An in vitro-in silico workflow for predicting renal clearance of environmental chemicals and drugs

Abstract

Accurate prediction of human renal clearance is essential for evaluating drug pharmacokinetics and environmental chemical risks, yet current methods often neglect rate-determining active transporter-mediated mechanisms. This study aimed to expand and validate a unified in vitro-in silico workflow for predicting renal clearance of both pharmaceuticals and per- and polyfluoroalkyl substances (PFAS) with varied elimination half-life ranges. We hypothesized that robust predictions of human renal clearance across diverse chemical classes can be achieved by combining human proximal tubule cell-based permeability/uptake assays with computational models of renal physiology. Human RPTEC/TERT1 cells and their OAT1-overexpressing variant were cultured in 96-well plates and Transwells to measure uptake, directional transport, and intracellular accumulation of 36 chemicals (28 PFAS, 7 drugs, 1 cosmetic ingredient). Time-course concentration data were used for either two-compartment (96-well) or three-compartment (Transwell) kinetic models. Permeability parameters were integrated into a physiologically-based kidney model for in vitro-to-in vivo extrapolation (IVIVE). A follow-up validation study with PFAS used independent experiments to derive similar predictions. Transwell-based three-compartment modeling yielded the most accurate absolute renal clearance predictions for rapidly eliminated drugs. For slowly cleared PFAS, simpler 96-well two-compartment modeling provided high correlation with observed human clearance, accurately distinguishing low-, medium- and high-clearance compounds; model predictions were consistently human health-protective. The PFAS validation study confirmed reproducibility of the approach. The proposed workflow is a conservative, scalable, mechanistically-informed and empirically-benchmarked approach for predicting renal clearance in humans. Transwell assays best support drug clearance estimation, whereas high-throughput 96-well formats enable reliable relative clearance ranking for PFAS, supporting both pharmaceutical development and environmental chemical risk assessment.

1. Introduction

Accurate prediction of human renal clearance is a central challenge for both drug development and chemical risk assessment (Ring et al., 2011). Renal elimination – the process by which the kidneys remove chemicals via urine – plays a major role in determining systemic exposure, bioaccumulation, efficacy, and potential adverse effects of chemicals. Traditionally, in vivo measurements in animal models have been the cornerstone for studying both overall toxicokinetics (OECD, 2010) and specifically renal clearance (Jansen et al., 2020). However, interspecies differences in renal physiology, transporter function/expression, and chemical metabolism can result in poor extrapolation to humans, demonstrating the need for advanced human in vitro and computational methods that more accurately predict human renal clearance for broadly diverse compounds, including drugs and environmental chemicals (Chang et al., 2022; Scotcher et al., 2016a).

Over the past two decades, the need to develop robust in vitro–in silico workflows to predict human renal clearance has grown more urgent. In pharmaceutical development, improved in vitro approaches are crucial to efficiently screen drug candidates for renal excretion profiles, potential drug-drug interactions at renal transporters, and human risk for nephrotoxicity (Scotcher et al., 2016b). For chemicals used in consumer products and industrial applications, the sheer scale of “data-poor” substances, numbering in the tens of thousands, demands predictive, high-throughput models that can estimate human pharmacokinetics in the absence of direct clinical data. This challenge is especially acute for per- and poly-fluoroalkyl substances (PFAS) that are known to vary considerably – from days to decades – in their human bioaccumulation potential (Dawson et al., 2023; East et al., 2023).

Current predictive frameworks for renal clearance draw on a range of experimental and computational models. Simple one- or two-compartment models, widely used in drug development and regulatory assessment, characterize pharmacokinetics with aggregate parameters such as volume of distribution and first-order elimination (Murphy, 2021). These models are often paired with in vitro data on plasma protein binding and permeability to predict renal clearance but typically assume passive glomerular filtration as the dominant elimination pathway (Jamei et al., 2009; Pearce et al., 2017). However, for many drugs and chemicals, active excretion and reabsorption via renal transporters in the tubules substantially influence net renal clearance, sometimes overriding passive filtration (Felmlee et al., 2013; Yin and Wang, 2016). To address these complexities, physiologically-based pharmacokinetic (PBPK) models have been developed that incorporate saturable transporter-mediated renal processes, as well as species, sex, and age dependencies (Chou and Lin, 2019; Huang et al., 2025). In vitro permeability and transporter data, often from cell lines overexpressing human renal transporters such as organic anion transporters 1 (OAT1) and 3 (OAT3), or multidrug and extrusion protein 1 (MATE1), are also used for in vitro–in vivo extrapolation (IVIVE) (Kumar et al., 2020). While these approaches are mechanistically useful, IVIVE models for renal clearance have been primarily developed and validated for pharmaceuticals, with limited translation to other classess of chemicals (Kunze et al., 2014; Neuhoff et al., 2013; Scotcher et al., 2016b). Despite these insights, most current high-throughput IVIVE or computational toxicokinetic models either neglect transporter-mediated processes or rely on parameterizations developed for poorly-related pharmaceutical scaffolds (Dawson et al., 2023; East et al., 2023). There is a pressing scientific need for experimental workflows that combine human-relevant in vitro transport measurements, particularly in systems expressing rate-determining renal transporters, with in silico models explicitly accounting for excretion and reabsorption to predict renal clearance for both drugs and data-poor environmental chemicals such as PFAS.

Although the landscape of in vitro and computational tools for renal clearance prediction has advanced rapidly, several challenges remain. Among them is a dearth of studies that systematically benchmark the performance of different in vitro cell systems and analytical approaches to chemical measurements for their ability to predict human renal clearance, especially for chemicals outside the traditional pharmaceutical domain (Lin et al., 2024; Sakolish et al., 2025). Therefore, the goal of this study was to extend a tiered in vitro–in silico workflow developed previously for PFAS (Lin et al., 2024; Sakolish et al., 2025) and to test and validate its utility for the prediction of human renal clearance for chemicals spanning both environmental and pharmaceutical domains. Specifically, we extended our previous studies by (i) comparing results from several of the human proximal tubule cell types that have been most frequently used for in vitro toxicity and transport studies – both an immortalized RPTEC line and its -OAT1 overexpressing variant (Aschauer et al., 2015); (ii) expanding the number of chemicals tested from 3 to 36, including both pharmaceuticals and PFAS; and (iii) including an external validation study where a set of PFAS was independently tested in 96-well plate model and analytical quantitations performed in two independent laboratories (with varied instrumentation and methods) to further test the robustness and reproducibility of the overall approach. Together, we demonstrate that our overall in vitro-in silico workflow is a robust, scalable, mechanistically-informed and empirically benchmarked strategy for generating human-relevant renal clearance predictions for both drugs and environmental chemicals.

