Consideration of the applicability domain of an in silico physiologically based pharmacokinetic model for improving the predictivity of food-related compounds

Abstract

The increasing demand for non-animal approaches in food safety assessment underscores the need for methods capable of predicting human internal exposure. We systematically evaluated the applicability of a physiologically based pharmacokinetic (PBPK) model—originally developed for pharmaceuticals and general chemicals, to predict physicochemical and pharmacokinetic parameters from chemical structure information and simulate human plasma profiles—to human internal exposure prediction. Food-related compounds were selected from published papers, and human single-dose pharmacokinetic studies were curated when all of the following criteria were met: 1) single administration in humans, 2) dose clearly specified as the target compound amount, 3) unchanged parent compound measured, 4) area under the curve (AUC) analyzed, and 5) no formulation/processing intended to control absorption. Using predicted parameters as PBPK inputs, peak plasma concentration (C_max) and AUC were simulated and compared with observed values. In the absence of applicability restrictions, prediction ratios (predicted value/observed value) ranged from 0.00970 to 825 for C_max and from 0.0000299 to 1628 for AUC; 20/36 C_max and 19/40 AUC data points were within the three-fold error range (0.33–3). Restricting the applicability domain to compounds with predicted intestinal absorption/bioavailability > 0.5 and excluding a specific chemical-space region improved predictivity, with 11/12 (91.6%) C_max and 9/12 (75.0%) AUC data points within the three-fold error range. Taken together, these results indicate that the in silico PBPK-based approach provides a practical and reasonably accurate estimate of human internal exposure for a subset of food-related compounds when an appropriate applicability domain is defined.

INTRODUCTION

The growing need for non-animal food safety assessment methods in line with the 3Rs (replacement, reduction, refinement) principles is driving a paradigm shift toward new approach methodologies (Ohta et al., 2023; Wood et al., 2025). Because risk assessment is fundamentally based on hazard and exposure, such non-animal approaches require not only alternative methods for hazard identification and point-of-departure derivation but also reliable tools to predict human internal exposure. Various non-animal approaches, including in silico, in chemico, and in vitro methods, have been developed (Sewell et al., 2016; Wanibuchi et al., 2019; Imai et al., 2024). In silico prediction is highly reproducible because the same algorithm produces identical outputs across facilities, and its utility has been evaluated in multiple safety assessment domains (Cronin et al., 2022). Physiologically based pharmacokinetic (PBPK) models are available for predicting the pharmacokinetics of food compounds; however, they generally require compound-specific parameters, and integrated approaches that predict internal exposure directly from chemical structure information are limited.

The Yamazaki laboratory has recently reported machine learning-based methods to predict human pharmacokinetics of pharmaceuticals and chemicals (Kamiya et al., 2022a; Adachi et al., 2022). These approaches provide a simple and reproducible prediction workflow for predicting physicochemical and pharmacokinetic parameters from chemical structure information and simulating internal exposure using a PBPK model. Their utility in risk assessment, including integration with toxicity evaluation, has been demonstrated (Kamiya et al., 2022b; Adachi et al., 2023). However, the models were developed for pharmaceuticals and general chemicals, and their applicability to food-related compounds has not been systematically investigated.

Therefore, in this study, we evaluated the predictive performance of this in silico PBPK-based approach for peak plasma concentration (C_max) and area under the curve (AUC) using curated human pharmacokinetic data for food-related compounds. Further, we analyzed compound characteristics associated with higher predictivity to clarify an applicability domain in the food-related chemical space.

MATERIALS AND METHODS

Literature search and inclusion criteria for studies on food-related compounds

Food-related compounds were selected from Science of Food Functionality (Nishikawa, 2008) and the Foods with Function Claims database in Japan (https://www.caa.go.jp/policies/policy/food_labeling/foods_with_function_claims/search/). PubMed and/or Google Scholar were searched to collect human pharmacokinetic studies. Publications were included in our analysis when they met all of the following criteria: (1) single-dose administration in humans, (2) dose clearly stated as the amount of the target compound, (3) unchanged parent compound measured, (4) AUC analyzed, and (5) no formulation or processing intended to control absorption.

