Insights into In Vitro Biokinetics Using Virtual Cell Based Assay Simulations

The Virtual Cell Based Assay (VCBA) is an in silico model that simulates the biokinetics of chemicals in in vitro test systems. VCBA simulations can indicate the degree to which the bioavailable concentration varies across chemicals and experimental conditions, thereby providing important contextual information for comparing the results of different in vitro toxicity experiments. The simulated results can also be used to support in vitro to in vivo extrapolation of toxicity data, especially when the VCBA is coupled to a physiologically based kinetic model. In this work, we selected 83 chemicals previously tested for in vitro cytotoxicity with a neutral red uptake (NRU) assay and used the respective in vitro data to optimize a toxicity and effects model simulating the 3T3 BALB/c cell line in a 96-well microplate with 5% serum supplementation. We then used the optimized parameters to simulate alternative experimental conditions. The simulations show the impact of different physicochemical properties on chemical fate of this diverse group of chemicals and how the different partitioning (to protein, lipid, and plastic) and kinetic (evaporation and degradation) events are intrinsically connected. The results of VCBA simulations were interpreted with respect to the applicability domain of the different QSARs incorporated in the model and the underlying assumptions and uncertainties


Introduction
The increasing need to test chemicals for risk assessment and the ethical and scientific concerns over animal experimentation have brought in vitro models into the limelight.They have been proposed as a more efficient and economical model that is compliant with the 3Rs principles (replacement, reduction, and refinement of animal testing).
Considering that toxicological events are mostly initiated at the cellular level (Ekwall, 1983;Schirmer, 2006), in vitro methods are not only relevant to screen for these events but also to better understand their mechanisms of action.Such information may help to identify other chemicals with similar mechanisms.In vitro methods encompass a wide range of cell types, culture methods, and experimental conditions.The chemical concentrations tested in these models represent concentrations that are bioavailable to the organ and not actual exposure doses.Therefore, to extrapolate observed in vitro toxic doses to respective in vivo exposure doses in a quantitative way (QIVIVE), in silico mod-Therefore, a better approximation of in vivo toxic concentrations may be obtained by modulating these in vitro biokinetic processes to determine the concentration that is effectively dissolved in the exposure medium and is unbound (free concentration).Indeed, chemicals that have higher cytotoxic potencies are affected more strongly by these biokinetic events than chemicals that have lower cytotoxicity potencies, as the amount of non-bioavailable chemical can easily surpass the bioavailable amount.In the case of less toxic chemicals, the impact of serum binding may be negligible if the nominal toxic concentration exceeds the binding capacity of the serum proteins (Gülden and Seibert, 2005).Heringa et al. (2004) showed that while nominal air, is an environment that is drastically different from the in vivo microenvironment of cells.Indeed, binding of chemicals to serum proteins and lipids, plastic, and evaporation (Kramer, 2010;Stadnicka-Michalak et al., 2014;Seibert et al., 2002;deBruyn and Gobas, 2007), have been broadly described to considerably impact the actual chemical concentration in in vitro systems.Moreover, the most commonly used dose metric in in vitro systems is the added (nominal) concentration, which does not reflect the amount of chemical per number of cells in the system.Gülden et al. (2001) showed that free concentration and toxic effects change with cell number, i.e. higher cell numbers in culture have higher EC 50 values.interconnected models of the VCBA are reported in (Zaldívar Comenges et al., 2017).More information on the VCBA can be found in (Paini et al., 2017b;Graepel et al., 2017;Sala Benito et al., 2017;Worth et al., 2017).
The overall goal of this work was to apply the VCBA to a large set of chemicals, for which in vitro acute oral toxicity data were already available, to clarify which experimental conditions and physicochemical properties have the greatest impact on in vitro chemical fate.A subset of 83 chemicals tested in international projects and validation studies (NICEATM/ECVAM validation study (NIH, 2006); the PF6 EU project ACuteTox 2 (Prieto et al., 2013a); ECVAM validation study (Prieto et al., 2013b)) was chosen.The complete list of chemicals used in these projects can be found in the EC-JRC Chemical Lists Information System, CheList 3 .In all of these projects, the BALB/c 3T3 cell line was used and cytotoxicity was measured with a neutral red uptake (NRU) assay.Hence, the VCBA model was set to simulate the conditions of the NRU assay with some experimental variability to characterize the uncertainties of the VCBA core model, and to estimate the dissolved or intracellular chemical concentration that can be used for different in vitro comparisons and as a point of departure for QIVIVE.