2. Materials and methods

2.1. Tested chemicals

The list of chemicals used in these studies is provided in Table 1. In the model development study, we included a total of 36 substances – 28 PFAS, 7 drugs and 1 cosmetic ingredient (benzophenone-4, BP-4). For these studies, compounds were procured as follows. Perfluorobutanoic acid (PFBA) and perfluoroheptanoic acid (PFHpA) were purchased from Apollo Scientific Ltd. (Stockport, England). Perfluoroundecanoic acid (PFUnDA) and perfluorononanoic acid (PFNA) were purchased from Oakwood Products Inc. (Estill, SC) and perfluorodecanoic acid (PFDA) was purchased from Sigma-Aldrich (St. Louis, MO). The following compounds were purchased from SynQuest Laboratories (Alachua, FL): 1 H,1 H,2 H,2H-perfluorooctane sulfonic acid (8:2 FTS), 1 H,1 H,2 H,2H-perfluorohexane sulfonic acid (6:2 FTS), 1 H,1 H,2 H,2H-perfluorobutane sulfonic acid (4:2 FTS), perfluorooctanesulfonic acid (PFOS), perfluorohexanesulfonic acid (PFHxS) and perfluoropentanoic acid (PFPeA). Perfluorooctanoic acid (PFOA), perfluorohexanoic acid (PFHxA) and perfluorobutanesulfonic acid (PFBS) were obtained from TCI America (Portland, OR). Isotopically labeled perfluorooctanoic acid (C¹³-PFOA; #MPFOA) was purchased from Wellington Laboratories (Guelph, Ontario, Canada) and used as the internal standard for all PFAS. Additional test articles and their corresponding internal standards were obtained from Sigma-Aldrich (St. Louis, MO): tobramycin (#PHR1079), indomethacin (#I8280), rifampin (#R7382), benzophenone-4 (BP-4; #09649), tenofovir (#SML1795), gentamicin (#G1397), streptomycin (#1623003), colchicine (#PHR1764), naproxen (#N8280), isotopically labeled caffeine (C¹³-caffeine; #C-082), diclofenac (#SML-3086), and repaglinide (#R9028).

Table 1.

Chemicals tested in this study. In the model development study, a panel of 22 compounds was evaluated for uptake and transport in RPTEC models. The list includes 14 per- and polyfluoroalkyl substances (PFAS) of varying chain lengths and functional groups, along with pharmaceutical compounds and the UV filter Benzophenone-4 (BP-4). The follow-up validation study included 23 PFAS with 9 overlapping with the primary study. Each compound is shown with its common abbreviation, chemical formula, CAS number, and the concentration tested.

Chemical Name	Chemical Abbreviation	Chemical Formula	CAS Number	Tested Concentration
Model Development Study
Perfluoroundecanoic acid*	PFUnDA	C11HF21O2	2058–94–8	5 μM
8:2 Fluorotelomer sulfonic acid	8:2 FTS	C10H5F17O3S	39108–34–4	5 μM
Perfluorodecanoic acid*	PFDA	C10HF19O2	335–76–2	5 μM
Perfluorooctanesulfonic acid*	PFOS	C8HF17SO3	1763–23–1	5 μM
Perfluorononanoic acid*	PFNA	C9HF17O2	375–95–1	5 μM
6:2 Fluorotelomer sulfonic acid	6:2 FTS	C8H5F13O3S	27619–97–2	5 μM
Perfluorooctanoic acid	PFOA	C8HF15O2	335–67–1	5 μM
Perfluorohexanesulfonic acid*	PFHxS	C6HF13SO3	355–46–4	5 μM
Perfluorobutanoic acid*	PFBA	C4HF7O2	375–22–4	5 μM
Perfluoroheptanoic acid*	PFHpA	C7HF13O2	375–85–9	5 μM
4:2 Fluorotelomer sulfonic acid	4:2 FTS	C6H5F9O3S	757124–72–4	5 μM
Perfluorohexanoic acid*	PFHxA	C6HF11O2	307–24–4	5 μM
Perfluoropentanoic acid	PFPeA	C5HF9O2	2706–90–3	5 μM
Perfluorobutanesulfonic acid*	PFBS	C4HF9O3S	375–73–5	5 μM
Tobramycin		C18H37N5O9	32986–56–4	1000 μM
Indomethacin		C19H16ClNO	53–86–1	10 μM
Rifampin		C43H58N4O12	13292–46–1	10 μM
Benzophenone–4	BP–4	C14H12O6S	4065–45–6	10 μM
Tenofovir (TFV)		C9H14N5O4P	147127–20–6	10 μM
Gentamicin		C21H43N5O7	1403–66–3	1000 μM
Streptomycin		C21H39N7O12	57–92–1	1000 μM
Colchicine		C22H25NO6	64–86–8	10 μM
Follow-up Validation Study Chemicals
3:3 Fluorotelomer carboxylic acid	3:3 FTCA	C6H5F7O2	356–02–5	10 μM
4H-Perfluorobutanoic acid	4H-PFBA	C4H2F6O2	679–12–9	10 μM
2 H,2 H,3 H,3H-Perfluorooctanoic acid	5:3 FTCA	C8H5F11O2	914637–73–1	10 μM
9H-Perfluorononanoic acid	9H-PFNA	C9H2F16O2	76–21–1	10 μM
Nonafluoro–3,6-dioxaheptanoic acid	NFDHA	C5HF9O4	151772–58–6	10 μM
Ammonium perfluorooctanoate	NH4PFOA	C8H4F15NO2	3825–26–1	10 μM
Perfluoro(2-ethoxyethane)sulfonic acid	PFEESA	C4HF9O4S	113507–82–7	10 μM
Perfluoroheptanesulfonic acid	PFHpS	C7F15SO3H	375–92–8	10 μM
Potassium perfluorohexanesulfonate	PFHxS-K	C6F13KO3S	3871–99–6	10 μM
Perfluoro–4-methoxybutanoic acid	PFMBA	C5HF9O3	863090–89–5	10 μM
Potassium perfluorooctanoate	PFOA-K	C8F15KO2	2395–00–8	10 μM
Potassium perfluorooctanesulfonate	PFOS-K	C8F17KO3S	2795–39–3	10 μM
Perfluoro–4-isopropoxybutanoic acid	PFPE–1	C7HF13O3	801212–59–9	10 μM
Perfluoro–3,6-dioxadecanoic acid	PFPE–5	C8HF15O4	137780–69–9	10 μM

For the follow-up model validation study with PFAS, DMSO stocks were obtained through a US EPA contract (Evotec, Inc., Branford, CT). Neat standards obtained from vendors (purities exceeding 95 %) were solubilized at a target concentration of 30 mM; lower concentrations were utilized if solubility issues were noted. For secondary source verification and quality control, a mixture containing 30 native substances (#PFAC30PAR; Lot: PFAC300621) and, for internal standard usage, a mass-labeled PFAS mixture (#MPFAC-24ES; Lot: MPFA-C24ES0424) were obtained from Wellington Laboratories (Guelph, Ontario, Canada). LCMS-grade acetonitrile was from Honeywell Research Chemicals (Morris Plains, NJ; LC015–4, Lot: EJ845-US) and ammonium acetate from Sigma-Aldrich (73594–100g-F).