Prediction of physicochemical and pharmacokinetic parameters in humans

Physicochemical properties and human pharmacokinetic parameters were predicted from chemical structure information (Simplified Molecular Input Line Entry System; SMILES and Mol files) using combined in silico tools and a reported machine learning model (Adachi et al., 2022; Kamiya et al., 2022a). The following parameters were obtained/predicted and used to derive PBPK inputs: cLogP, logD at pH 6.0 and 7.5, fraction unbound in plasma (f_u,p), absorption rate constant (k_a), volume of distribution (V1), hepatic intrinsic clearance (CL_h,int), and fraction of intestinal absorption and fraction escaping intestinal metabolism (F_aF_g). Canonical SMILES were obtained from PubChem (https://pubchem.ncbi.nlm.nih.gov/). Mol files with 2D-coordinate Mol files were generated from the canonical SMILES using OpenBabel (https://www.cheminfo.org/). cLogP (XLOGP3) values were obtained using SwissADME (https://www.swissadme.ch/). LogD (pH 6.0, 7.5) values were predicted using ChemAxon LogD Predictor (https://docs.chemaxon.com/display/docs/calculator. org/logd-plugin. md). f_u,p (“fup human”) values were obtained using DruMAP (https://drumap.nibiohn.go.jp/). Human pharmacokinetic parameter k_a, V1, CL_h,int, and F_aF_g values were predicted using the LightGBM-based model, as described previously (Adachi et al., 2022; Kamiya et al., 2022a), with SMILES, f_u,p, logD_6.0, and logD_7.5 as inputs.

PBPK model simulation and prediction of human C_max and AUC

Human plasma concentration–time profiles after oral administration were simulated using the reported PBPK model (Kamiya et al., 2022a). The model comprises compartments for the small intestine, liver, and kidneys, and a central compartment. While detailed compound-specific mechanisms are not explicitly included, the model reportedly provides consistent predictive performance across diverse chemicals. Model inputs included compound name, administered dose, AUC analysis time window, body weight, f_u,p, logP, k_a, V1, CL_h,int, and F_aF_g. Other pharmacokinetic and physiological parameters were calculated as previously described (Kamiya et al., 2022a). All input parameters used in this study are summarized in Table S1.

Calculation of chemical space

Chemical space (CS) was calculated to visualize structural diversity in two dimensions, as previously described (Adachi et al., 2024). In brief, molecular descriptors were computed from SMILES, followed by dimensionality reduction (e.g., principal component analysis) to map compounds onto a 2D plane. The plane was divided into 25 blocks, which were assigned CS numbers ranging 0–24.

Data analysis

For each compound, prediction ratios were calculated as predicted/observed values for C_max and AUC. Following prior work (Kamiya et al., 2022a), predictions within a three-fold error range (0.33–3) were considered acceptable. Ratios >3 were categorized as overprediction, and those <0.33 as underprediction.

RESULTS

Curated dataset of food-related compounds

In total, 35 curated food-related compounds were collected from 40 publications (Table 1). Doses ranged from 0.0039 mg (β-sitosterol) to 760 mg (α-linolenic acid; ALA). Observed C_max and AUC values ranged from 0.102 to 45550 ng/mL and from 0.0130 to 68091 μg·h/mL, respectively. Predicted physicochemical properties ranged widely (e.g., molecular weight ranged from 165 [phenylalanine] to 648 [menaquinone-7], and logP from –4.56 [thiamine hydrochloride] to 16.7 [menaquinone-7]). Details on physicochemical properties are provided in Table S1.

Table 1. Food-related compounds used in this study.

Predictivity for all curated compounds

Prediction results for C_max and AUC are summarized in Table 2. Prediction ratios ranged from 0.00970 (eicosapentaenoic acid; EPA) to 825 (daidzein) for C_max and from 0.0000299 (docosahexaenoic acid; DHA) to 1628 (fisetin) for AUC. Within the threefold range, 20/36 data points were acceptable for C_max and 19/40 for AUC. For C_max, 10 and six data points were over- and under-predicted, respectively; for AUC, eight and 13 data points were over- and under-predicted, respectively.

Table 2. Comparison of observed and predicted C_max and AUC values after oral administration in humans.

