In Vitro -Based Prediction of Human Plasma Concentrations of Food-Related Compounds

Efforts have been made to replace animal experiments in safety evaluations, including in vitro -based predictions of human internal exposures, such as predicting peak plasma concentration (C max ) values for xenobiotics and comparing these values with in vitro -based toxicity endpoints. Herein, the authors predicted the C max values of food-related compounds in humans based on existing and novel in vitro techniques. In this study, 20 food-related compounds, which have been previously reported in human pharmacokinetic or toxicokinetic studies, were evaluated. Human induced pluripotent stem cell-derived small intestinal epithelial cells (hiPSC-SIEC) and Caco-2 cells, HepaRG cells, equilibrium dialysis of human plasma, and LLC-PK1 cell monolayer were used to assess intestinal absorption and availability, hepatic metabolism, unbound plasma fraction, and secretion and reabsorption in renal tubular cells, respectively. After conversion of these parameters into human kinetic parameters, the plasma concentration profiles of these compounds were predicted using in silico methods, and the obtained C max values were found to be between 0.017 and 183 times the reported C max values. When the in silico -pre- dicted parameters were modified with in vitro data, the predicted C max values came within 0.1-10 times the reported values because the metabolic activities of hiPSC-SIECs, such as uridine 5’-diphospho-glucuronosyl transferase, are more similar to those of human primary enterocytes. Thus, combining in vitro test results with the plasma concentration simulations resulted in more accurate and transparent predictions of C max values of food-related compounds than those obtained using in silico -de- rived predictions alone. This method facilitates accurate safety evaluation without the need for animal experiments.


Introduction
Conventional risk assessment methods rely on animal testing.Alternative testing methods focus on hazard identification and characterization without the use of animals (EFSA, 2014;Gocht et al., 2015;Kojima, 2019).A new paradigm called next generation risk assessment (NGRA) was proposed for quantitative risk assessment without animal testing.NGRA is defined as an exposure-led, hypothesis-driven risk assessment approach that integrates new approach methodologies (NAMs), i.e., in silico, in chemico, and in vitro methodologies (Dent et al., 2018).Using NGRA, human physiological conditions can be reproduced more closely by using human physiologically based kinetic (PBK) models rather than relying on conventional in vitro-based toxici-procured from Fujifilm Wako.Acrylamide, fenitrothion, and β-estradiol were purchased from Sigma-Aldrich (St. Louis, MO, USA).Bisphenol S, picloram, and raloxifene were purchased from Tokyo Chemical Industry (Tokyo, Japan).Daidzein and genistein were purchased from LC Laboratories (Woburn, MA, USA).Quercetin, diclofenac, telmisartan, and troglitazone were purchased from Nacalai Tesque (Kyoto, Japan).
2.2 Membrane permeability assay using hiPSC-SIECs and Caco-2 cells hiPSC-SIECs were thawed and seeded at an initial density of 1 × 10 5 cells/well in Matrigel-coated 24-well cell culture inserts (Merck Millipore, Burlington, MA, USA) and then maintained for 9-13 days according to the manufacturer's instruction.Preplated Caco-2 cells were shipped on day 18 after seeding into 24-well cell culture inserts and maintained with DMEM supplemented with 1% NEAA, 10% FBS, and 50 U/mL penicillin/ streptomycin for an additional 3-7 days according to the suppliers' instructions.Transepithelial electronic resistance (TEER) values were measured with Millicell ERS-2 (Merck Millipore) immediately before and after the membrane permeability assays to ensure maintenance of the membrane barrier during the assays.
The assays were conducted according to a previously described method (Kitaguchi et al., 2021).Briefly, hiPSC-SIECs (n = 2) and Caco-2 cells (n = 3) were washed thrice with apical and basal transport buffers (HBSS containing 10 mM MES and 4.5 g/L glucose at pH 6.5 and 7.4, respectively) in chambers and incubated at 37°C for at least 30 min.Membrane transport assays were performed at 37°C for 30, 60, and 90 min after replenishing the apical chambers with transport buffer containing each substrate at a final concentration of 10 µM.The solution was collected from the basal chambers at 30 min intervals, diluted with an equal volume of acetonitrile, and stored at -20°C.Unreacted substrates were measured using ultraperformance liquid chromatography-tandem mass spectrometry (UPLC-MS/MS).
The apparent permeability coefficients (P app ) were calculated using the following equation: Eq. 1 where dQ/dt, A, and C 0 represent the amount of permeated compound per unit of time, the surface area of the transport membrane, and the initial compound concentration in the donor chamber, respectively.