2.2. Cell culture methods

The human telomerase reverse transcriptase-immortalized renal proximal tubule epithelial cell line (hTERT-RPTEC) and its OAT1 transporter-overexpressing variant (hTERT-RPTEC-OAT1) were purchased from ATCC (#CRL-4031, #CRL-4031-OAT1; Manassas, VA). These cells were cultured as recommended by the vendor in DMEM/F12 (#30–2006, ATCC) supplemented with the “hTERT Immortalized RPTEC Growth Kit” (#ACS-4007, ATCC), with a final concentration of 0.1 mg/mL Geneticin^™ Selective Antibiotic (G418 Sulfate; Gibco, Waltham, VA). RPTECs were cultured at 37°C and 5 % CO₂ in Tissue Culture-treated 75 cm² flasks (#CLS430641U, Corning Inc, Corning, NY) prior to seeding onto 96-well Transwells (#3381; Corning Inc) or 96 well plates (#3904, Corning Inc). On the day of seeding, Transwells were coated with 30 μL of 100 μg/mL fibronectin (Sigma-Aldrich # F1141–5MG) in phosphate buffered saline (PBS) for 1 h at 37°C. After coating, excess solution was removed, and RPTECs were seeded onto the Transwells at a density of 294,000 cells/mL in cell culture medium (22,050 cells in 75 μL media), and allowed to attach for 4 h at 37°C. After incubation, 235 μL of medium was added to the bottom “receiver” plate. For 96-well plates, matrix coating was not necessary and RPTECs were seeded at a density of 213,333 cells/mL in cell culture medium (32,000 cells in 150 μL media) on standard tissue culture plates. After cell seeding, cultures were maintained at 37°C and 5 % CO₂ for 6 days with media changes every 48 h prior to treatment with test articles in 96 well plates, or for 2 days in Transwells, which was time frame observed for maximal Transepithelial Electrical Resistance (TEER). Barrier formation on Transwells was confirmed via TEER testing to ensure monolayer formation using the EVOM Manual system (#EVM-MT-03–01; World Precision Instruments, Sarasota, Florida) with the HTS probe (#STX100C96, World Precision Instruments). Transwell cultures with TEER less than 65 Ω×cm² were excluded from testing due to sub-optimal barrier integrity.

2.3. Chemical transport studies

To evaluate chemical transport across the epithelial cell layer and potential intracellular retention, chemicals were added to cell culture medium at final concentrations shown in Table 1. In Transwell cultures, medium was either added to the bottom well (235 μL, basolateral surface, B-to-A transport) or top well (75 μL, apical surface, A-to-B transport) of the RPTEC cultures. Fresh RPTEC culture medium was added to the opposite (recipient) well in either treatment configuration. In 96-well plates, 150 μL of chemical-containing medium was added to each well. After treatment, cultures were returned to the incubator until sampling. Media were sampled (25 μL) from the top and bottom wells of Transwells, or from the 96-well plate at 4 and 24 h. After the final timepoint, cells were rinsed with PBS, then lysed with M-PER^™ Mammalian Protein Extraction Reagent (30 μL, #78503; Thermo-Scientific, Waltham, MA) for 10 min at room temperature to determine intracellular accumulation of test compounds. Media and lysate samples were stored at 20°C until further analysis.

2.4. Cell viability assessment

To evaluate cell viability and barrier integrity after treatment with test chemicals, TEER was measured in Transwell cultures, and lactate dehydrogenase (LDH) activity was measured in the media sampled from all culture conditions. TEER was measured using an epithelial volta-meter with Ag/AgCl probes as detailed in section 2.1.2. Resistance values (Ω) were recorded, and TEER was calculated by subtracting the resistance of blank membranes and multiplying by the cell culture area of the Transwells (0.143 cm²). Results are reported as Ω×cm². Additionally, lactate dehydrogenase (LDH) activity was measured from all culture conditions after 24 h of exposure using a biochemical assay kit (#ab102526; Abcam, Cambridge, UK) following the manufacturer’s protocol.

2.5. Sample preparation for analytical assays

For the model development set of experiments, samples were prepared using liquid-liquid extraction with protein precipitation. In brief, 25 μL of medium was combined with 50 μL of chilled acetonitrile (#A998–4, Fisher Scientific) containing the appropriate internal standard for each analyte: 0.1 μM C¹³-PFOA was used for all PFAS; 1 μM gentamicin for tobramycin; 5 μM naproxen for indomethacin; 1 μM C¹³-caffeine for tenofovir, rifampin, and colchicine; 5 μM diclofenac for BP-4; and 1 μM repaglinide for streptomycin and gentamicin. Samples were vortexed, centrifuged, and the resulting supernatant was transferred to a 1.5 mL microcentrifuge tube. Supernatants were evaporated to dryness using a Savant^™ SpeedVac^™ concentrator (#SPD1010, ThermoFisher), then reconstituted in 25 μL of mobile phase A (Table S1). Reconstituted samples were transferred to autosampler vials fitted with 200 μL glass fused inserts (Ibis Scientific, Las Vegas, NV) and stored at −20°C until LC-MS/MS analysis.

For the follow-up validation study of PFAS, samples were vortexed and centrifuged briefly, after which a 100 μL aliquot was mixed with 200 μL ice-cold acetonitrile containing the mass-labeled PFAS standard mixture (MPFAC-24ES, Wellington Laboratories) and vortexed for 5 s. After 20 min on ice, the samples were centrifuged for 15 min at 12500 ×g. A 50 μL aliquot of the supernatant was diluted with 100 μL of the initial mobile phase condition (see method description below) for the mass spectrometry run.

2.6. Analytical chemistry methods

The initial quantitative analysis was performed using a triple quadrupole mass spectrometer (Agilent 6470, Santa Clara, CA) coupled with an Agilent 1290 Infinity II liquid chromatography (LC) system. Chromatographic separation was achieved using a ZORBAX SSHD Eclipse Plus C18 column (3.0 × 50 mm, 1.8 μm; #959757–302, Agilent) with an inline guard column (2.1 × 5 mm, 1.8 μm; #821725–901, Agilent). Ionization mode, injection volumes, instrument parameters, mobile phase compositions, run times, and chromatographic gradients are detailed in Table S1.

For the validation study of PFAS, quantitation of analytes was performed by ultra-high-performance liquid chromatography-tandem mass spectrometry (UPLC-MS/MS). Specifically, extracted medium and cell lysate samples were analyzed using a Waters ACQUITY I-Class UPLC coupled to a Xevo TQ-XS or TQ-S micro triple quadrupole mass spectrometer (Waters Corporation, Milford, MA). The UPLC system was plumbed with the Waters’ PFAS Solution Installation Kit with solvent lines containing polyether ether ketone, stainless-steel filter, and an isolator column (ACQUITY BEH C18 column 2.1 × 50 mm). Chromatographic separation was carried out using a Waters CORTECS T3 reversed-phase column (3 mm × 100 mm, 2.7 μm), a flow rate of 0.4 mL/min, and a binary mobile phase gradient with mobile phases A (95:5, 2.5 mM ammonium acetate: acetonitrile) and B (95:5, acetonitrile: 2.5 mM ammonium acetate). The gradient program was 11 min total and programmed as follows: 2 % B (1.2 min), 2–30 % B (0.8 min), 30–55 % B (4 min), 55–75 % B (2 min), 75–99 % B (1 min), 99–2 % B (0.4 min), 2 % B (1.4 min), and 2 % B (0.2 min). 10 μL of each sample was injected regardless of chromatography employed. The Waters Xevo triple quadrupole mass spectrometer (MS) was operated in unispray negative mode (US-). Parameters varied depending on instrument (Table S2). Previously optimized, multiple reaction monitoring (MRM) transitions were used for each unique PFAS (Table S3) and often included a quantitation and confirmation transition. Internal standards were also optimized (Table S4). Blank matrix samples as well as instrument blanks were included to monitor for any potential PFAS contamination or low-level responses associated with the matrix and/or instrumentation. Samples were thawed to room temperature, vortexed briefly, and resuspended in 98:2 mobile phase A: mobile phase B in polypropylene autosampler vials. A 20-point calibration curve was prepared in a similar sample matrix that consisted of Dulbecco’s modified eagle medium (DMEM), ranging from concentrations of 10,000 nM to 2.78 nM. The calibration curve matched the dilutions of samples. For each assay set, solvent and matrix blanks were included to evaluate possible contamination from laboratory operation and method performance. Blanks were assessed every sixth sample of the sample analysis worklist. All blanks were determined to have a concentration to be less than half the estimated method detection limit (US EPA, 2016). Additionally, the recoveries of labeled standards were within 75–125 %.