Relationships between compound characteristics and predictivity

The relationships between prediction ratios and selected parameters (CS, logP, k_a, F_aF_g, CL_h,int, and V1) are shown in Fig. 1a–1f. Predictivity tended to be lower for compounds located in a specific CS region (CS = 0 in this analysis; examples included 10-gingerol, ALA, DHA, docosapentaenoic acid (DPA), EPA, lutein, menaquinone-7, zeaxanthin, α-carotene, and β-carotene) and for compounds with predicted F_aF_g < 0.5 (e.g., 10-gingerol, 6-shogaol, ALA, chlorogenic acid, DHA, DPA, epigallocatechin gallate (EGCG), EPA, lycopene, thiamine hydrochloride, trans-resveratrol). No clear trends were observed for log P, k_a, CL_h,int, and V1.

Fig. 1

Relationships between predictivity (predicted/observed value ratios for C_max and AUC) and compound characteristics, including (a) CS; Chemical space, (b) logP, (c) k_a, (d) F_aF_g, (e) CL_h,int, and (f) V1.

Predictivity after applicability domain restriction

Based on the above findings, predictivity was re-evaluated after restricting the applicability domain using CS and F_aF_g criteria (Fig. 2). While the number of compounds within the applicability domain decreased to approximately one-third of the original dataset (12 compounds; 12 publications for both endpoints), the proportion within the three-fold range increased from 55.6% to 91.6% for C_max (11/12 data) and from 47.5% to 75.0% for AUC (9/12 data). Among the remaining outliers within the restricted set, fisetin was overpredicted for both C_max and AUC (452-fold and 1628-fold, respectively), zeaxanthin was slightly overpredicted for AUC (3.11-fold), and piperine was underpredicted for AUC (0.199-fold).

Fig. 2

Comparison of predictivity before and after applying an applicability domain for food-related compounds (dashed lines indicate 0.33–3-fold range).

DISCUSSION

Momentum toward non-animal approaches is growing not only in cosmetics but also in pharmaceuticals, general chemicals, and pesticides (ICH S5(R3), 2020; EU, 2016a, 2016b; EPA, 2018; EFSA, 2016; Masjosthusmann et al., 2020; Han, 2023; Weiner et al., 2024). The food sector is following a similar trend worldwide (FDA, 2017; EC, 2024; Cronin et al., 2025). In Japan, the Food Safety Commission has promoted the adoption of new approach methodologies through workshops and roadmaps (Tachibana et al., 2019; FSCJ, 2024). In thislight, the International Life Sciences Institute Japan AAT Project has endorsed alternative approaches for systemic toxicity assessment in the food domain (Ohta et al., 2023).

We systematically evaluated an in-silico PBPK-based approach for predicting human internal exposure from chemical structure information. We curated human pharmacokinetic data on food-related compounds and collected C_max data for 32 compounds (from 36 publications) and AUC for 35 compounds (from 40 publications). When applied to the entire dataset, the predictive performance of the model was lower than that reported for the original dataset used for model development (Adachi et al., 2023). Acceptable predictivity was more likely for compounds with predicted F_aF_g > 0.5 and located outside a specific CS region (CS ≠ 0 in this analysis).

Although mechanistic interpretation is generally challenging for machine learning-based methods, F_aF_g > 0.5 is suggestive of compounds with relatively favorable intestinal absorption and limited intestinal first-pass loss. Food-related compounds may exhibit absorption and metabolism characteristics distinct from typical pharmaceuticals, including low membrane permeability and susceptibility to intestinal metabolism (Ishii et al., 2019; Kikuchi et al., 2022), which could contribute to reduced predictivity. The CS criterion excluded many highly lipophilic nutrients (e.g., DHA, EPA, carotenoids) which often exhibit high persistence in humans (Lin et al., 2022; Richelle et al., 2004; Morifuji et al., 2020) and may deviate from the domain represented in the model training set.

Overall, our findings suggest that the in silico PBPK-based approach may be useful, particularly for early-stage screening, provided that an appropriate applicability domain is defined. For compounds outside the domain, further refinement of prediction strategies may require higher-tier methods such as in vitro assays, consistent with next-generation risk assessment concepts (Baltazar et al., 2020).