Estimation of chemical parameters
The following physicochemical parameters are required to run the VCBA: 1.Molecular weight (MW; g/mol); 2. Molecular diffusion volume (SVcomp, indicated as atomic diffusion in Zaldívar Comenges et al. ( 2017)) 3. Molar volume (MV; cm 3 /mol); 4. Henry law constant (HLC; Pa×m 3 /mol); 5. Degradation rates in water and air (s -1 ); 6.The logarithm of the octanol-water partition coefficient (log-K ow ) (unitless).The physicochemical parameters of the 83 chemicals were collected between July 2016 and October 2016.Molecular weight was retrieved from the EPA Chemical Dashboard 4 , and water and air degradation rates were derived from the half-life reported by EPI Suite™ (US Environmental Protection Agency) by applying the following formula: k deg = ln/t (1/2) .Molar volume was retrieved from ACD/Labs 5 software predictions retrieved from Chemspider 6 , and when not available it was extrapolated from the MW through a simple linear equation, as indicated in Proença et al. (2017).Molecular diffusion volume (dimensionless), required for fugacity prediction, was calculated by the Fuller semi-empirical method as the sum of the specific atomic diffusion volumes, discounting the volume of each aromatic/heterocyclic ring (Fuller EC 50 values depend on serum percentages, the freely available EC 50 values are roughly the same.Ideally, this in vitro biokinetics translation of nominal to free concentration of chemicals can be integrated into PBK models (Gajewska et al., 2015;Paini et al., 2017a) and quantitative adverse outcome pathways (qAOPs) (Madureira et al., 2014).
The Virtual Cell Based Assay (VCBA) is an in silico model developed as part of the EU FP7 COSMOS project 1 (SEUR-AT-1).It simulates in vitro chemical fate and intracellular concentrations to determine the free concentration that is required to cause perturbations in cells (Zaldívar Comenges et al., 2010, 2011).Briefly, to run the VCBA, specific input parameters for chemicals, cell types, and experimental set-up are needed.The VCBA consists of 4 interconnected models (Fig. 1): 1 Fate and transport model, which is based on each chemical's physicochemical properties.It describes the dynamic mass balance of a given chemical, including its partitioning between headspace (gas exchange equations), plastic, and serum lipid (deBruyn and Gobas, 2007;Kramer, 2010), as well as chemical degradation; 2 Cell partitioning model, which describes cell uptake/excretion, and the intracellular partitioning between lipid, protein, and aqueous fractions, which depends on the chemical characteristics and cell type composition; 3 Cell growth and division model, which is simulated by a 4-stage (G1, S, G2, and M cell cycle phases) approach using a Leslie matrix.Depending on the cell type, cell growth and division can be adjusted (Paini et al., 2017b).4 Toxicity and effects model, which merges the experimental in vitro concentration-response curve with the cell growth and division model to fit the toxicity parameters, i.e., the no-effect concentration (NEC) and killing rate (Kr).Briefly, VCBA simulates the cell growth for all nominal concentrations of the experimental concentration-response curve and it calculates the intracellular concentration for each nominal concentration.First, a random threshold is used at which the NEC starts to cause cell death (Kr), hence obtaining the percentage of viability for all nominal concentrations.Then, in order to minimize the error between the experimental and the predicted concentration-response curve, the sum of errors is calculated.
Next, the NEC and Kr, are changed to create another concentration-response curve and the error is calculated.Finally, using the optimum function in R (general-purpose gradient optimization function) and following the optimization approach described by Zaldívar and Baraibar (2011), NEC and Kr are changed to decrease the sum of errors.Further, the VCBA considers the experimental set-up, which includes the well shape and size, the volume of medium, and the amount of supplemented serum and, thus, protein and lipid content in the media.The mathematical equations describing the four diction-model development.In the absence of such predictions, the average of the other predictive tools in Chemspider was used.

Cell line parameters
The values used (Tab. 1) were taken from our previously published work (Zaldívar Comenges et al., 2017) except for the intracellular protein concentration, which was recalculated and set to 4.4 mol/m 3 , based on protein density, protein fraction in the cell, and cell volume.