2.7. In vitro–in silico method for predicting renal clearance

The approach for estimating human in vivo renal clearance based on in vitro data was adapted from the previously published method (Lin et al., 2024). Briefly, the method integrates experimental measurements and computational modeling to extrapolate in vitro findings to in vivo outcomes. The workflow involves: (1) quantifying time-course chemical concentrations in both cell and media compartments, using 96-well plates and Transwell plates, with the latter including sampling from both apical and basolateral chambers; (2) determining in vitro permeability by fitting the data to a multi-compartmental model; and (3) extrapolating the derived permeability to predict human renal clearance using a physiologically-based kidney model.

2.7.1. Bayesian modeling of Transwell-based kinetics for estimating in vitro permeability

A three-compartment model was developed to describe chemical transport within the Transwell plates, representing the apical (upper chamber; analogous to the tubular lumen), cellular, and basolateral (lower chamber; analogous to the kidney blood side) compartments. A schematic representation of this model is shown in Fig. 1. The model’s structure and the corresponding system of differential equations were adopted from (Lin et al., 2024). To derive chemical-specific permeability parameters time-course concentration data from bidirectional transport assays (i.e., where chemical was added to either the apical (A-to-B) or basolateral (B-to-A) compartment) were fitted to the model simultaneously. The resulting parameters include permeability coefficients between the cell and basolateral compartments (P₁ and P₂), and between the cell and apical compartments (P₃ and P₄) (Fig. 1).

Fig. 1.

Schematic overview of experimental design and modeling workflow. hTERT-RPTEC and hTERT-RPTEC-OAT1 cells were used in 96-well plates and Transwells to measure compound concentrations in cell lysates (24 h only) and media (4 and 24 h). Data were integrated into two- and three-compartment models to estimate media-to-cell ratios and transport dynamics. Predicted renal clearance values were then compared with reported human in vivo data.

Bayesian Markov Chain Monte Carlo (MCMC) methods were applied for parameter estimation. The first step in the MCMC process included assigning prior distributions to the parameters of interest (P₁ to P₄) and defining the likelihood function. As described previously (Lin et al., 2024), no prior information was available for the permeability; thus, non-informative lognormal priors were used with a geometric mean (GM) of 10⁻³ and a geometric standard deviation (GSD) of 10⁵. The likelihood, meaning the residual error model, assumed log-normal distribution of residuals, and the error’s GSD was assumed to follow a log-uniform distribution.

Convergence of the MCMC chains was evaluated using four independent chains. The initial iterations of each chain were discarded as burn-in, while the remaining iterations were used to derive the posterior distributions. The number of iterations and the length of the burn-in phase were determined based on convergence assessment, using the Gelman–Rubin criterion (Gelman and Rubin, 1992), which needs the potential scale reduction factor $\left(\widehat{R}\right)$ to be lower than 1.2.

Finally, to evaluate model performance, simulated concentration-time profiles (based on the posterior permeability values as input into the three-compartmental model) were compared with experimentally observed data. Simulated values were visualized as median concentrations with associated 90 % confidence intervals. Model fit performance was assessed by calculating the mean absolute deviation (MAD, Eq. 1) and mean error (ME, Eq. 2) of log₁₀-transformed concentrations between observed and predicted values, calculated by:

$$MAD=\frac{\sum \left|\text{log}10\left(C_{\mathit{\text{pred}}}\right)-\text{log}10\left(C_{obs}\right)\right|}{n}$$ (1)

$$ME=\frac{\sum \text{log}10\left(C_{\mathit{\text{pred}}}\right)-\text{log}10\left(C_{obs}\right)}{n}$$ (2)

where C_pred and C_obs were predicted and observed concentrations, respectively, and n is the number of data points. Additionally, the Pearson’s and Spearman’s correlation coefficient was calculated to evaluate the strength of the monotonic relationship between predicted and observed concentrations.

2.7.2. Physiologically-based renal clearance model

To extrapolate the in vitro permeability to the prediction of in vivo renal clearance, we used the physiological-based kidney-specific model developed by (Lin et al., 2024) that captures chemical kinetics within the proximal tubule and cell, as well as surrounding vasculature. Briefly, the model structure reflects key physiological features of the human kidney such as blood flow through the renal artery into the glomerulus where a portion is filtered into the proximal tubule lumen and reflected in the glomerular filtration rate (GFR). Chemicals in their unbound form are distributed between blood and tubular fluid and the mathematical formulation of the model was detailed previously (Lin et al., 2024) with a final equation for estimating unbound renal clearance (CL_{renal, free}, Eq. 3) shown in Fig. 1:

$$C{L}_{\mathit{\text{renal}},\mathit{\text{free}}}=\frac{Q_K•Q_{PT}}{\left(Q_{PT}+\frac{\left(Q_K-Q_{PT}\right)}{\mathit{\text{Pratio}}}\right)}$$ (3)

where Q_K and Q_PT represent the renal arterial blood flow and proximal tubule flow rate, respectively. The permeability ratio (P ratio) is defined as: $\frac{P_2•P_4}{P_1•P_3}$. Here, P₁ to P₄ are the permeability coefficients estimated from the Bayesian MCMC simulation, representing bidirectional transport across cellular membranes.

In the 2D culture format (96-well plates), where the basolateral side of the cells is affixed to the well bottom, it was assumed that P₁ = P₂. Under this assumption, the permeability ratio (Eq. 4) simplifies to:

$${\mathit{\text{Pratio}}}_{2D}=\frac{P_4}{P_3}=\frac{C_{\mathit{\text{Lumen}}}}{C_{\mathit{\text{Cell}}}}$$ (4)

Although Transwell data allow for full application of the three-compartment model and the complete in vitro–in silico workflow, we also explored whether a simplified 2D-like treatment of Transwell data (i.e., meaning only using cell and donor-side concentrations) could also generate reasonable renal clearance predictions.

Subsequently, to compute the total renal clearance (CL_renal,total, Eq. 5), the clearance for free chemicals was adjusted by the fraction of unbound chemical in plasma (f_ub):

$$C{L}_{\mathit{\text{renal}},\mathit{\text{total}}}=C{L}_{\mathit{\text{renal}},\mathit{\text{free}}}•f_{ub}$$ (5)

The f_ub values, which are chemical-specific, are provided in Table S5. Human kidney physiological parameters used in the model include a renal blood flow (Q_K) of 29.6 L/h (equivalent to 7.1 × 10⁶ mL/day) based on a 70-kg individual (Chou and Lin, 2019). Because the proximal tubule accounts for approximately 70 % of water reabsorption, the proximal tubule flow (Q_PT) was set as 30 % of the GFR (i.e., 6.77 L/h × 0.3 = 4.9 × 10⁴ mL/day) (Chou and Lin, 2019; Feraille et al., 2022). Finally, the predicted renal clearance values were compared to median of in vivo human clearance data sourced from literature, as detailed in the Table S5 and Table S6. Model performance was evaluated using both Pearson and Spearman correlation coefficients.