In conclusion, an in silico PBPK-based approach to predict human internal exposure from chemical structure information showed value for a subset of food-related compounds when an applicability domain was defined. For compounds with predicted F_aF_g > 0.5 and outside a specific CS region (CS ≠ 0), predicted C_max and AUC values were generally within 0.33–3-fold of observed values. These findings suggest that, within an appropriately defined applicability domain, this approach may provide a simple and reasonably accurate tool for estimating human internal exposure to food-related compounds.

ACKNOWLEDGMENTS

We are grateful to Prof. Dai Nakae (Faculty of Health Care and Medical Sports, Teikyo Heisei University) for his insightful advice and kind support. We thank Prof. Hiroshi Yamazaki (Laboratory of Drug Metabolism and Pharmacokinetics, Showa Pharmaceutical University) for his valuable advice in conducting this study.

Funding

This research did not receive any financial support.

Conflict of interest

The authors declare that there is no conflict of interest.

Data availability

The data in this study are included in the article/supplementary materials. Contact the corresponding author(s) directly to request the underlying data.

Author contributions

Conceptualization: Takuya Kikuchi, Kazutoshi Saito, Yuki Takahashi, Ayako Yamada

Investigation: Takuya Kikuchi, Kazutoshi Saito, Yuki Takahashi, Ayako Yamada

Supervision: Kazutoshi Saito

Visualization: Takuya Kikuchi, Yuki Takahashi, Maki Ishibashi

Writing – original draft: Takuya Kikuchi, Kazutoshi Saito, Yuki Takahashi, Maki Ishibashi

Writing – review & editing: Takuya Kikuchi, Kazutoshi Saito, Norie Arai

Funding acquisition: None

Ethical approval and consent to participate

Not applicable.

Patient consent for publication

Not applicable.

