36 ethers, 19 alcohols, 12 esters, 142 aliphatic hydrocarbons (compounds without a cyclic structure), and 43 aromatic hydrocarbons.
Table 2 shows two examples of structurally similar pairs of chemicals along with the partition coefficients for humans. The partition coefficients of the first pair, 1,1,1,2,2,2-hexachloroethane and 1,1,1,2,2-pentachloroethane, are similar. In contrast, the second pair, ethoxyethane and 1-methoxypropan-2-ol, have similar molecular weights, possess the same number of carbon atoms, and contain the same ether group, but their blood/air and liver/blood partition coefficients are different. These coefficients were approximately 1,000 times smaller and 600 times larger for ethoxyethane compared with those for 1-methoxypropan-2-ol, respectively. This difference may be attributable to the presence of the hydroxyl group in 1-methoxypropan-2-ol. The first pair of chemicals can be considered to be similar compounds belonging to the same group, whereas ethoxyethane and 1-methoxypropan-2-ol should be assigned to separate groups for further analysis. Similar considerations were applied to other functional groups (such as the hydroxyl and nitro groups). Thus, the chemicals were categorized based on the differences between the intermolecular interactions of functional groups or substructures that might contribute to membrane permeability and tissue distribution.
The chemicals listed in the database were classified based on intermolecular interactions. Linear regression analysis was performed between the physicochemical descriptors and partition coefficients for the chemicals in each category for humans and rats. The results indicated that for all 178 chemicals (82 chemicals in I, 76 chemicals in IIA, and 20 chemicals in IIB; category III was excluded from further analysis because it contained only few chemicals) (Table S2), there was a strong negative correlation between human blood/air partition coefficient and predicted value of Henry’s law constant. For chemicals in categories IIA and IIB, each negative correlation was strong (giving coefficients of determination of 0.88 and 0.94, respectively) within specific ranges of the logarithm of the predicted Henry’s law constant (Pa m3/mol) (−0.44 – 3.99 and −2.37 – 0.25, respectively). A scatter plot of the logarithm of blood/air partition coefficient against the logarithm of the predicted value of Henry’s law constant for the chemicals in category IIA for humans is shown in Fig. 1. The dataset is compiled in Table S3 and Script S1 of the Supplementary material. The coefficients of determination obtained by LOOCV of these categories (0.88 and 0.93 for IIA and IIB, respectively) were similar to those of linear regression (data not shown). Scatter plots and further details of the chemicals in categories I and IIB for humans are included in the supplemental material as Figs. S1 and S2, respectively; Tables S4 and S5, respectively; and Scripts S2 and S3, respectively. When a similar analysis for humans was performed using the empirical values of Henry’s law constant derived from EPI Suite, similar negative correlations were obtained (Figs. S3–S5, Tables S6–S8, and Scripts S4–S6). It was considered that the values of the coefficients could be predicted using their regression analyses within specific ranges defined as applicability domains.
Correlation between logarithm of the blood/air partition coefficient for humans and predicted value of Henry’s law constant for chemicals in category IIA of the database (denoted by open triangles); the solid triangle represents furan, which was used as a test chemical in this study.