2.8. Computational software

The three-compartment model for fitting Transwell data was implemented using the Markov Chain Monte Carlo (MCMC) algorithm through GNU MCSim (version 6.1.0) within the R programming environment. Subsequent data processing and visualization were carried out in R (version 4.1.2) using RStudio, with support from several packages including rstan (v2.21.7), tidyverse (v1.3.2), plyr (v1.8.9), and ggplot2 (v3.4.4). All raw datasets and code used for modeling and analysis are publicly accessible in the GitHub repository: https://github.com/hsingchiehlin/CLrenal_pred_22chem.

2.9. Data availability

Data for all experiments reported herein can be accessed through the EveAnalytics database at the links below. The 96-well plate RPTEC parent and RPTEC-OAT1 datasets are available at https://eve.eveanalytics.com/assays/assaystudy/1384/ and https://eve.eveanalytics.com/assays/assaystudy/1385/, respectively. Corresponding data for the Transwell RPTEC parent and RPTEC-OAT1 studies can be found at https://eve.eveanalytics.com/assays/assaystudy/1387/ and https://eve.eveanalytics.com/assays/assaystudy/1386/.

3. Results

An overview of the study design is illustrated in Fig. 1. The overall workflow was adapted from (Lin et al., 2024) and is divided into two major components: (1) in vitro transport experiments and (2) in silico prediction of human renal clearance using in vitro-derived data. The main processes include: (i) establishing transport assays using kidney proximal tubule cells cultured in 96-well plates and Transwell systems, (ii) applying a multi-compartment modeling approach to fit the observed concentration data, and (iii) estimating human renal clearance using a physiologically based kidney model.

For this study, we selected drugs that had renal clearance ranging from hours to days, a sunscreen active ingredient Benzophenone-4 (BP-4) that has been used recently as a case study of next-generation risk assessment to compare a non-animal approach with a traditional safety assessment based on historical animal data for which only predicted human C_max levels are available (Baltazar et al., 2025), and several PFAS with carbon chain lengths from 4 to 11 and known human renal clearance and T_1/2 (from days to weeks to years) (Table 1 and Fig. 2). As could be seen from Fig. 2A, for PFAS, a very strong correlation exists between human renal clearance and T_1/2. By contrast, for drugs, variation in pathways other than renal elimination contributing to total clearance result in no common rate-determining relationship between the two. For PFAS, multiple toxicokinetic studies, as reviewed by (East et al., 2023; Han et al., 2012; Niu et al., 2023), have documented the inverse correlation between the carbon:fluorine chain length and renal clearance in both humans and animals. Fig. 2B shows this relationship for the PFAS in our study, confirming a significant inverse relationship but with high variance, indicating that predictions based solely on the physico-chemical properties of PFAS may have considerable uncertainty. Therefore, our studies aimed to determine if a new approach method (NAM)-based workflow can provide greater confidence in predictions of human body burden for various PFAS and other chemicals.

Fig. 2.

Relationships between renal clearance and compound properties. (A) Correlation between median predicted renal clearance (CL, mL/kg/day) and reported human half-life (T_½, days) across test compounds. (B) Correlation between median observed renal clearance and carbon chain length (C#) for PFAS only. Red dashed lines indicate regression fits, with dotted lines representing 95 % confidence intervals. Squares denote pharmaceutical compounds and circles denote PFAS.

3.1. Cell viability and barrier integrity after chemical treatments

To ensure that test chemical exposures did not compromise cell viability or monolayer integrity, cytotoxicity and barrier function were assessed across all RPTEC cultures. Based on our previous comparative analysis of various culture modalities and cell types, we selected human TERT-immortalized RPTEC cells and their OAT1-overexpressing variant for these studies (Sakolish et al., 2025). We also previously showed that static Transwell cultures and 96-well plates can be used to provide experimental chemical transport and/or intracellular accumulation data suitable for the proposed experimental-computational workflow (Lin et al., 2024). To reduce complexity of the proposed workflow, we conducted Transwell experiments only with hTERT-RPTEC-OAT1 cells. LDH levels in untreated (i.e., medium-only) and vehicle-treated cultures were similar across both platforms and cell types with similar baseline levels. Transwell cultures also demonstrated expected (Srinivasan et al., 2015) TEER values for RPTEC cultures (~180 Ω×cm²), consistent with the formation of confluent epithelial monolayers (Fig. 3A).

Fig. 3.

Cytotoxicity and barrier integrity in RPTEC cultures following compound treatment. (A) Baseline LDH activity and barrier integrity of vehicles prior to treatment. (B) LDH release (reported as % of vehicle) after 24 h treatment with PFAS and reference compounds across plate and Transwell configurations. TAB (positive control) induced marked cytotoxicity, while most treatments showed no significant increase compared to vehicle. See Table S7 for the data. (C) TEER (% of pre-treatment) in Transwells demonstrated maintained barrier function under most conditions, with significant decreases observed for the TAB positive control and for selected PFAS when added basolateral-to-apical (B→A). Asterisks denote significance (p < 0.05) by one-way ANOVA with Dunnett’s post hoc test. Box-and-whisker plots are shown with the box showing the inter-quartile range and whiskers = min–max (n ≥ 3 for 96-well plate studies, n ≥ 4 for Transwell studies).

Following test compound exposures for 24 hrs, LDH release was not significantly elevated relative to vehicle controls for the majority of test articles across all platforms, indicating that the selected concentrations were non-toxic (Fig. 3B and Table S7). As expected, the positive control (tetraoctyl ammonium bromide, TAB) consistently induced significant LDH release across all conditions, demonstrating the sensitivity of the assay. Among test compounds, only perfluorooctanoic acid (PFOA) showed a small but significant increase in LDH release in OAT1-expressing RPTECs exposed in 96-well plates. In the Transwell model, TEER measurements taken 24 hrs post-treatment (Fig. 3C) revealed no significant decrease in barrier integrity for most compounds, with the exception of PFUnDA and PFDA that elicited small but significant decrease in TEER when these compounds were applied from the basolateral side (B→A). The TAB positive control caused a pronounced reduction in TEER, consistent with its established cytotoxic action. These results confirm that the test compounds were administered at sub-to non-toxic concentrations, preserving cell viability and monolayer integrity, which was critical for enabling subsequent chemical transport studies.