REFERENCES

•  Adachi,  K.,  Shimizu,  M. and  Yamazaki,  H. (2022): Updated in silico prediction methods for fractions absorbed and key input parameters of 355 disparate chemicals for physiologically based pharmacokinetic models for time-dependent plasma concentrations after virtual oral doses in humans. Biol. Pharm. Bull., 45, 1812-1817.
•  Adachi,  K.,  Nakano,  H.,  Sato,  T.,  Shimizu,  M. and  Yamazaki,  H. (2023): Liver and plasma concentrations of food chemicals after virtual oral doses extrapolated using in silico estimated input pharmacokinetic parameters to confirm reported liver toxicity in rats. Biol. Pharm. Bull., 46, 1133-1140.
•  Adachi,  K.,  Shimizu,  M.,  Shono,  F.,  Funatsu,  K. and  Yamazaki,  H. (2024): Octanol/water partition coefficients estimated using retention times in reverse-phase liquid chromatography and calculated in silico as one of the determinant factors for pharmacokinetic parameter estimations of general chemical substances. J. Toxicol. Sci., 49, 127-137.
•  Baltazar,  M.T.,  Cable,  S.,  Carmichael,  P.L.,  Cubberley,  R.,  Cull,  T.,  Delagrange,  M.,  Dent,  M.P.,  Hatherell,  S.,  Houghton,  J.,  Kukic,  P.,  Li,  H.,  Lee,  M.Y.,  Malcomber,  S.,  Middleton,  A.M.,  Moxon,  T.E.,  Nathanail,  A.V.,  Nicol,  B.,  Pendlington,  R.,  Reynolds,  G.,  Reynolds,  J.,  White,  A. and  Westmoreland,  C. (2020): A next-generation risk assessment case study for coumarin in cosmetic products. Toxicol. Sci., 176, 236-252.
•  Cronin,  M.T.,  Enoch,  S.J.,  Madden,  J.C.,  Rathman,  J.F.,  Richarz,  A.N. and  Yang,  C. (2022): A review of in silico toxicology approaches to support the safety assessment of cosmetics-related materials. Comput. Toxicol., 21, 100213.
•  Cronin,  M.T.,  Berggren,  E.,  Camorani,  S.,  Desaintes,  C.,  Fabbri,  M.,  Fabrega,  J.,  Herzler,  M.,  Ingram,  J.D.,  Lacasse,  K.,  Louhimies,  S.,  Maxwell,  G.,  Schutte,  K.,  Sobanski,  T.,  Streck,  G.,  Terron,  A. and  Worth,  A.P. (2025): Report of the European Commission workshop on “The roadmap towards phasing out animal testing for chemical safety assessments”, Brussels, 11-12 December 2023. Regul. Toxicol. Pharmacol., 161, 105818.
• EFSA Panel on Plant Protection Products and their Residues (PPR) (2016): Guidance on the establishment of the residue definition for dietary risk assessment. EFSA J., 14, e4549.
• Environmental Protection Agency (EPA). (2018): Strategic plan to promote the development and implementation of alternative test methods within the TSCA program. U.S. EPA, Washington, DC. https://www.epa.gov/sites/default/files/2018-06/documents/epa_alt_strat_plan_6-20-18_clean_final.pdf.
• European Union (2016a): How to use alternatives to animal testing to fulfil the information requirements for REACH registration. Practical guide. Publications Office of the European Union, Luxembourg. https://op.europa.eu/en/publication-detail/-/publication/e54fc06b-9ffd-11e6-868c-01aa75ed71a1/language-en.
• European Union (2016b): New approach methodologies in regulatory science – proceedings of a scientific workshop (Helsinki, 19–20 April 2016). Publications Office of the European Union, Luxembourg. https://data.europa.eu/doi/10.2823/543644.
• Food and Drug Administration (FDA). (2017): FDA’s predictive toxicology roadmap. U.S. FDA, Silver Spring, MD. https://www.fda.gov/media/109634/download.
• Food Safety Commission of Japan (FSCJ). (2024): Roadmap for the advancement of risk assessment methodologies. FSCJ (Japan). https://www.fsc.go.jp/chousa/roadmap.html.
•  Han,  J.J. (2023): FDA Modernization Act 2.0 allows for alternatives to animal testing. Artif. Organs, 47, 449-450.
• International Council for Harmonisation (ICH) S5. (R3) (2020): ICH Guideline on detection of toxicity to reproduction for medicinal products & toxicity to male fertility. ICH, Geneva. https://database.ich.org/sites/default/files/S5-R3_Step4_Guideline_2020_0218.pdf.
•  Imai,  T.,  Ishigamori,  R.,  Naruse,  M.,  Ochiai,  M.,  Maru,  Y.,  Hippo,  Y. and  Totsuka,  Y. (2024): Bridging toxicological properties of environmental chemicals between animals and humans using healthy organoid systems. J. Toxicol. Sci., 49, 425-434.