A positive correlation (coefficient of determination obtained using linear regression analysis: 0.54) between the fat/blood partition coefficient in humans and log D at pH 7.4 was obtained for 64 chemicals in category IIA within a specific range of log D at pH 7.4 (–0.20 – 2.97) (Fig. 2, Table S9, and Script S7). The LOOCV-obtained coefficient of determination was 0.58 (data not shown). The scatter plots, chemicals in category I, and chemicals in category IIB for humans are provided in Figs. S6 and S7, Tables S10 and S11, and Scripts S8 and S9, respectively, of Supplementary material. However, for liver/blood, richly perfused tissue/blood, and poorly perfused tissue/blood partition coefficients, weak correlations with log D at pH 7.4 in each category were observed. Therefore, the chemicals were subcategorized based on their structural similarities. In the case of liver/blood partition coefficient, the positive correlation (coefficient of determination obtained using linear regression analysis: 0.67) was strong for the 10 chemicals in category IIA for humans that contain only carbon, hydrogen, and oxygen in their molecular structures (Fig. S8, Table S12, and Script S10). Additionally, the coefficient of determination obtained using LOOCV was 0.71 (data not shown). Alternatively, as liver/blood partition coefficients within a specific range of log D at pH 7.4 (−0.28 – 3.96) for the chemicals in category IIA for humans were similar irrespective of differences in molecular structure, it was considered that the values of these coefficients could be predicted using their average values (Fig. 3 and Table S13). As shown in Fig. 3, the average value of liver/blood partition coefficient for chemicals in category IIA was 1.55 (log10 1.55 = 0.19). For categories I and IIB for humans, these average values of this coefficient were 2.95 and 1.10, respectively (Figs. S9 and S10 and Tables S14 and S15). For richly perfused tissue/blood and poorly perfused tissue/blood partition coefficients in category IIA for humans, it was possible to predict the value of the coefficient adopting the average values of the coefficient within the applicability domains of log D at pH 7.4 (−0.28 – 3.96 and −0.20 – 3.96, respectively) (Figs. S11 and S12 and Tables S16 and S17). Scatter plots and further details of the chemicals in categories I and IIB for humans are presented in Figs. S13–S16 and Tables S18–S21.
Correlation between logarithm of the fat/blood partition coefficient for humans and log D at pH 7.4 for chemicals in category IIA of the database (denoted by open triangles); the solid triangle represents furan, which was used as a test chemical in this study.
Correlation between logarithm of the liver/blood partition coefficient for humans and log D at pH 7.4 for chemicals in category IIA of the database (denoted by open triangles); the solid triangle represents furan, which was used as a test chemical in this study.
After analyzing these results, a case study for humans was conducted to predict the parameter values for a test chemical, furan, based on the values for chemicals in the database. This compound belongs to category IIA and had a predicted value of 2.74 for the logarithm of Henry’s law constant and an estimated log D of 1.30 at a pH of 7.4. The values of these parameters were in the middle of the ranges found in the datasets, and good predictability and high reliability were expected using a quantitative structure–property relationship (QSPR) and average values and cases in which a good correlation with blood/air and the fat/blood partition coefficients was found for humans (Figs. 1 and 2). Therefore, the values of these two coefficients for furan were predicted based on QSPR. The value of liver/blood partition coefficient was obtained by applying this approach to data available for 10 similar chemicals consisting of carbon, hydrogen, and oxygen atoms and by taking the average of values for the chemicals in categories IIA for humans. For richly perfused tissue/blood and poorly perfused tissue/blood coefficients, values were obtained by averaging the values for the chemicals in category IIA, without using QSPR. The predicted values of the partition coefficients were within 95% confidence intervals and close to those previously reported in the literature (Table 3) (Kedderis and Held, 1996).
The parameter predictions described above were also conducted for other categories and species. For rats, we predicted the log blood/air and log fat/blood partition coefficients for benzene (category I), 4-methylpentan-2-one (category IIA), and 2-propan-2-yloxyethanol (category IIB). The predicted values of the log blood/air partition coefficients were 1.47 (benzene), 2.10 (4-methylpentan-2-one), and 4.98 (2-propan-2-yloxyethanol), whereas the reported values are 1.18, 1.90, and 4.06, respectively (Bois et al., 1991; Meulenberg and Vijverberg, 2000). The predicted values of the log fat/blood partition coefficients were 1.36 (benzene), 0.84 (4-methylpentan-2-one), and −0.57 (2-propan-2-yloxyethanol), whereas the reported values are 1.52, 0.82, and −0.76, respectively (Bois et al., 1991; Meulenberg and Vijverberg, 2000). As all predicted values were close to the reported ones, these results show the effectiveness of our prediction method regardless of the category or species. Further details of partition coefficients in rats are presented in Figs. S17–S31, Tables S22–S36, and Scripts S11–S16 of the Supplementary material.