3.2. Uptake and transport of test compounds by RPTECs

For the high-throughput experimental model of proximal tubule transport, RPTECs (parental and OAT1-overexpressing lines) were cultured in traditional 96-well plates and exposed to test compounds for 24 h. The amounts of each test compound were evaluated in both cell lysates and media (Fig. 4A–B). Overall, recovery patterns were similar between the two cell types with the exception of tenofovir, a known OAT1 substrate (Kohler et al., 2011). Tenofovir showed enhanced accumulation in the OAT1-overexpressing cells. Longer-chain PFAS (PFUnDA, 8:2 FTS, PFDA) tended to accumulate more in cells as compared to shorter-chain PFAS, the latter largely remained in the cell culture medium. In Transwell cultures of the hTERT-RPTEC-OAT1 cells, recovery was measured from 3 compartments – donor and recipient media, and cell lysates. Notably, several longer-chain PFAS (PFUnDA, 8:2 FTS, PFDA, and PFOS) displayed divergent recovery patterns depending on dosing orientation, with differences between apical (A→B; Fig. 4C) and basolateral (B→A; Fig. 4D) treatment, suggesting directional uptake. By contrast, most other compounds showed limited transport regardless of treatment orientation. The data on recovery amounts (pmol) in each compartment – two in 96-well plates and 2 or 3 in Transwells – were used as the input for the compartmental modeling of renal clearance.

Fig. 4.

Compound recovery in RPTEC cultures across platforms. Amounts of each test compound recovered (pmoles) after 24 h treatment in (A) hTERT-RPTEC (96-well plates), (B) hTERT-RPTEC-OAT1 (96-well plates), (C) hTERT-RPTEC-OAT1 Transwells dosed apical-to-basolateral (A→B), and (D) hTERT-RPTEC-OAT1 Transwells dosed basolateral-to-apical (B→A). Recovery was measured from donor media (white), cell lysates (orange), and recipient media (purple). Results highlight differences in distribution between compartments and across culture formats. Box-and-whisker plots are shown with the box showing the inter-quartile range and whiskers = min–max (n ≥ 3 for 96-well plate studies, n ≥ 4 for Transwell studies).

To further illustrate the differences in test compound distribution among compartments, Fig. 5 presents ratios of measured concentrations. In 2D 96-well plate monolayer cultures (Fig. 5A), PFAS demonstrated a clear carbon:fluorine chain-length-dependent enrichment in cells, whereas shorter-chain PFAS and most pharmaceuticals remained partitioned in the medium. Transwell cultures (Fig. 5B) demonstrated similar trends, with long-chain PFAS exhibiting marked intracellular accumulation. Directional transport analysis (Fig. 5C) revealed that most compounds were retained on the donor side when applied apically (A→B), whereas basolateral administration (B→A) often resulted in greater transfer to the apical (e.g., recipient) compartment, consistent with polarized secretion rather than absorptive uptake in these proximal tubule–like cells. Notably, tenofovir again showed higher intracellular association, in line with its rate-determining status as an OAT1 substrate.

Fig. 5.

Donor to recipient ratios across culture formats. (A) Ratio of cell (lysate) to media concentrations after 24 h exposure in hTERT-RPTEC (Parent) and hTERTRPTEC-OAT1 (OAT1) cultured in 96-well plates. (B) Cell to media (donor) ratios in hTERT-RPTEC-OAT1 Transwells following apical-to-basolateral (A→B) or basolateral-to-apical (B→A) dosing. (C) Ratio of recipient to donor media concentrations in Transwells for both dosing orientations. The red dashed line denotes equilibration (equal concentration in donor and recipient media). Values to the right of the line indicate higher levels in the recipient compartment (cells or recipient media), suggesting selective transport, while values to the left indicate retention in the donor compartment, consistent with barrier function or limited transport.

3.3. Transwell-based compartmental model fitting using MCMC algorithm

For each chemical, a unified set of permeability parameters was estimated by simultaneously fitting the chemical concentration data from both “A→B” and “B→A” assays using a compartmental model. The raw concentration data for each replicate, along with the statistical information used for model fitting, are available in the GitHub repository. All simulations were conducted using Markov Chain Monte Carlo (MCMC) sampling and achieved convergence within 300,000 iterations per chain, with outputs recorded every 30 iterations. For posterior inference, the first one-fifth of iterations from each chain were discarded as burn-in. All parameters yielded potential scale reduction factors $\left(\widehat{\text{R}}\right)$ below 1.2, indicating adequate convergence across the four independent chains. The final posterior distributions and detailed $\widehat{\text{R}}$ values for all parameters are also provided in the GitHub repository (see Section 2.7). Model performance is illustrated in Figure S1, which compares simulated and observed concentrations in both apical/basolateral wells and cell lysates. The simulated values showed strong agreement with the experimental data, with mean absolute deviations ranging from 0.02 to 0.73 and mean errors from −0.06–0.13. Pearson and Spearman correlation coefficients were generally above 0.8, except for the Spearman coefficients of PFUnDA, PFDA, 8:2 FTS, and PFOS, which ranged from 0.65 to 0.74. These results demonstrate the robustness of the model and support the use of the derived posterior permeability distributions for predicting in vivo human renal clearance for each chemical compound.

3.4. Predictions of renal clearance by IVIVE – model development study

The permeability ratio (P ratio) is the key parameter in the calculation of renal clearance. It is defined as the ratio of secretion to reabsorption permeability. In 2D cell cultures, it can also be calculated as the ratio of concentrations in the medium to those in the cells. Figure S2 presents the P ratios determined either from cellular and medium concentrations in 96-well plates or from posterior distributions derived via compartmental modeling of Transwell-derived data. We also applied the P ratio calculation method used for 2D culture to Transwell assays by only considering concentrations in the donor chamber and within the cells. In general, the P ratios derived from 96-well plates showed greater differences across chemicals compared to those from the three-compartment model applied to Transwell data. However, when the data from a Transwell model was treated similarly to 2D cultures, the P ratio trends aligned closely with those from the 96-well plates, regardless of whether “A→B” or “B→A” transport assays were used. These P ratios served as inputs for the physiologically based kidney model to predict human renal clearance.

Fig. 6 compares the predicted total renal clearance (y-axes in all plots) from our in vitro–in silico approach with in vivo values obtained from the literature (x-axes in all plots), separately for data from 96-well plates and Transwells. For data from 96-well plates (Fig. 6A), predictions for PFAS exhibited very high correlations for both parent and OAT1-overexpressing RPTEC cells (r = 0.93 and ρ = 0.98 for parental cells; r = 0.92 and ρ = 0.97 for OAT1-overexpressing cells), but with consistent underprediction by about one order of magnitude. For drug predictions, however, correlations were lower with higher variance, with those using data from parent cells showing somewhat higher correlation than those using OAT1-overexpressing cells (r = 0.56 and ρ = 0.39 for parental cells; r = 0.23 and ρ = 0.18 for OAT1-overexpressing cells). This discrepancy can be largely attributed to the small sample size and an effect of the data for tenofovir. Specifically, the predicted renal clearance of tenofovir using OAT1-overexpressing cells was significantly lower than that using parental cells, which corresponded to a roughly 100-fold lower P ratio (Figure S2). This suggests that OAT1 over-expression increases intracellular accumulation of tenofovir, reducing its predicted renal elimination.

Fig. 6.

Comparison of predicted (this study) and observed (published) human renal clearance values. Predicted total renal clearance (CL_renal,total) from in vitro–in silico models was compared with reported in vivo values for PFAS (black circles) and pharmaceutical compounds (purple squares). (A) Predictions from 96-well plate studies using two-compartment models with hTERT-RPTEC (filled icons) and hTERT-RPTEC-OAT1 cells (open icons). (B) Predictions from hTERT-RPTEC-OAT1 Transwell assays using three-compartment modeling (left) or two-compartment modeling for apical-to-basolateral (middle) and basolateral-to-apical (right) dosing. (C) Same as (A) for the validation experiment. Dashed black lines represent regression fits, and the red dotted line indicates perfect agreement. Pearson (r) and Spearman (ρ) correlation coefficients are shown for PFAS (black) and pharmaceutical compounds (purple). See Table S8 for the data.