Plasma protein binding assays for predicting the volume of distribution
The unbound fraction in human plasma (f up ) was determined by the equilibrium dialysis method with a Rapid Equilibrium Dialy-dent on the data sets used to create the prediction models.These prediction models are often built on data from pharmaceuticals due to the availability of detailed pharmacokinetic measurements in humans.As pharmacokinetic measurements regarding certain food-related chemicals is scarce, the direct applicability of these models to predicting the pharmacokinetics of food-related compounds is limited.An in vitro-based approach is required to overcome the above-mentioned issues as such an approach includes mechanistic information obtained using cellular and/or tissue-mimicking assay systems.
Previously, the authors have demonstrated the membrane permeability of food-related compounds using uridine 5'-diphospho-glucuronosyl transferase (UGT)-mediated intestinal metabolism in human-induced pluripotent stem cell-derived small intestinal cells (hiPSC-SIECs).These cells are more predictive of human physiology than Caco-2 cells (derived from cultured human colon cancer cells), which are considered the gold standard (Kitaguchi et al., 2021).
This study mimicked typical human ADME processes using existing and novel in vitro techniques, converted the in vitro data into human PBK parameters, and performed simulations using a human PBK model to evaluate the pharmacokinetic parameters of food-related compounds.Finally, the prediction accuracy of the Cmax values of 20 food-related test compounds was demonstrated by comparison with corresponding plasma concentrations in humans.
All drugs, food-related compounds, and metabolites listed below were dissolved in DMSO and stored at -20°C until further use.Curcumin, bisphenol A, 2-amino-3,8-dimethylimidazo[4,5-f] quinoxaline (MeIQx), and 7-hydroxycoumarin were 7.4 for the basolateral (1 mL) and 6.8 for the apical (0.2 mL) compartments.The cells on the inserts were washed thrice with prewarmed assay buffer, and then preincubated with buffer in both compartments for at least 30 min.Subsequently, the buffer from the apical chamber was aspirated and replaced with buffer containing test compound (1 µM).After incubation for 30 and 60 min at 37°C (n = 3), aliquots (0.1 mL) from both compartments were sampled and quantified with UPLC-MS/MS.
The P app values in basolateral to apical (P app,BA ) and apical to basolateral (P app,AB ) directions were determined using Eq. 1.The P app,BA and P app,AB values were upscaled to human intrinsic renal clearances (CL r,int,BA and CL r,int,AB ) using the following equation: where the surface of a human proximal tubule is calculated as the product of its length l PT (1.5 cm), its diameter d pt (7 × 10 -3 cm), and the number pi (π = 3.14).n neph is the number of nephrons per kidney (1.5 × 10 6 ), n kid the number of kidneys per human (2), and BW is the average human body weight (70 kg).

Data conversion from in vitro data to human PBK parameters Determination of absorptive constant (k a )
The experimental P app values of Caco-2 cells were converted to the effective permeability coefficient (P eff ) using GastroPlus (ver.9.8.2., Simulations Plus, Lancaster, CA) P eff converter with human P eff reported compounds.k a (10 -3 min -1 ) was derived from P eff according to the previous literature (Sugano, 2012) with the following equation: where DF is the degree of flatness of the gastrointestinal tract (1.7), and R GI is the radius of the gastrointestinal tract (1.5 cm).
Determination of fraction escaping intestinal metabolism (F g ) F g was estimated using a previously described method (Michiba et al., 2022) using the following equation: where CL AtoB,+inh and CL AtoB,-inh represent the apical-to-basolateral transcellular transport clearance in the absence and presence of a metabolic inhibitor, respectively.In this study, Caco-2 cells were assumed to possess negligible metabolic activity (Kitaguchi et al., 2021), whereas other factors such as transporter-and paracellular-mediated permeation were assumed to be similar in Caco-2 cells and hiPSC-SIECs.These assumptions suggested that the relationship of CL AtoB,+inh / CL AtoB,-inh ratio and P app of Caco-2 cells/P app of hiPSC-SIECs ratio are equal, and these values were substituted in the above equation.
sis device (Thermo Fisher Scientific) according to the manufacturer's instructions.Briefly, human plasma (Takara Bio) was dialyzed against phosphate buffer (pH 7.2) using a membrane with a nominal cutoff of 8 kDa at 37°C for 4 h with gentle agitation (n = 3).The final concentration of each compound added to the plasma was 2 or 20 µM depending on the detection sensitivity of UPLC-MS/MS.Equal volumes of either buffer (to the plasma samples) or compound-free plasma (to the buffer samples) were added, followed by protein precipitation with acetonitrile (performed four times) and filtration with FastRemover (GL Science, Tokyo, Japan) for UPLC-MS/MS.
The f up of each compound was calculated as the concentration of the buffer sample divided by the concentration of the plasma sample.