The cell division rate parameter does not consider cellcell contact growth inhibition; this cell line reaches confluence at 50,000 cells/cm 2 , after which the cells halt their growth.However, the initial cell numbers and assay time used describe a non-confluency scenario.

Experimental set-up
The experimental parameters, such as the different well-plates and serum percentages used, are illustrated in Table 2.These conditions were based on those used in the scope of the ACute-et al., 1966ACute-et al., , 1969)).For chemicals requiring atom diffusion volumes, and for which the atom increments were not available, an extrapolation from MW was also made.A local sensitivity analysis of the VCBA showed that variations of the atom diffusion volumes and the molar diffusion only slightly impact the predicted unbound medium concentration results, thus there is no need for highly accurate parameters (Proença et al., 2017).
Both experimental and predicted values of the HLC were available in both EPIsuite and Chemical Dashboard.Priority was given to experimental values; if more than one value was found, the average was used.When no experimental data was available, priority was given to EPIsuite predictions, first the Group Method, then the Bond Method, and finally Chemical Dashboard, which uses several prediction models and presents the average value.
LogK ow values were researched in both Chemspider (Royal Society of Chemistry) and Chemical Dashboard, again giving priority to experimental values, and using the average when suitable.When predicted values had to be used, priority was given to Chemical Dashboard, which uses a revised database in pre- trol response (which is included as a response of 100% at 0 µM).The experimental set-up that was used to obtain this data was based on a 48-hour assay with 5% serum and in a 96-well plate.
After the NEC and Kr were optimized by minimizing the residuals between the experimental and theoretical response curves, they were used within the model to further study the impact of different set-ups on the concentration of chemical in the various compartments, by keeping them constant but changing the experimental set-up parameters.
The code 8 was developed using R software, version 3.3.1,with cell growth given in hours and differential equations solved using a discretization in seconds.This specific time was chosen to speed up optimization, as cell growth/death in minutes would require VCBA to run more cycles, slowing down the optimization process.Moreover, since cell cycle phases take several hours rather than minutes, running the cycle in minutes would Tox project for 3T3 cells: initial number of 1680 cells, 48 hours in a 96-well plate with 10% FBS in 100 µl medium (DB-ALM protocol 139 7 ).Then, the volume of medium and cell number in the different in vitro set-ups were selected in proportion to the surface area.The densities of the phases were set to: water = 1000g/l, proteins = 1350 g/l, and lipids = 900 g/l, as previously reported in Zaldívar Comenges et al. (2017).
Optimization and running of the VCBA NEC and Kr are both chemical-and cell line-specific, and were obtained using the experimental results, as explained above in the introduction to the VCBA model.
Optimization of NEC and Kr was done by calculating the partitioning of each chemical and integrating it with the experimental concentration-response curves from the ACuteTox project, consisting of eight concentrations and a value linked to the con-  ing to the cut-off value in Equation 3, less than 10% of the chemical was unbound in the medium for the 14 chemicals with a logK ow higher than or equal to 4.46 (phenantrene has a logK ow of 4.46).For different logK ow ranges, different elements of the in vitro system will be the main sequesters of the chemical.Observing Figure 3 from bottom to top, from 1.88 (triethylene glycol dimethacrylate) to higher logK ow values, the unbound amount decreases as the amount of protein-bound chemical increases until logK ow 3.83 (endosulfan), where lipid gradually binds more chemical, eventually sequestering chemical from protein binding.Although there is significant binding to plastic, it is never more than 18%.
Regarding evaporation, the VCBA model indicates that only for the two chemicals with higher HLC values, benzene (HLC = 562 Pa×m 3 /mol) and xylene (HLC = 525 Pa×m 3 /mol), the percentage of chemical that evaporates is greater than 10%.In contrast, for 1,2,4-trichlorobenzene, with an HLC of 144 Pa×m 3 /mol, only 1% is predicted to evaporate.