For results based on data from Transwells (Fig. 6B), the best performance was for drug predictions based on three-compartment modeling, which exhibited good correlations (r = 0.68 and ρ = 0.60) as well as low variance, representing a substantial improvement over results from 96-well plates. When treating Transwell data with a two-compartment approach (i.e., treating the system similarly to 96-well plates), predictions of drug renal clearance were comparable to those from 96-well plate studies, with low correlations (r = 0.25 and ρ = 0.26 for A→B data; r = 0.23 and ρ = −0.03 for B→A data) and high variance. By contrast, Transwell data on PFAS did not markedly improve prediction accuracy or precision. For instance, the use of three-compartment modeling with Transwell data led to reduced correlation coefficients for PFAS (r = 0.69 and ρ = 0.55), with a worsening trend towards overprediction for PFAS with observed renal clearance below 1 mL/kg/day. When treating PFAS Transwell data with a two-compartment approach, although correlations were high (r = 0.95 and ρ = 0.94 for A→B data; r = 0.92 and ρ = 0.95 for B→A data), there was a notable trend towards worsening underprediction as renal clearance values became smaller.

3.5. Predictions of renal clearance by IVIVE – validation study for PFAS

To test the robustness of our IVIVE workflow, we conducted a follow-up validation study using a slightly different but largely overlapping set of PFAS. These compounds were selected based on the availability of the analytical methods in another laboratory but representing the range in chain length and physico-chemical properties. In vitro experiments were performed by another scientist and quantitation of PFAS in cell medium and cell lysates were performed in a second (external) laboratory employing the same type of targeted quantitation (i.e., LC-MS/MS), although on distinct instrumentation (Agilent vs. Waters, see Methods for details) and using varied ionization modes and other instrument parameters (see Section 2.5 – validation study; and Table S1 for instrumental settings between studies). Based on the outcomes of the model development study showing that 2D cell cultures are sufficient for prediction of renal clearance of PFAS (Fig. 6A–B), this validation study was conducted only in 96-well plate format, with in silico analyses using the two-compartment model approach. Similar to the model development study results, the in vitro–in silico predicted renal clearance shows very high correlation with observed in vivo human values (Fig. 6C) and very strong ability to discriminate slower and faster renal clearance. However, the absolute results from the validation study shows a greater degree of underprediction as compared to the published data – about 100-fold rather than 10-fold. For PFAS that overlapped between the model development and validation studies, there is a high correlation between predicted renal clearance values (Spearman rho=0.92, Pearson r²=0.97). These validation data provide support for the robustness of our overall workflow for predicting relative values of PFAS clearance (Table 2 and Table S8).

Table 2.

Predicted relative renal clearance and human half-life for the PFAS tested in these studies using PFOA as index chemical. See Table S8 for the clearance values and data.

Compound	CAS#	Relative Renal Clearance Factor (Predicted CLrenal/CLrenal,PFOA)		Predicted human half-lifeb
		Model Development Studya	Validation Studya
PFPeA	2706–90–3	500	-	Days
PFBA	375–22–4	500	300	Days
4H-PFBA	679–12–9	-	200	Days
3:3 FTCA	356–02–5	-	60	Days
PFBS	375–73–5	2000	20	Days/Months
PFMBA	863090–89–5	-	10	Months
PFHxA	307–24–4	200	6	Days/Months
4:2 FTS	757124–72–4	10	-	Months
6:2 FTS	27619–97–2	5	-	Months
5:3 FTCA	914637–49–3	-	3	Months
NFDHA	151772–58–6	-	3	Years
PFOA	335–67–1	1 c	-	Years
PFOA-K	2395–00–8	-	1 c	Years
9H-PFNA	76–21–1	-	0.9	Years
PFPE–1	801212–59–9	-	0.8	Years
PFEESA	113507–82–7	-	0.8	Years
PFHpA	375–85–9	1	0.4	Years
PFHxS-K	3871–99–6	-	0.4	Years
PFHxS	355–46–4	3	0.2	Years
PFNA	375–95–1	0.9	0.3	Years
NH4PFOA	3825–26–1	-	0.4	Years
8:2 FTS	39108–34–4	0.2	-	Years
PFPE–5	137780–69–9	-	0.1	Years
PFOS	1763–23–1	0.1	0.06	Years
PFOS-K	2795–39–3	-	0.07	Years
PFHpS	375–92–8	-	0.04	Years
PFDA	335–76–2	0.04	0.07	Years
PFUnDA	2058–94–8	0.002	0.01	Years

^a,See Methods for the description of the PFAS studies. Relative clearance factors were rounded to 1 significant digit.

^b,Predicted human half-life values were based on dividing index chemical half-life (PFOA or PFOA-K, with half-life of 3.14 years from (Chiu et al., 2022)) by relative renal clearance, which assumes all PFAS have same volume of distribution and fraction of total clearance that is renal. Days:30 days; Months: 30–364 days; Years: ≥ 365 days.

^c,Index chemical

4. Discussion

The overall goal of this study was to determine whether a unified experimental-computational workflow can be devised to accurately predict human renal clearance of both drugs and non-pharmaceuticals with pharmacokinetics that involve active transport in the kidney. This challenge is particularly critical for chemically diverse and environmentally persistent classes such as PFAS. Thousands of PFAS are in commercial use or found in environmental media, yet human toxicokinetic data are available for only a handful (East et al., 2023). PFAS exhibit a range of elimination half-lives from days to many years, differences that reflect chain length, functional group type, protein binding potential, and, critically, their interactions with renal transporters in the proximal tubule (Pizzurro et al., 2019; Schulz et al., 2024; Smeltz et al., 2023). For some PFAS, such as PFOS, efficient renal reabsorption, rather than poor filtration, appears to explain their long persistence and bioaccumulation in humans (Fujii et al., 2015; Harada et al., 2005). However, experimental human data to determine plasma half-life (T_1/2) is available for only a relatively small number of PFAS, and chemical property- and/or computational-based approaches have low precision or can only provide “binning” into broad categories (Dawson et al., 2023).

Taken together, our results suggest the use of complete Transwell data combined with three-compartment modeling provides the most accurate predictions of renal clearance, having strong concordance with in vivo data both for relative and absolute predictions across chemicals, but only for more rapidly (hours to days) cleared compounds. This model is likely to be most relevant for small molecule pharmaceuticals, which are not intended to remain in systemic circulation over months or years. Although absolute renal clearance values for such compounds were slightly underpredicted in our model by on average about 3-fold, they are no more uncertain than allometric scaling of drugs from animal studies, which have been found to have geometric fold errors of 2- to 7-fold depending on which species is/are used (Lombardo et al., 2013). Thus, as preclinical data supporting an investigational new drug (IND) application for pharmaceuticals, this approach will have promise if combined with other IVIVE-based methods to address absorption, metabolism, and plasma protein binding.