Metabolic stability assay using HepaRG cells for predicting hepatic clearance
HepaRG cell culture and metabolic stability assays were performed as described previously (Bonn et al., 2016).Briefly, the cryopreserved HepaRG cells were thawed and 100 µL cell suspension (0.72 × 10 6 cells/mL) was seeded into each well of flat-bottomed 96-well plates coated with type 1 collagen.They were allowed to attach for approximately 24 h in a CO 2 incubator, and the medium was renewed with HepaRG culture medium at room temperature.After 60 h, the medium was renewed with HepaRG serum-free induction medium, and 24 h later, the cells were exposed to the compounds for 0.25, 0.5, 1, 2, 4, 8, and 24 h (n = 3).After the incubation, the supernatants from each well were collected and diluted three times with ice-cold acetonitrile containing 0.8% formic acid.The samples were centrifuged at 4°C for 10 min at 10,000 g, and aliquots of the clear supernatant were diluted with equal amounts of water and kept at 4°C until UPLC-MS/MS analysis.
In vitro intrinsic hepatic clearance (CL h,int ) was calculated from the parent compound loss data according to Bonn et al. (2016) using the following equation: Eq. 2

Bidirectional membrane permeability assay using LLC-PK1 cells for predicting urinary clearance
These assays were conducted as described by Kunze et al. (2014) with minor modifications.LLC-PK1 cells were cultured in medium 199 (Thermo Fisher Scientific) supplemented with 10% FBS at 37°C in a 5% CO 2 incubator.The medium was changed every second to third day.The cells were cultivated for a maximum of 2 months and were passaged 16 times.Bidirectional transport assays were conducted in 24-well cell culture inserts (Corning, Corning, NY).The LLC-PK1 cells were seeded into 24-well cell culture inserts with a microporous polyethylene terephthalate membrane (0.4 µm pore size; 0.33 cm 2 surface area) at a density of 4 × 10 4 cells/cm 2 .The cells were maintained for four days as described above.The experiments were performed in modified Krebs buffer adjusted to pH

Determination of elimination constant (k e )
The k e (10 -3 min -1 ) was calculated using the following equation: Eq. 11 Prediction of in silico physicochemical and human kinetic parameters All physicochemical and kinetic parameters (pKa, LogP, k a , first pass, k e , and V d ) were directly predicted using ACD/Percepta (2017.2).Prediction of LogP and pKa are from QSAR (quantitative structure activity relationships) model on the basis of GALAS (global, adjusted locally according to similarity) modeling methodology (Sazonovas et al., 2010).Prediction of k a and V d are from previously reported QSAR models (Reynolds et al., 2009;Lanevskij et al., 2010), and k e and first pass are from the empirical model based on fragment structure descriptors and partial least squares regression statistics (personal communication).

Prediction of human plasma concentrations using simulation software
The PK Explorer module in ACD/Percepta (2017.2) was used to predict human plasma concentration profiles.The module consisted of differential equations involving solubility in the gastrointestinal tract, k a , first-pass metabolism in the gut and liver (1 − F g × F h ), V d , and k e .The solubility in the gastrointestinal tract was predicted from aqueous solubility corrected for the solubilizing effect of bile salts, and the other parameters were replaced with the in silico-predicted values described above.The predictions of human plasma concentrations were based on the compartmental absorption and transit model as previously reported (Yu and Amidon, 1999) (personal communication).