Although chemical partitioning has a high impact on its unbound concentration after 48 hours, degradation also plays a role and is considered in the VCBA model.The ratio between nominal (total initial) amount and final amount (sum of the amount of chemical that is unbound, bound to protein, lipids, plastic, in headspace and intracellular after the 48 h assay period) was assumed to be an indication of the chemical's effective degradation: Nominal Amount The percentage of degraded chemical and related descriptors are shown in Table 3.In spite of both chemicals not having especially high degradation rates, xylene stands out with the highest degradation (38%), followed by ethanol (14%).Since the degradation rate is in general higher in air than in water, volatile compounds will tend to degrade faster, which highlights the importance of the HLC as an indicator of in vitro fate.It is also noteworthy that the overall degradation of a chemical does not depend solely on specific degradation rates but also on its partitioning behavior.The percentage degradation has no apparent correlation with air degradation, while for water degradation there is a trend.These degradation rates will only affect chemicals that are unbound, either in the aqueous or the gas phase.Chemical bound to protein, lipid, or plastic is considered to be "protected" from degradation in the VCBA.Analyzing all these parameters together (Tab.3) using a color scale pattern shows that the actual degradation of chemicals in an in vitro experiment is a multifactorial process.

Influence of the experimental set-up on chemical partitioning
As binding to plastic, extracellular lipid and protein affects the amount of unbound chemical so greatly, it was explored how different in vitro set-ups further influence the partitioning behavior.
Sixteen chemicals from the initial set of 83 were selected to cover a wide range of logK ow values (from dimethyldioctadecylammonium chloride with 8.392 to ammonium chloride with -2.68), not necessarily better resemble the physiological growth.The R code for running the simulations has been included into a KNIME workflow to automate the model and make it easier to use without the need for programming skills.
Statistical analysis R studio and GraphPad Prism 3.0 (San Diego, CA) were used to plot and analyze the data.

Results
The 83 substansces with a range of physicochemical properties (logK ow = -4.5 to 10.3; HLC = 9.66×10 -34 to 1.74×10 3 ; and MW = 41 to 823) were modelled in the VCBA, with their NEC and Kr optimized against the experimental concentration-response curves, as presented in the supplementary file 9 .Figure 2 illustrates, for 1,2,3,4-tetrachlorobenzene and diethyl phthalate, how the VCBA fits the toxicological parameters against the experimental data, and how it models chemical fate and the effect of several concentrations on cell number.Although the fitting was optimal for some chemicals (the lowest sum of the residual squares was 15 for glycerol), the error was much higher for others (the lowest sum of the residual squares was 5313 for benz(a) anthracene).
Modelling the relation between cell number and chemical concentration (Fig. 2), accounting for both growth and toxicity effects, best illustrates how viability decreases with time.The plot for intracellular concentration of diethyl phthalate (Fig. 2) indicates a virtually immediate cell uptake, in which the slight decrease over time is accompanied by a decrease also in dissolved concentration.The partitioning of chemical to plastic, serum lipid and protein is also instantaneous (not shown here); hence these events are not the cause of the decrease of concentration over time.Since both 3T3 BALB/c cells and the VC-BA model do not include chemical metabolism, the causes of this decrease are evaporation and/or degradation.In fact, after 48 hours exposure there was a loss of 14% of the chemical mass due to diethyl phthalate degradation and less than 0.1% mass in headspace.

Influence of logK ow and HLC in chemical partitioning
The binding of chemicals to lipids, protein and plastic is modelled as achieving equilibrium instantaneously, and is thus not considered a rate-limiting step; however, it does affect the bioavailable/unbound concentration during the whole assay duration as lipids, protein and plastic act as chemical reservoirs.Figure 3 shows the extracellular partitioning after 48 hours, where the nominal concentration is the experimentally obtained EC 50 value, hence a concentration that induces cell death.Even though the model requires additional chemical descriptors, which vary widely among the chemicals, this influence is clearly observed by sorting the chemicals from the highest to the lowest logK ow and plotting their modelled partitioning (Fig. 3).Accord-  VCBA simulates that the lipid content of dead cells is released into the medium.It is evident for the same chemicals that the 6-well plate, followed by the 24-well plate, is the plate configuration with the highest unbound chemical concentration, possibly due to higher plastic binding found for the 96-well plate (results not shown).
Figure 4B shows a lower percentage of chemical within the cells for the 0% FBS condition than for 5% and 10%.Because this percentage reflects the amount in all cells, this means more chemical is being taken up by the cells, inducing toxicity, and thus decreasing the number of cells.In fact, the concentration of chemical in the few live cells remaining in the 0% FBS condition is higher.
