Optimization of Bottom‐Up PBPK Model Development in SIMCYP via Retrospective Analysis of Clinical Human PK Data

ABSTRACT

Physiologically‐based pharmacokinetic (PBPK) modeling has become a major tool in drug discovery and development. Here, we describe the bottom‐up PBPK modeling approaches employed at AbbVie using Simcyp Simulator and evaluate the impact of three system parameters, GI physiology, P‐gp Relative Expression Factor (REF), and recombinant CYP enzyme (rCYP) intersystem extrapolation factor (ISEF), independently and in combination, on PBPK prediction performance through retrospective analysis of 8 clinical assets. Overall, the application of New GI physiology resulted in a considerable improvement in the prediction of oral absorption for most compounds compared to the Original GI physiology (C_max: 76% vs. 43% within 3‐fold) when using the default P‐gp REF (1.5) and adjusted ISEF. Decreasing P‐gp REF to 0.5 resulted in additional improvement in the predictions of C_max for P‐gp substrates (86% within 3‐fold). The observed plasma exposure‐time profiles and AUC_INF are better predicted using the default rCYP ISEF values instead of individually adjusted values (48% vs. 43% within 3‐fold) when using the Original GI and default P‐gp REF (1.5). The combination of optimized parameters (New GI physiology, P‐gp REF of 0.5 and rCYP default ISEF) predicted the plasma exposures (AUC_INF and C_max) within 3‐fold for 81% and 86% of the tested simulations, respectively. In conclusion, the present study proposes an optimized strategy for bottom‐up PBPK model development in Simcyp Simulator. Retrospective comparison with observed clinical PK data is vital for model verification as well as further improvement in prospective predictions for future drug candidates.

Study Highlights.

• What is the current knowledge on the topic?

  • ○
    Bottom‐up physiologically based pharmacokinetic (PBPK) modeling is routinely applied in drug discovery to inform first‐in‐human study design and drug–drug interaction risk. However, the performance of bottom‐up PBPK models varies depending on the complexity of physicochemical and ADME properties of a molecule.
• What question did this study address?

  • ○
    In order to optimize the strategy of bottom‐up PBPK model development, a retrospective analysis of clinical assets was conducted focusing on the impact of three system parameters (GI physiology, P‐gp REF and rCYP ISEF) in Simcyp Simulator on oral absorption and elimination predictions.
• What does this study add to our knowledge?

  • ○
    We provide the first validation of the New GI physiology in Simcyp Simulator, which improves C_max predictions over the Original physiology. System‐specific REF value further improved oral absorption predictions for P‐gp substrates, and default ISEF values for rCYPs provided the most accurate AUC_INF predictions.
• How might this change clinical pharmacology or translational science?

  • ○
    An optimized strategy for bottom‐up PBPK model development is proposed. The new strategy is expected to improve the performance of prospective predictions for future drug candidates.