Assessment of absorption and Fg using Caco-2 cells and hiPSC-SIECs
Figure 1 shows the P app values obtained using Caco-2 cells and hiPSC-SIECs of 20 food-related compounds.The P app values in hiPSC-SIECs were 3.1-53 times lower than those in Caco-2 cells for quercetin, curcumin, daidzein, daidzin, genistein, genistin, bisphenol A, bisphenol S, and ferulic acid, which are known UGT-metabolized compounds.However, the P app values of cyanidin 3-glucoside and caffeic acid in hiPSC-SIECs were 5.6-7.7 times higher than those in Caco-2 cells.Furthermore, P app values were determined using Caco-2 cells and hiPSC-SIECs for the reported UGT-metabolized drugs in the small intestine, including raloxifene, diclofenac, telmisartan, and troglitazone.As shown in Table 1, the P app values for these drugs were 1.2-15 times lower in hiPSC-SIECs than in Caco-2 cells.
The F g values of these drugs were calculated using the conversion formula (Michiba et al., 2022).As shown in Table 1, the calculated F g values of the four drugs with reported F a (fraction absorbed from the intestine) × F g or F g values were comparable to the previously reported F a × F g or F g in humans (Nakamori et al.,

Determination of hepatic clearance (CL h ) and fraction escaping hepatic metabolism (F h )
CL h (mL/min/kg) was estimated based on a previous study (Bonn et al., 2016) using the following equation: Eq. 6 where f u,b is a fraction of unbound compound in the blood and was calculated as per a previous report (Hallifax et al., 2010), and f u,inc is a fraction of unbound compound after in vitro incubation predicted from a previous report (Kilford et al., 2008).
The F h was calculated from CL h and f u,b using the following equation according to a well-stirred model.

Determination of renal clearance (CL r )
The CL r (mL/min/kg) is a composed process involving glomerular filtration clearance (CL r,fil ), tubular secretion clearance (CL r,sec ), metabolic clearance, and reabsorption of a fraction of the drug from the tubule fluid back into the blood (f reab ).Assuming that the contribution of renal metabolic clearance is negligible, CL r,org can be expressed using the following equations (Kunze et al., 2014): where GFR is the glomerular filtration rate [1.79 mL/(min • kg)] and Q r,b is the renal blood flow rate [17.14 mL/(min • kg)].

Determination of the volume of distribution (V d )
The V d (L) was calculated according to previous reports (Øie and Tozer, 1979;Waters and Lombardo, 2010) using the following equation: where f ut is the fraction unbound in tissues and R E/I is the extravascular: intravascular ratio of binding proteins (usually 1.4 for albumin).V P , V E , and V R refers to the volumes of plasma, extracellular fluid, and remainder fluid with values of 0.0436, 0.151, and 0.38 L/kg, respectively, in humans.f ut was predicted according to a previous study (Berellini and Lombardo, 2019) using the following equation: Eq. 10 where f i,7.4 is the cationic fraction ionized at pH 7.4.

Comparison between the in silico-predicted and in vitro-derived values for human PBK parameters
The in vitro-derived parameters were converted to human PBK parameters (k a , first-pass metabolism in the gut and liver, k e , and V d ) with a previously reported conversion formula described in the Materials and Methods section.The obtained values were then compared with the in silico-predicted values (Fig. 2A-D).

Comparison between in silico-predicted and in vitro-derived values for Cmax
Figure 3 shows the results of human pharmacokinetic prediction using the PBK parameters obtained above.Detailed data is provided in Table S3 1 .When the plasma concentrations of the food-related test compounds were predicted using only in silicopredicted parameters, the Cmax values of these 20 compounds were between 0.017 and 183 times those of the reported Cmax values.When the in silico-predicted parameters were replaced with the in vitro data obtained in this study and reported physicochemical parameters, the predictivity of Cmax values did not improve (ranging from 0.19 to 333 times the reported Cmax values).Better predictivity was obtained using the F g values from hiPSC-SIECs: the predicted Cmax values of the 20 test compounds were almost within a factor of 10 of the reported Cmax values (0.092-10.1).