Although the VCBA simulates a very low percentage of unbound chemical for highly lipophilic chemicals, lipophilicity also determines uptake and bioaccumulation.Therefore, high-including the two chemicals with significant evaporation (benzene and xylene).By using the previously optimized NEC and Kr values, the compounds' fates were simulated for experiments in 96, 24, and 6-well plates with 0, 5, and 10% serum.
Xylene had the highest percentage of degradation (Fig. 4B) for all in vitro set-ups.However, a higher % of serum does reduce degradation by decreasing the amount of free chemical in medium or in headspace.Figure 4A also shows the impact of the different percentages of serum supplementation and the microplate configuration on the final unbound amount of chemical in medium relative to the nominal amount.As expected, less serum is related to a higher unbound amount of chemical.However, for highly lipophilic compounds (logK ow ≥ 5.76), there is virtually no unbound fraction in the medium, even with 0% serum, mainly due to plastic and lipid binding.It is noteworthy that lipids present in the 0% serum condition are derived from dead cells, as the The VCBA is a flexible model that is capable of adjusting to the variety of experimental set-ups used and facilitates data integration and interpretation.The impact of several biokinetic processes on the chemical concentrations at the end of a 48-hour assay was studied here.The model predictions are a direct result of the model equations for the three main partition constants for lipid, protein, and plastic binding, which are the following: Plastic partition constant: Kp = 10 (0.97*logKow-6.94) (2) Protein partition constant: Ks < -10 (vals-1.178)-1.31, if logKow < 1.09 vals = 0.57×logKow + 0.69, if 1.09 ≤ logKow ≤ 4.6 (3) logKow -1.3, if logKow > 4.6 Lipid partition constant: Kl < -10 (1.25*logKow-3.70)(4) The binding processes to plastic, protein and lipid will indirectly compete by removing dissolved unbound chemical from the medium.LogK ow is a common descriptor of unspecific binding in chemical-biological systems (Caron and Ermondi, 2008) and is er cell permeability compensates to some extent for the lack of bioavailable chemical, as observed in Figure 5A.In fact, by dividing the chemicals into those with logK ow < 2.5 (hydrophilic) and those with higher values (lipophilic), a different relation of intracellular to unbound amount becomes apparent (Fig. 5B).Chemicals with logK ow < 2.5 show a linear correlation, while chemicals with logK ow > 2.5 have more scattered intracellular amounts of chemical.Still, for the same unbound amount, lipophilic chemicals (logK ow > 2.5) show a higher intracellular amount than less lipophilic chemicals.
The more lipophilic the chemical, the more the intracellular amount of chemical per cell varies between experimental set-ups; decreasing with higher serum concentrations but with little difference between the different well geometries.For a few of the most lipophilic chemicals, the intracellular concentration increased such that 0% FBS was enough to kill all the cells (with NEC and Kr previously optimized).Evaporation also seems to play a role for intracellular concentration, as both benzene and xylene show the lowest intracellular amounts in the 6-well plate configuration.

Discussion
VCBA simulations for 83 chemicals were optimized against experimental concentration-response curves.While the VCBA was able to fit the experimental data approximately for most of these chemicals, resulting in small errors, the fit was poor for some chemicals.This was due to their experimental concentration-response curves deviating from the classical form (an inverted logistic sigmoidal curve).Accordingly, when choosing the chemical set, five chemicals were excluded for having concentration-response curves that were considered inadequate { cal in the headspace (at 5% serum).Several reports (ISO, 2006;OECD, 2000) have indicated that chemicals with 1 Pa×m 3 /mol are likely to evaporate, while chemicals with 100 Pa×m 3 /mol (which would include 1,2,4-trichlorobenzene) are considered highly volatile.Therefore, we expected higher percentages in the headspace for 1,2,3,4-tetrachlorobenzene, pentachlorobenzene, diethyl phthalate, p-benzoquinone, endosulfan, phenanthrene, ethyl chloroacetate, acetonitrile, benzaldehyde, tetramethylthiuram monosulphide, tert-butyl hydroperoxide, benz(a)anthracene, pyrene, and acetophenone, while, in fact, modelling predicted that less than 0.8% of any of these chemicals would evaporate.Furthermore, since evaporation is linked to degradation, the concentrations of volatile chemicals can drop even further during the 48-hour incubation.High evaporation and degradation was found for xylene and benzene in the 6-well plate owing to its higher ratio of headspace to medium volume.For non-volatile chemicals, the 24-well plate seems to retain the highest concentration of unbound chemical in the medium as there is less plastic binding (clearer in 0% serum conditions), which is related to a relatively lower ratio of plastic surface exposed to medium volume.When choosing the conditions to be simulated, care was taken to maintain the proportions across the experimental set-ups, but differences were still found among the microplate geometries.Hence, when comparing experimental conditions that do not maintain proportions, even greater differences would be expected between the different microplates.