The applicability domains of the models for humans and rats to predict partition coefficients from these descriptors are summarized in Table S37.
Use of the database for predictions of Vmax and KmMetabolic parameters are often expressed by Vmax or Km. For Vmaxc of 28 chemicals for humans in the database, containing the types of structural fragments described in “Prediction of Vmax and Km” in the “MATERIALS AND METHODS” section, the range of adjusted coefficients of determination of 0.790 – 0.800 was obtained for the multiple regression analyses of log Vmaxc as the objective and number of fragments of 12 structural fragments as explanatory variables; the contribution to the parameter value made by each type of fragment was also determined. A similar analysis was performed for Km of 29 chemicals for humans in the database; in this case, the range of adjusted coefficient of determination was 0.565 – 0.590. In most cases, the ratio of predicted value to the reported value for the two metabolic parameters was within 0.1 – 10. Similar results were obtained in rats as in humans. The contributions and number of each type of structural fragment together with an analysis of predicted and reported values are given as the Supplementary material (Figs. S32–S43, Tables S38–S73 and Scripts S17–S28).
Chemicals used to calculate the contribution of each structural fragment mainly included halogenated aliphatic hydrocarbons, aromatic hydrocarbons, and aliphatic hydrocarbons. As a first case study for humans, 1,1,2-trichloroethylene was used to test predictions of log Vmaxc and log Km values. This chemical contains 3 (1 carbon–carbon double bond, 1 hydrogen of carbon–carbon double bond, and 3 chlorine atoms) of the 12 structural fragments described earlier, and according to information available in PubChem, it is primarily metabolized by P450 (Hibino et al., 2013). The values for this chemical were predicted using the Free-Wilson approach. The predicted values of each parameter were close to the reported values, and the ratios of the predicted to reported values were within 0.1 – 10 (Table 4 and Figs. S32 and S35) (Fisher, 2000).
Then, Vmaxc and Km were predicted for styrene and hexane. The ratios of the predicted values to reported values were within 0.1 – 10 (Table 4, Fig. S33, S34, S36, and S37) (Fisher et al., 1997; Ali and Tardif, 1999). Furthermore, similar calculations were performed for rats with dibromo(chloro)methane, toluene, and buta-1,3-diene; the ranges of adjusted coefficients of determination for the multiple regression analyses of log Vmaxc and Km were 0.917 – 0.918 and 0.601 – 0.617, respectively. The ratios of the predicted to reported values were within 0.1 – 10 (Table 4, Fig. S38–S43) (Luciene da Silva et al., 1999; Tardif et al., 1993; DeJongh et al., 1998; Kohn and Melnick, 1993).
Some chemicals in the database have multiple reported values of Vmaxc and Km for humans or rats, and this variation is as high as a factor of 20 for 50%–90% of the chemicals listed. Thus, ratios of 0.1 – 10 for the predicted to reported values are reasonable.
We compiled a database of PBK parameters including the physicochemical (blood/air, tissue/blood, and tissue/air partition coefficients) and metabolic (Vmax and Km) parameters of industrial chemicals whose inhalation exposure is a concern. The parameter values were collected from the literature on PBK models of humans, rats, and mice. Most chemicals in the database are inhalable VOCs and relevant to industrial health and TK.
Concerning TK of nonpharmaceutical chemicals, Sayre et al. (2020) published a database primarily consisting in vivo pharmacokinetic parameters (changes in blood concentration, Cmax, T1/2, etc.) for different administration routes (oral, inhalation, etc.) and species (humans, mice, rats, etc.) for 144 chemicals; PBK parameters were included for some of these chemicals. On comparing the two databases, 42 chemicals appeared in both. The lack of overlap between the databases may be because they were compiled for different purposes. By constructing the database in this study, additional chemical information was compiled.