Discussion
Prediction of human pharmacokinetics is a key factor for NGRA.Recently, the application of safety evaluation methods without conducting animal experiments has been in high demand, not only for excipients of cosmetics but also for gen-Moreover, the predicted values for the physicochemical parameters, Log P and pKa, were compared with those available in the literature (Fig. 2E,F).The details of these values are provided in Table S1 and S2 1 .
For the first pass and k e , no apparent relationships were observed between the in silico-predicted and in vitro-derived values (Fig. 2B,C).In particular, the k e of ochratoxin A, quercetin, and chlorpyrifos varied considerably, possibly due to the significantly higher plasma protein binding rate (> 99.9%) than the in silico-predicted values (Tab.S1 1 ).
The in vitro-derived V d values were lower than the in silico-predicted values (Fig. 2D).Especially, the in vitro-derived V d for BHT, BHA, and chlorpyrifos were more than 10-fold lower than the in silico-derived values (248 vs. 5390, 56 vs. 574, and 43 vs. 1680 L/70 kg, respectively).Moreover, the physicochemical parameters (Log P and pKa) were consistent with the in silico-predicted values (Fig. 2E,F).Recently, Kikuchi et al. (2022) reported an allometry-based calculation of human F a × F g values for food-related compounds using existing pharmacokinetic data in rats.This study calculated the F a × F g values for quercetin and genistein, which were 0.0165 and 0.0375, respectively, comparable to predicted F g values (0.025 and 0.041, respectively).The possible limitation of this calculation method is the assumption that characteristics other than UGT metabolism (e.g., transporter expressions and functions, transcellular permeability, and paracellular permeability) are similar between Caco-2 cells and hiPSC-SIECs.However, the authors did not confirm these characteristics.For example, TEER, which indicates the tightness of the cell-to-cell connection, is significantly different between Caco-2 cells and hiPSC-SIECs.However, the authors reported that the F a × F g and P app slopes of the 20 known drugs are almost the same, regardless of the difference between the TEER values.Both cells have higher TEER values than the observed values in the human small intestine (Kitaguchi et al., 2021).Other possible differences are expressions and functions of uptake and efflux transporters (Press and Di Grandi, 2008).Further studies are required to confirm these assumptions.Since UGT inhibitors, which cover all types of isoforms, have yet to be reported, the estimation of F g using UGT inhibitors is limited.Therefore, it is challenging to validate the similarity in the characteristics between Caco-2 cells and hiPSC-SIECs.Hence, additional experimental UGT inhibitor-independent methods, such as UGT knockdown, may be required for validation.Daidzin and genistin, 7-O-glycosides of daidzein and genistein, respectively, are deglycosylated in the small intestine (Murota et al., 2002).Therefore, the differences in the P app values of these compounds may be due to the variation in UGT metabolism and glycosidase activity between Caco-2 cells and hiPSC-SIECs.Further studies must discuss the differences in the P app values of daidzin and genistin because the glycosidase activity of Caco-2 cells and hiPSC-SIECs is not well characterized.
HepaRG cells were used for the prediction of F h and CL h because the metabolic activity of these cells is similar to that of cryopreserved human hepatocytes (Zanelli et al., 2012;Bonn et al., 2016).Although cryopreserved human hepatocytes are used as a gold standard to predict CL h , these cells were not used in this study because they exhibit significant lot-to-lot variation (Araki et al., 2016).Further improvement might be needed to predict CL h using pooled cryopreserved human hepatocytes, preserving their metabolic activities, e.g., by growing them in 3-dimensional cultures (Lauschke et al., 2019).
V d was predicted by conducting equilibrium dialysis of human plasma for obtaining experiment-and physicochemical parameter-based f ut values (Berellini and Lombardo, 2019).Compared to the experimental f up , the in silico-predicted f up was between 1% and 100%.Values less than 1% were not observed.Highly protein-bound compounds (ochratoxin A, BHT, BHA, and chlorpyrifos) showed significantly higher plasma protein binding rates (> 99.9%) than the in silico-derived values.These differences might affect the prediction of V d values.For example, the in silico-derived V d of BHT was much higher than the in vitro-based V d (5390 L vs. 248 L, respectively).eral chemicals, including food-related compounds.Since food-related compounds are orally ingested, predicting their gastrointestinal absorption is essential for predicting human pharmacokinetics.However, reports on the pharmacokinetics of food-related compounds and pharmacokinetics-based predictivity studies without animal testing in humans are limited.This study evaluated the predictivity of Cmax values using recently reported alternative testing methods and compared it with human pharmacokinetic data of food-related compounds from existing literature.
Caco-2 cells have microvilli-like structures and can be used to predict the oral absorption of compounds that undergo transcellular permeation by passive diffusion (Lea, 2015).Owing to this property, Caco-2 cell monolayers have been used to predict gastrointestinal absorption.These cells also express intestinal efflux transporters such as P-glycoprotein, multidrug resistant protein 2, and breast cancer resistant protein (Lu et al., 2022).However, Caco-2 cells are colon cancer-derived cells and have characteristics that are different from the human small intestine such as an extremely high TEER value and limited expression of metabolic enzymes, including CYP3A and small intestine-specific UGT isoforms (Kabeya et al., 2020;Kitaguchi et al., 2021).In line with this, Punt et al. (2022) reported the predictivity of human plasma Cmax values of 44 pharmaceuticals and food-related compounds with certain overestimation.They discussed the reasons behind the overestimation compared to existing ADME information and identified several compounds that undergo intestinal metabolism.The prediction of F g was assumed to be a key factor.
Recently, new cells were developed, and human-derived samples were used for F g prediction (Michiba et al., 2022).The authors reported significant differences in the membrane permeation rates for a group of compounds undergoing glucuronidation in Caco-2 cells and hiPSC-SIECs.The expression and metabolic activities of UGT isoforms in hiPSC-SIECs were similar or higher than those in human small intestinal cells (Kitaguchi et al., 2021).As several test compounds used in this study are known glucuronidated substrates, an attempt was made to apply the first-pass metabolism in the gut by comparing the permeability data of Caco-2 cells and hiPSC-SIECs.The P app values of hiPSC-SIECs were 3.1-53-fold lower than those in Caco-2 cells.Previously, the authors of this study had reported that the relationships between P app values and F a or F a × F g of hiPSC-SIECs correlated better than those in Caco-2 cells.However, regression equation-based F a × F g calculation did not correlate well with the reported F a × F g or F g (data not shown).Therefore, the conversion formula put forth by Michiba et al. (2022) was used to calculate F g .They calculated CYP3A-derived F g based on the difference in the permeability of human intestinal organoid-derived cells with and without CYP3A4 inhibitor.It was assumed that Caco-2 cells have negligible UGT metabolic activity and that other factors are similar in Caco-2 cells and hiPSC-SIECs.Based on these assumptions, F g was calculated, which was found to be comparable to the reported F g of drugs in humans.not be a problem in evaluating their safety as previously discussed (Punt et al., 2022).
Addition of further ADME-related processes might increase the prediction accuracy of our method.For example, excretion in bile might be an important factor, especially for carboxylic and anionic compounds (Luo et al., 2010).Another possible factor might be the stability of food-related compounds in gastrointestinal fluid.Unlike pharmaceuticals, it is necessary to evaluate the stability of food-related compounds in digestive fluid prior to the absorption process.
Thus, although the consideration of in vitro-based parameters is important, improved simulation software for analyzing human plasma concentration is also needed.ACD/Percepta is a simplified compartment-based model that does not include the stability of compounds in the gastrointestinal tract and does not consider the different elimination processes, such as metabolism in the liver, renal elimination, and biliary excretion separately.Without modifying the PK Explorer module, its use can limit the prediction accuracy.The authors experienced this limitation while investigating catechins in the in vitro assays, where catechins were difficult to evaluate because of their instability at physiological pH (data not shown).
Out of all the test compounds, only ochratoxin A had underpredicted Cmax values (0.092-fold).This compound was reported to have extremely low f up values (6 × 10 -5 ), and its detection sensitivity was close to the detection limit of UPLC-MS/MS.This experimental variability may lead to a difference in predicted V d (11.7 L) compared to reported V d (6.9 L), which might lead to lower prediction of Cmax value.Punt et al. (2022) reported that predicted Cmax values of ochratoxin A were lower than the observed values; thus, the underestimation of ochratoxin A was not unique in our prediction method.Further research and refinement is needed to resolve this issue.
In conclusion, compared with in silico-derived predictions, the authors obtained more accurate and transparent predictions for the Cmax values of food-related compounds by appropriately combining ADME-related in vitro tests and simulation of plasma concentrations.This method can enable accurate safety evaluation of food-related compounds without the need for animal experiments.