•  Ishii,  S.,  Kitazawa,  H.,  Mori,  T.,  Kirino,  A.,  Nakamura,  S.,  Osaki,  N.,  Shimotoyodome,  A. and  Tamai,  I. (2019): Identification of the catechin uptake transporter responsible for intestinal absorption of epigallocatechin gallate in mice. Sci. Rep., 9, 11014.
•  Kamiya,  Y.,  Handa,  K.,  Miura,  T.,  Ohori,  J.,  Kato,  A.,  Shimizu,  M.,  Kitajima,  M. and  Yamazaki,  H. (2022a): Machine learning prediction of the three main input parameters of a simplified physiologically based pharmacokinetic model subsequently used to generate time-dependent plasma concentration data in humans after oral doses of 212 disparate chemicals. Biol. Pharm. Bull., 45, 124-128.
•  Kamiya,  Y.,  Miura,  T.,  Kato,  A.,  Murayama,  N.,  Shimizu,  M. and  Yamazaki,  H. (2022b): Plasma concentration profiles for hepatotoxic pyrrolizidine alkaloid senkirkine in humans extrapolated from rat data sets using a simplified physiologically based pharmacokinetic model. Drug Metab. Bioanal. Lett., 15, 64-69.
•  Kikuchi,  T.,  Shigemura,  S.,  Ito,  Y. and  Saito,  K. (2022): Determination of human F_aF_g of polyphenols using allometric scaling. J. Toxicol. Sci., 47, 409-420.
•  Lin,  X.L.,  Baisley,  J.,  Bier,  A.,  Vora,  D. and  Holub,  B. (2022): Transgenic canola oil improved blood omega-3 profiles: a randomized, placebo-controlled trial in healthy adults. Front. Nutr., 9, 847114.
•  Masjosthusmann,  S.,  Blum,  J.,  Bartmann,  K.,  Dolde,  X.,  Holzer,  A.K.,  Stürzl,  L.C.,  Keßel,  E.H.,  Förster,  N.,  Dönmez,  A.,  Klose,  J.,  Pahl,  M.,  Waldmann,  T.,  Bendt,  F.,  Kisitu,  J.,  Suciu,  I.,  Hubenthal,  U.,  Mosig,  A.,  Leist,  M. and  Fritsche,  E. (2020): Establishment of an a priori protocol for the implementation and interpretation of an in-vitro testing battery for the assessment of developmental neurotoxicity. EFSA Support. Publ., EN-1938.
•  Morifuji,  M.,  Ichikawa,  S.,  Kitade,  M.,  Fukasawa,  T.,  Asami,  Y.,  Manabe,  Y. and  Sugawara,  T. (2020): Exopolysaccharides from milk fermented by lactic acid bacteria enhance dietary carotenoid bioavailability in humans in a randomized crossover trial and in rats. Am. J. Clin. Nutr., 111, 903-914.
• Nishikawa (2008): Science of Food Functionality. Sangyo Gijutsu Service Center, Tokyo.
•  Ohta,  S.,  Kimura,  S.,  Maejima,  D.,  Saito,  K.,  Tanaka,  Y.,  Kojima,  S.,  Kawaguchi,  T.,  Kikuchi,  T.,  Seki,  T.,  Chung,  H.,  Funaki,  R.,  Kuhara,  T.,  Hori,  H. and  Ito,  Y. (2023): Report on 2021 International workshop for non-animal approaches in the food sector (Japan): current status and avenues for further research. Altern. Anim. Exp., 40, 350-356.
•  Richelle,  M.,  Enslen,  M.,  Hager,  C.,  Groux,  M.,  Tavazzi,  I.,  Godin,  J.P.,  Berger,  A.,  Métairon,  S.,  Quaile,  S.,  Piguet-Welsch,  C.,  Sagalowicz,  L.,  Green,  H. and  Fay,  L.B. (2004): Both free and esterified plant sterols reduce cholesterol absorption and the bioavailability of β-carotene and α-tocopherol in normocholesterolemic humans. Am. J. Clin. Nutr., 80, 171-177.
•  Sewell,  F.,  Edwards,  J.,  Prior,  H. and  Robinson,  S. (2016): Opportunities to apply the 3Rs in safety assessment programs. ILAR J., 57, 234-245.
•  Tachibana,  K.,  Kass,  G.E.,  Ono,  A.,  Yamada,  T.,  Tong,  W.,  Doerge,  D.R. and  Yamazoe,  Y. (2019): A Summary Report of FSCJ Workshop “Future Challenges and Opportunities in Developing Methodologies for Improved Human Risk Assessments”. Food Saf. (Tokyo), 7, 83-89.
•  Wanibuchi,  S.,  Yamamoto,  Y.,  Sato,  A.,  Kasahara,  T. and  Fujita,  M. (2019): The amino acid derivative reactivity assay with fluorescence detection and its application to multi-constituent substances. J. Toxicol. Sci., 44, 821-832.
•  Weiner,  A.M.,  Irijalba,  I.,  Gallego,  M.P.,  Ibarburu,  I.,  Sainz,  L.,  Goñi-de-Cerio,  F.,  Quevedo,  C. and  Muriana,  A. (2024): Validation of a zebrafish developmental defects assay as a qualified alternative test for its regulatory use following the ICH S5(R3) guideline. Reprod. Toxicol., 123, 108513.
•  Wood,  A.,  Atienzar,  F.,  Basili,  D.,  Coulet,  M.,  Fernandez,  R.,  Galano,  M.,  Marin-Kuan,  M.,  Montoya,  G.,  Piechota,  P.,  Punt,  A.,  Reale,  E.,  Wang,  S. and  Hepburn,  P. (2025): Countdown to 2027 - maximising use of NAMs in food safety assessment: closing the gap for regulatory assessments in Europe. Regul. Toxicol. Pharmacol., 162, 105863.