In the current version of the VCBA, cell uptake of chemicals is based entirely on passive diffusion; active transport is not considered.This uptake rate is based on the specific cell surface and the permeability equation suggested by Yazdanian et al. (1998): Log permeability = −1.1711+ 0.98 logK ow − 0.0011MW (5) This equation considers the molecular size and the lipophilicity (logK ow ) of the molecule, which constitute two of the five rules of Lipinski (Lipinski et al., 1997), as essential parameters.The contribution of logK ow to the uptake will compensate, to some extent, the lower concentration of unbound chemical in the medium.Other chemical characteristics, such as the van der Waals surface area, also influence cell uptake.However, calculating these requires powerful computational methods and therefore, for the sake of a higher throughput model, they were not considered here (Yazdanian et al., 1998).As no distinction was made between neutral and charged chemicals, the uptake of charged compounds at pH 7.4 will be overpredicted.
The global applicability of the VCBA, which allows any chemical to be simulated as long as the physicochemical characteristics are known, is based on several assumptions that we briefly highlight here.Firstly, the QSARs used in the VCBA to predict partition coefficients for lipids, serum, and plastic were based mostly on logK ow .Other physicochemical properties that are important for fate, such as charge and whether the chemical is an H-donor/H-acceptor, were not included (Goss and Schwarzenbach, 2001).Besides, as described below, these QSARs were derived from specific sets of chemicals.
For plastic, Kramer (2010) analyzed seven polycyclic aromatic hydrocarbons (PAHs) with logK ow values ranging from 3.33 therefore frequently used in QSARs, including those used in the VCBA.The greater contribution of logK ow to lipid binding than to plastic binding will always dictate a higher binding constant to lipids.Similarly, for proteins, the relationship between logK ow and protein binding depends on the logK ow value.The binding constant for protein is always higher than for plastic and lipids up to a logK ow value of about 6.For very lipophilic compounds, the binding constants for lipids and proteins are similar.The total binding to each element depends on this binding constant but also on the concentration of proteins and lipids in the medium and on the plastic surface area.
It is noteworthy that although the lipids in medium with 0% serum should be one eighth of the lipids in 5% serum after 48 hours, they still have a high binding capacity, as observed in Figure 4.However, two considerations indicate that the 0% serum plot is most likely far from reality.Firstly, at 0% serum, most cell lines halt growth and thus the fraction of lipids in the medium would be expected to be even smaller.Secondly, saturation is not included in the simulation.Saturation is not a straightforward phenomenon for lipids, especially because the interaction of lipids with chemicals occurs in the form of aggregates rather than individually (Balaz, 2009).Modelling saturation could furthermore increase the differences among the different percentages of serum supplementation.Although the extent of lipid binding should alert to careful consideration of the chemical's kinetics, high lipid binding may also indicate a mode of toxicity (Armitage et al., 2014) based on narcosis, which can also occur via external cell effects, without complete cell uptake.
While partitioning to plastic does seem (experimentally) to change with nominal concentration (Kramer, 2010), saturation of this process is very likely and was not modelled.Indeed, Wilmes et al. (2013) showed evidence of plastic binding saturation for cyclosporin A. Likewise, protein and lipid binding saturation was also not modelled, and therefore these VCBA estimations are representative of the maximum concentration which will partition to non-aqueous phases.To include saturation for these phases, additional experimental measurements would have to be performed, such as determination of the maximum number of binding sites on serum protein, or the maximum concentration of chemical that can be bound to plastic (Kramer, 2010).