References
Araki, T., Iwazaki, N., Ishiguro, N. et al. (2016) 2019).An accurate in vitro prediction of human VDss based on the Øie-Tozer equation and primary physicochemical descriptors.3. Analysis and assess-For predicting CL r , LLC-PK1 cell-based renal tubular secretion and reabsorption assays were used, as previously reported (Kunze et al., 2014).Daidzein, genistein, and caffeic acid were predicted to be the possible renal secreted compounds (CL r,fil < CL r,org ), whereas others were identified as reabsorptive compounds (CL r,fil > CL r,org ).Daidzein and genistein have longer half-lives in patients with end-stage renal disease (Fanti et al., 1999).Although renal excretion of caffeic acid was not reported, it can potentially inhibit human organic anion transporter 1 and 3 (Uwai et al., 2011).The limitation of this method is that the expression and function of some transporters (e.g., organic anion transporters) was insufficient and contributions of renal secretion and reabsorption were relatively smaller than the glomerular filtration ratio.Therefore, further improvement of in vitro-based prediction methods might be essential.As LLC-PK1 cells are porcine-derived and might not accurately represent the human RPTEC cell line, human renal proximal epithelium tubular cells-derived cell lines, such as RPTEC-TERT1, SA7K-clone, and ciPTEC, might be better choices for improving predictivity.
Finally, the in silico-predicted Cmax values of 20 food-related compounds were in the range of 0.017-187 times the reported human Cmax values.By replacing the in silico-predicted PBK parameters with in vitro-based ones, the prediction accuracy was improved.Then the predicted Cmax values were almost within 10-fold of those reported in the literature, and half of the compounds were within 3-fold.Previously, Terasaka et al. (2022) reported that most of the predicted Cmax values of 150 drugs classified as Class 1a and 2 by the extended clearance classification system were within a 6-fold range.Punt et al. (2022) reported that the predicted Cmax values were within a 5-fold range for 34 of 44 drugs and food-related compounds.Therefore, the authors considered that the prediction accuracy of our in vitro-based prediction method was comparable to those previously reported using in vitro and in silico-based prediction methods.
The in vitro-based prediction method employed in our study has two advantages: The first is its focus on predictive methods for food-related compounds with minimal usage of pharmaceutical datasets, which might broaden the applicability domain compared to those of drug-based datasets.The second is its transparency in mimicking each ADME process in vitro, which could be more agreeable to safety evaluation assessors emphasizing transparency in the prediction process.
However, the limitation of this method is the limited number of compounds (n = 20) compared with other reports.Further evaluation of compounds might confirm the reliability of our prediction method.
Maximum fold error of the prediction is essential because a large prediction difference might lead to over-or underestimation of the compounds for safety evaluations.Therefore, our in vitro-based prediction method was superior to the in silico-predicted method because the prediction variability was much lower.Although the Cmax values of certain food-related compounds evaluated in this study tended to be overestimated, this might data represents the mean ± S.D. of n = 3 (Caco-2 cells) or means of n = 2 (hiPSC-SIECs).a Nakamori et al., 2012.The values are calculated using the hepatic blood flow rate of 20.7 mL/min/kg.b Nishimuta et al., 2011.c Kikuchi et al., 2022.The values are from the allometric scaling-based-Fa × Fg calculation method.