By contrast, for slowly cleared compounds such as PFAS, it is more challenging to make predictions of absolute renal clearance values. Consistent with our previous study (Lin et al., 2024), relative differences in renal clearance between low- and high-clearance PFAS can be readily distinguished using a relatively simple and high-throughput experimental design with 96-well plate experiments to derive culture medium and cell lysate concentrations followed by a two compartment model. We also found that one reasonable general approach, when no a priori knowledge is available for the renal clearance of a test compound, is to start with a Transwell model, but implement a tiered experimental and modeling approach. First, one can analyze intra-cellular concentrations of a compound and if significant intracellular accumulation is observed, the donor medium compartment concentrations would suffice as additional input for a two-compartment model using donor and cellular concentrations. If intracellular concentrations are found to be low as a fraction of total amount tested, both donor and recipient measurements will need to be determined for a three-compartment model.

This study used two RPTEC cell models, a transformed cell line and its OAT1-overexpressing variant, both widely used in pharmacokinetic studies (Aschauer et al., 2015; Wieser et al., 2008). A number of immortalized RPTEC cell lines have been used in toxicology studies (Bajaj et al., 2018), and our previous work employed the RPTEC/TERT1 line to successfully demonstrate the proof of concept for an in vitro–in silico workflow for predicting PFAS clearance (Lin et al., 2024). However, these cells lack robust expression of many renal transporters (Sakolish et al., 2025) that are critical for the disposition of various drugs and chemicals (Hagos and Wolff, 2010). Therefore, given that OAT1 is among the most important drug transporters in the proximal tubule, we tested the utility of an OAT1-overexpressing variant of the RPTEC/TERT1 cell line (Nigam et al., 2015) by incorporating this additional cell model into our study design. As expected, we observed differences in transport for tenofovir, a known substrate for OAT1 (Parker et al., 2023); however, no effect was observed for the other compounds. For PFAS, PFOA and PFOS have been extensively studied with respect to the role of renal transporters (Fujii and Harada, 2025). Previous studies have shown that multiple transporters may contribute to the efficient reabsorption of other PFAS in the human kidney, and that transporter specificity depends on molecular weight and lipophilicity (Ryu et al., 2024). Based on our data, we reason that an initial screen using an immortalized RPTEC line is a sensible first step. However, inclusion of overexpressing variants of these lines may have limited utility for substances whose transport depends on multiple transporters or on transporters for which such variants are not available.

While this proposed workflow may be most practical for screening purposes, for highly accumulating compounds, the two-compartment model based on the Transwell, or 96-well plate data will only provide relative ranking of clearance values. For these types of substances, it may still be possible to make predictions of absolute renal clearance levels by including an “index chemical” in each experiment, similar to Toxic Equivalent Factors for dioxin and dioxin-like compounds or Relative Potency Factors (RPFs) for polycyclic aromatic hydrocarbons (Van den Berg et al., 1998). For instance, using PFOA as an “index chemical,” a Relative Renal Clearance Factor (RRCF) can be calculated for other PFAS by dividing their renal clearance predictions by the prediction for PFOA in the same experiment. Although either Transwell or 96-well plate (2D) data can be used, our results suggest that the 96-well data would be preferred, because Fig. 6A (96-well plate data) shows a correlation slope for PFAS closer to one, better representing relative clearance throughout the very wide range of predicted and observed renal clearance. By contrast, Fig. 6B (Transwell two-compartment data) shows that the correlation slope is steeper than one, with more underprediction for lower clearance PFAS. This proposed RRCF approach is illustrated in Table 2, including the results from both the model development and follow-up validation studies across all 28 tested PFAS, with PFOA set as the index chemical (with RRCF=1). To facilitate interpretation, Table 2 also makes estimates of overall half-life in humans by dividing the PFOA half-life by the RRCF. This illustration shows a clear distinction between relatively rapidly cleared PFAS with half-lives of days to months, and very slowly cleared PFAS with half-lives of years, and that this discrimination is largely consistent between the main and validation studies.

Our study has a number of limitations. The first limitation of the approach proposed herein, particularly for estimating (relative or absolute) renal clearance for a large number of compounds, is that the analytical chemistry remains a bottleneck. In particular, traditional targeted quantitation methods may require custom method development depending on the instrumentation available in each laboratory, and complex sample preparation and analysis steps, which can take a significant amount of time, particularly when evaluations include a diverse chemical set, a large number of samples, or when accuracy and precision are required (Vervoort et al., 2021). However, our proposed in silico modeling approach, particularly when using 96-well plates, only requires relative quantitation; therefore, faster non-targeted analytical methods such as Ion Mobility Spectroscopy-Mass Spectrometry, may be used (Dodds et al., 2025; Ibrahim et al., 2016; Teri et al., 2025). These methods should be tested for their applicability and integration into our in vitro-in silico workflow. Second, for drugs, we only tested a relatively small number of substances for comparisons with renal clearance predictions. Future studies will be needed to evaluate the accuracy of our approach with a larger set of pharmaceuticals. Third, we only conducted a follow-up validation study for PFAS, and the robustness of the data for pharmaceuticals needs to be further investigated. Fourth, our ability to evaluate model performance for PFAS is considerably hampered by the fact that only a handful of PFAS have been studied for the extent of metabolism or have measured human clearance values. Even when available, human PFAS clearance values have high degree of uncertainty and inter-individual variability. We believe that the data provided by our study is already the largest available dataset comparing predicted and observed renal clearances for PFAS, one that is unlikely to be substantially enlarged in the near future because of the lack of human data. Fifth, it has been shown recently that biliary/fecal clearance can be a major contributor to overall clearance of PFOA and PFOS (Andersson et al., 2025); thus, even though the data on fecal clearance of PFAS is now limited to PFOA and PFOS, only accounting for renal clearance may underestimate estimates of total clearance, resulting in over-prediction of bioaccumulative potential for PFAS. Finally, for both drugs and non-pharmaceuticals, it should be noted that our in vitro experiments used a single donor-derived cells, so they do not address population variability and additional research is needed to determine the extent of inter-individual variability in renal clearance and xenobiotic metabolism.

In conclusion, the present study extends the domain of applicability and demonstrates robustness of our previously proposed in vitro–in silico workflow developed to improve the prediction of human renal clearance (Lin et al., 2024; Sakolish et al., 2025). We systematically evaluated several human proximal tubule cell types and in vitro platforms, from high-throughput 2D well plates to traditional Transwells, and extended the approach from three to 36 chemicals representing both drugs and PFAS and having a very wide range of human half-lives. We found results consistent with our initial studies, with Transwell model leading to more accurate absolute renal clearance predictions, but only for relatively rapidly cleared compounds such as drugs. The higher throughput 2D model provided accurate relative renal clearance predictions for slowly cleared compounds such as PFAS. We also conducted a validation study where a set of PFAS was independently tested in the 96-well plate model and a different analytical method. Together, the results of this study demonstrate that our overall in vitro-in silico workflow is a robust, scalable, mechanistically informed, and empirically benchmarked strategy for generating human-relevant renal clearance predictions for both drugs and environmental chemicals.

Appendix A. Supporting information

Supplementary data associated with this article can be found in the online version at doi:10.1016/j.tox.2025.154336.

Data availability

All raw datasets and code used for modeling and analysis are publicly accessible in the GitHub repository mentioned in section 2.8 and external database mentioned in section 2.9