Although the absence of serum leads to a higher unbound chemical concentration in medium, evaporation consequently increases for xylene and benzene in all simulated in vitro set-ups.This is in line with Kramer et al. (2012), who experimentally verified that serum retained phenanthrene in solution.In the same article, a possible deficiency of the model in capturing evaporation was reported, since it uses the HLC obtained/predicted for 20-25°C as a proxy of evaporation, while experiments are often run at higher temperatures such as 37°C (Kramer et al., 2012).Furthermore, the HLC is dependent on chemical solubility and vapor pressure increases with temperature.It was reported that the HLC increased by about 50% for chlorobenzene, chlorobiphenyls, and polycyclic aromatic hydrocarbons (Ten Hulscher et al., 1992) when the temperature was increased from 20 to 30°C.
Evaporation of phenantrene and 1,2,4-trichlorobenzene has been experimentally verified (Kramer et al., 2012), while our model points to only 0.03% and 2.74% of the respective chemi- The best way to integrate the predicted concentrations as a point of departure for extrapolation in risk assessment is still being studied.Can we assume that the unbound concentration reflects the bioavailable concentration, and can this be used as a point of departure for QIVIVE?In this work, the unbound concentration in the medium after 24-48 hours is assumed to be the "toxic dose", while some chemicals cause their toxic effect by being bound to lipids (narcosis) or proteins.Moreover, other dose metrics such as C max (maximum concentration) and the area under the concentration-time curve (AUC) are used in concentration-response studies (Groothuis et al., 2013), especially for repeated long-term exposures, which can be modelled by VCBA.The most appropriate dose metric for comparison with in vivo toxicity is an issue that should be further evaluated in the future.
To increase confidence in VCBA simulations, an initial step should be to check some of the simulations experimentally.It would be especially insightful to test compounds representing a wide range of physicochemical properties (e.g., logK ow and HLC).For HLC, the use of a correction factor such as van't Hoff may help to avoid under-prediction of evaporation (Armitage et al., 2014).Besides, the high uncertainty derived from modelling compounds outside the applicability domain of the QSARs included in the VCBA is an issue that needs to be addressed and, ideally, the applicability domains should be widened.Additionally, running local or global sensitivity analysis and uncertainty analysis (UASA) will increase model credibility.
The most direct use of the VCBA, even with all of its uncertainties, is to compare the results of in vitro toxicity assays.This can indicate whether partitioning or biokinetic processes explain differences between toxicity assays, and also whether assays of different durations actually produce the same results once the NEC and Kr are optimized.Moreover, the VCBA can be run without toxicity data when designing experiments: It can indicate whether a chemical is likely to evaporate, especially considering experiments with higher serum supplementation.Moreover, it gives indications whether a 100% mass balance can be expected after a chosen assay duration, and in case a specific range of concentrations is to be tested, what nominal concentrations in a specific medium correspond to respective unbound, bioavailable, or intracellular concentrations.
In this work, the VCBA is used to model a fibroblast cell line, where little metabolism and active transport will occur, mitigating the in silico model limitations.When simulating metabolically competent cell lines, such as HepaRG, it should be assessed whether the test chemical is metabolized or whether it is a substrate of specific transporters present in the cell line (Clerbaux et al., 2019).Furthermore, being a dynamic model capable of predicting different dose metrics (final concentration, C max , etc.) in an in vitro assay, the VCBA can be used to integrate toxicity data obtained by multiple approaches.
In general, the use of the VCBA allows increased confidence in the use of in vitro models for chemical risk assessment.

References
Andersen, M. E. and Krishnan, K. (1994).Physiologically based pharmacokinetics and cancer risk assessment.Environ Health to 6.13 (and HLC from 0.034 to 45 Pa×m 3 /mol), measuring the binding constants at 1% maximum water solubility, while for lipids, Jonker and van der Heijden (2007) analyzed 13 PAHs with logK ow values ranging approximately from 4.5 to 7. To develop the protein partitioning QSAR, chemicals other than PAHs were also considered, as the study was based on six independent studies including 36 chemicals with logK ow ranging from -1.3 to 5.1.PAHs are a family of neutral, non-polar chemicals; hence it is not clear to what extent VCBA outputs for chemicals with different characteristics are reliable.As an example, we considered protein binding as a non-specific interaction, which is not necessarily true for polar, charged, and more lipophobic chemicals (Gülden and Seibert, 2005).For metallo-organic and inorganic chemicals, this uncertainty increases further.
Similar to the QSARs included in the VCBA, some predictive tools and methods for the physicochemical input parameters, such as the ones in EPIsuite and Fuller's method for calculation of SVcomp, are based on organic chemicals.Fuller's method is less precise for inorganic compounds than for organics (error < 10%) (Tang et al., 2014).