Fig
Fig. 3: Comparison between the predicted and observed Cmax values from 20 food-related compounds The in silico predicted Cmax values, in vitro-based Cmax values with in silico predicted first pass, and in vitro-based Cmax values divided by the reported human Cmax values from previous literature were plotted.The dotted and dashed lines represent the predicted values that are within 3-fold and 10-fold, respectively.*Cmax values of daidzin and genistin were predicted as daidzein and genistein.

: The comparison between in silico- predicted and in vitro- derived or reported values for evaluating kinetic and physicochemical parameters
The in silico-predicted and in vitro-derived values of the test compounds were plotted for the following parameters: A) absorption rate constant (ka) from Caco-2 cells, B) first pass, C) elimination rate constant (ke), and D) volume of distribution (Vd); the dotted and dashed lines represent the values that are within 3-fold and 10-fold, respectively.The in silicopredicted and reported values were plotted against E) partition coefficient (Log P) and F) acidity constant (pKa); the dotted lines represent the values that are between -1 and 1.The compounds plotted are as follows: Dze, daidzein; Gse, genistein; Que, quercetin; GA, gallic acid; CA, caffeic acid; FA, ferulic acid; C3G, cyanidin 3-glucoside; Cur, curcumin; Caf, caffeine; AA, acrylamide; BHA, butylated hydroxyanisole; BHT, butylated hydroxytoluene; AFB, aflatoxin B1; OTA, ochratoxin A; BPA, bisphenol A; BPS, bisphenol S; CLP, chlorpyrifos; PLR, picloram.