Most of the chemical-specific parameter values for the rat PBK models that were collected in this study were measured values, whereas some of the parameter values used in human PBK model were estimated from rat values. For instance, Koizumi measured rat tissue/blood and human blood/air partition coefficients for trichloroethylene using the vial equilibrium method and calculated human tissue/blood partition coefficient by dividing rat tissue/air partition coefficient by human blood/air partition coefficient (Sato and Nakajima, 1979; Koizumi, 1989). This was possibly due to difficulty in obtaining human tissue. The differences between these methods should be taken into account depending on the purpose of use of the database.
Partition parameter values can differ considerably even for chemicals with similar molecular structures (Table 2). Therefore, analyzing the relationship between the descriptors and properties based on chemical categories to develop methods of predicting partition parameters is helpful. In this study, we investigated the relationship between descriptors and partition coefficients by classifying chemicals in the database based on molecular interactions possibly affecting partition to blood or tissue.
Figure 1 shows a negative correlation between Henry’s law constant, an air/water partition index, and human blood/air partition coefficient. This result is reasonable owing to the definition of Henry’s law constant and the composition of blood, which primarily comprises water (91%–92%) other than blood cells and platelet components (Mathew et al., 2023). Most chemicals belonging to categories I (e.g., aliphatic and non-fluorine halogenated hydrocarbons) and IIA (e.g., ether and ketone), with weak intermolecular interactions, tended to show large values of Henry’s law constant (thus, a relatively small blood/air partition coefficient), whereas most chemicals in category IIB (e.g., alcohols and amides), with strong intermolecular interactions, tended to show a small Henry’s law constant (a relatively large blood/air partition coefficient) (Figs. 1, S1, and S2; Tables S3–S5; and Scripts S1–S3). Therefore, by classifying chemicals based on intermolecular interactions and using the negative correlation between the blood/air partition coefficients and Henry’s law constant, one can predict the blood/air partition coefficient of a target chemical. Because the chemicals in categories I and IIA tend to be highly volatile, they are probably less soluble in water-soluble mucosal components such as the alveolar lining; therefore, they can easily reach the deep alveoli; some chemicals would enter the bloodstream and circulate throughout the body. Because of the strong intermolecular interactions present in the chemicals in category IIB, the volatility of these chemicals is low. Once these are inhaled and reach the deep alveoli, they would rapidly enter the bloodstream and circulate throughout the body.
Kramer et al. (2016) predicted the human blood/air partition coefficients based on the molecular structural information of 90 VOCs (coefficient of determination between the predicted and empirical data = 0.95). Wang and Xu (2020) developed a software-based multiple regression model for predicting the human blood/air partition coefficients of 143 VOCs (coefficient of determination between the predicted and empirical data = 0.79–0.92). In this study, the human blood/air partition coefficients were predicted by a rather simple method that linear regresses each category using the Henry's law constants calculated by EPI Suite (Figs. 1, S1, and S2; Tables S3–S5, Scripts S1–S3), but which nonetheless provides a similar correlation against the empirical values (coefficient of determination = 0.81; data not shown) as the above two reports. This simple method can be employed in a screening-level assessment process, and it is also easy to improve the models by adding newly acquired data.
Figure 2 shows that for chemicals in category IIA, there is a positive correlation between human fat/blood partition coefficient and value of log D at pH 7.4. This value is related to the lipophilicity of chemical at pH 7.4. Therefore, reasonably, a positive correlation between this descriptor and the fat/blood partition coefficient should exist, suggesting the value of fat/blood partition coefficient can be predicted based on this quantitative relationship.