Despite the assumptions of the VCBA, its simulations show to what extent the unbound concentration can deviate from the nominal concentration as a result of partitioning, evaporation or degradation.This emphasizes the importance of using in vitro fate data in the interpretation of in vitro studies and for in vitro to in vivo extrapolations.As in Gülden and Seibert (2003), the model can be used to separate chemicals into classes to distinguish compounds that are more prone to binding and that, therefore, have bioavailable concentrations that are different from the nominal concentrations.In addition, it can indicate how comparable in vitro toxicity experiments are for a given chemical.In this work, unlike in Gülden and Seibert (2003), a separation of chemicals of low and high in vitro potency is not made, but on the other hand, plastic binding, evaporation, degradation, and cell dilution are included.
Although not shown here, it is noteworthy that using the VC-BA final unbound concentrations did not improve, in general, the prediction of in vivo acute oral toxicity (Proença et al., 2017).It is unclear whether this is due to the uncertainties of the VCBA and/ or to the lack of in vitro to in vivo extrapolation (e.g., by PBK modelling).The in vivo dose used here is the nominal oral dose, but a chemical's in vivo kinetics comprise additional processes such as metabolism and absorption through the gastrointestinal tract, which are not accounted for in these comparisons.The cell line used in these assays has no metabolic competence; hence it will not correctly predict compounds that are bio-activated or that are more easily cleared or excreted when metabolized.It is noteworthy that the higher protein-binding ability of lipophilic compounds also includes a higher affinity to enzymes involved in xenobiotic metabolism.Thus, these chemicals tend to be metabolized more commonly than polar compounds (Pirovano et al., 2012), which further explains a lack of a trend between false predictions and high logK ow values.However, a more accurate comparison of in vivo unbound blood plasma concentrations and in vitro unbound medium concentrations would require PBK modelling for all 83 compounds, a task which extends beyond the scope of this specific work.

Fig. 1 :
Fig. 1: Schematic representation of the 4 interconnected models of the virtual cell based assay (VCBA) used to simulate the kinetics of a test chemical Kr, the rate constant for the killing rate increases as a function of the chemical concentration; NEC, no effect concentration for survival 7 https://ecvam-dbalm.jrc.ec.europa.eu/methods-and-protocols/search/139(accessed 08.03.2019).8doi:10.14573/altex.1812101s1

Fig. 2 :
Fig. 2: Optimization of NEC and Kr, and modelling chemical fate and effects in the cell population for 1,2,3,4-tetrachlorobenzene and diethyl phthalate In each panel, the top left plot shows the optimization of the concentration-response relationship by comparing cell viability after 48 hours (black lines) against experimental data (red dots).The top right plot shows the effect of the chemical's concentration (Molar or gram per gram wet weight) relative to the number of cells during the 48-hour assay for several nominal concentrations.For 1,2,3,4-tetrachlorobenzene the simulated concentrations ranged from 0 to 0.00232 M and for diethyl phthalate from 0 to 3.6×10 -4 M. The remaining plots show the unbound, intracellular, lipid and protein bound concentrations over time.For all plots except the first, 10 concentrations were modelled and are represented by shades of blue from lower (light color) to higher concentrations (darker color).
Fig. 3: Calculated percentage of partitioning of the 83 chemicals after 48 hours, relative to the nominal concentration Figure adapted from Proença et al. (2017); logKow, logarithm of octanol/water partition coefficient

Fig. 4 :
Fig. 4: Percentages of the final partitioned amounts (after 48 hours) relative to the nominal amount (amount indicated in the supplementary file 9 ) for different experimental set-ups A) Final unbound amount in medium for different experimental set-ups.B) Percentage of amounts partitioned into the components of the in vitro assay for a 96-well plate experiment with 0, 5, and 10% FBS.FBS, fetal bovine serum

Fig. 5 :
Fig. 5: Amounts of intracellular chemicals (moles) per cell in different experimental set-ups A) Amount of intracellular chemicals for several experimental set-ups.B) Percentage of chemicals per cell vs unbound in medium (for 96-well geometry, 5% FBS).FBS, foetal bovine serum; logKow, the logarithm of octanol/water partition coefficient