However, according to Figs. 1 and 2, the correlation between human blood/air partition coefficient and Henry’s law constant is weak when logarithm of the constant is ≥4 (Pa m3/mol) (Fig. 1); the correlation between human fat/blood partition coefficient and log D at pH 7.4 becomes weak when log D is ≥3 (Fig. 2). Therefore, to make accurate predictions of partition coefficients of a target chemical using the QSPR approach, the values of the descriptors should be within the strong correlation range. In fact, the guidance contained in the (Quantitative) Structure–Activity Relationship ((Q)SAR) Assessment Framework of the Organisation for Economic Co-operation and Development emphasizes that the descriptor values of test chemicals should be within the relevant ranges (Organisation for Economic Co-operation and Development, 2023). Therefore, in our case studies, we chose test chemicals with coefficient values that lay in the middle of these ranges for each dataset; as expected, these chemicals exhibited good predictability.
No strong correlation was found between tissue/blood partition coefficient and log D at pH 7.4. However, the correlation between human liver/blood partition coefficient and log D at pH 7.4 was high for structurally similar chemicals in category IIA (Fig. S8). This implies that predictions of tissue/blood partition coefficient could be made by applying QSPR based on a dataset consisting of chemicals structurally similar to the target chemical through subcategorization. Alternatively, as human tissue/blood partition coefficient values were found to occupy a relatively narrow range despite structural differences between the chemicals studied, the value of this parameter could be predicted by averaging known parameter values (Figs. 3 and S9–S16). Similar trends were observed in rats as in humans (Figs. S17–S31). These results indicated certain relationships between physicochemical property values and partition coefficients.
In this study, we showed that our new prediction methods work for other categories and species. Along with the database published here, these predictions of inhalation exposure parameters will improve our ability to analyze various chemicals.
The log Vmaxc and log Km values were predicted using a group-contribution method called the Free-Wilson approach. The same approach was adopted by Price and Krishnan (2011), who determined the relative contributions of 11 molecular structural fragments in 53 VOCs to the Vmax and Km values of rats and developed a model for predicting the TK of a 10-VOC mixture following inhalation exposure, and by Watanabe-Matsumoto et al. (2022), who predicted the Vmax and Km of ETBE in humans after inhalation exposure.
Moreover, using this approach, we constructed models for structurally diverse test chemicals and verified their predictivity. The ratios of the calculated to reported values were found within 0.1 – 10, implying good predictions (Table 4). These results suggested that the Free-Wilson approach could be applied to other chemicals that fulfill the requirements of our training set (“Prediction of Vmax and Km” in the “MATERIALS AND METHODS”). To clarify the range of fragment type and number for good prediction, further studies on more chemicals are needed.
Watanabe-Matsumoto et al. (2022) performed a preliminary study on one chemical from a small dataset with the aim of predicting TK parameters and used them for construction of the appropriate PBK model. This study used an expanded dataset of 203 chemicals and further demonstrated the applicability of PBK parameter prediction. Our results are expected to contribute to the construction and validation of PBK models in the future.
Read-across is a method of grouping similar chemicals with available toxicity data and predicting the toxicity of a target chemical for which these data are not available. It is crucial to group chemicals based on the similarity of toxicity mechanisms and chemical structure. When predicting systemic toxicity, grouping based on TK similarity should be considered to increase the confidence of predictions (Schultz et al., 2019). The parameter values in the database developed in this study could be used as the basis of this grouping.
Altogether, herein, a database of chemical-specific PBK parameters was developed for inhalation chemicals. Based on these data, parameter values needed for the PBK modeling of chemicals were successfully predicted. This approach can potentially predict the parameters of other VOCs not included in the database. Moreover, a PBK model constructed with the predicted parameters would be able to predict the TKs of inhalation chemicals lacking data.
The authors acknowledge the assistance of Yuriko Meiseki with data input for the database. The authors also acknowledge the assistance of Taeko Maruyama-Komoda, with whom they had fruitful discussions and who provided valuable comments. This work was supported by Health and Labour Sciences Research Grants (H30-Chemistry Destination-005 and 21KD2005) from the Ministry of Health, Labour and Welfare of Japan.
Conflict of interestThe authors declare that there is no conflict of interest.