Abbreviations

ADAM

advanced dissolution, absorption and metabolism model

ADME

absorption, distribution, metabolism, and excretion

AFE

absolute fold error

AUC_INF

area under the plasma concentration time curve to infinity

CL_int

intrinsic clearance

CL_int,u,CYP

unbound intrinsic clearance in rCYP

CL_int,u,hep

unbound intrinsic clearance in human hepatocytes

CL_{int,u,matched}

unbound intrinsic clearance from rCYP matched to human hepatocytes

C_max

maximum plasma concentration

C_Obs

observed plasma concentration

C_Pred

predicted plasma concentration

CSR

critical supersaturation ratio

CYP

cytochrome P450

ER

extraction ratio

f _a

fraction absorbed

FIH

first‐in‐human

f _m,additional

fraction metabolized by non‐CYP enzymes

f _m,met

fraction metabolized by a CYP enzyme in hepatocytes

f _m,rCYP

fraction metabolized by a rCYP

f _u,gut

fraction unbound in the gut

f _u,inc

fraction unbound in the incubation

f _u,mic

fraction unbound in microsomes

f _u,p

fraction unbound in plasma

GI

gastrointestinal

IR

immediate release

ISEF

intersystem extrapolation factor

ISEF_adjusted

adjusted intersystem extrapolation factor

ISEF_default

default intersystem extrapolation factor

IVIVE

in vitro‐to‐in vivo extrapolation

J _max

maximal velocity of transport processes

k _a

absorption rate constant

K _m

Michaelis–Menten constant for transport or enzymatic processes

K _p

tissue:plasma partition coefficient

LogKm:w

logarithm of bile micelle:buffer partition coefficient

MAFE

median absolute fold error

MechPeff

mechanistic passive regional permeability predictor

PBPK

physiologically‐based pharmacokinetic

P _eff,man

human intestinal passive permeability

P‐gp

P‐glycoprotein

PK

pharmacokinetics

PRC

precipitation rate constant

P _trans,0

intrinsic transcellular permeability

R _BP

blood to plasma ratio

rCYP

Gentest Supersome Human Recombinant CYP

REF

relative expression factor

UBL

unstirred boundary layer

V _max

maximal velocity of enzymatic processes

V _ss

volume of distribution

σ

standard deviation

1. Introduction

Physiologically‐based pharmacokinetic (PBPK) modeling is a major tool in drug discovery and development. Common applications include first‐in‐human (FIH) dose projections, drug–drug interaction (DDI) assessments, and clinical trial design [1, 2, 3]. PBPK modeling combines mechanistic and physiological models with drug‐specific parameters to allow dynamic prediction of human pharmacokinetics (PK) [1, 2, 4]. Software packages, such as Simcyp Simulator, connect different organs in the body with a mathematical framework considering the physiology such as organ size and blood flow [5]. Due to its mechanistic nature, a PBPK model can evaluate the impact of intrinsic factors (e.g., age, sex, ethnicity, genetics and disease state) and extrinsic factors (e.g., concomitant medications, smoking and diet) on PK [1, 2]. Additionally, a PBPK model enables the dynamic simulation of absorption, distribution, metabolism, and excretion (ADME) considering time‐dependent processes, including saturation of absorption/elimination processes, precipitation of a molecule in the intestinal lumen, and inhibition and induction of enzymes and/or transporters.

During PBPK model development, drug‐specific parameters can be assembled using three distinct approaches: bottom‐up, top‐down, or middle‐out [2]. Top‐down and middle‐out approaches leverage available human PK data in order to either obtain drug‐specific parameters by fitting a PBPK model to observed data (top‐down) or optimize certain parameters to better explain the observation (middle‐out). Prediction performance of these models is high since they are validated and qualified during model development [4, 6]. In contrast, a bottom‐up PBPK model is built solely from in vitro ADME and preclinical in vivo PK parameters. Such models are primarily used to inform FIH study design and DDI risk before human data becomes available. The performance of bottom‐up models varies depending on the complexity of physicochemical and ADME properties of a molecule [2, 7, 8]. Given the widespread use of PBPK modeling during drug discovery, an optimization of the bottom‐up PBPK modeling strategy is warranted. Several publications have become available over the last decade illustrating the validity and accuracy of different PBPK modeling approaches [3, 5, 9, 10, 11]. For example, in vitro‐to‐in vivo extrapolation (IVIVE) of physicochemical properties and in vitro ADME parameters were evaluated by developing preclinical PBPK models, followed by the incorporation into bottom‐up human PBPK models for individual compounds (PBPK‐IVIVE approach) [10]. Others explored an integration of the Wajima method to PBPK modeling in order to improve the distribution volume (V _ss) and PK profile prediction [3] or a bottom‐up PBPK modeling using human in vitro data and appropriate intersystem scaling factors (e.g., relative transporter expression factors) [5, 9, 11]. These approaches resulted in reasonable performance of bottom‐up human PK predictions in general. However, such research often describes only a snapshot in time without considering the impact of changes to modeling software and approaches.

One recent update in a modeling software, Simcyp Simulator, was the introduction of the New gastrointestinal (GI) physiology in version 20 [12, 13, 14, 15]. Human fraction absorbed (fa) has often been underpredicted in Simcyp Simulator particularly for low solubility compounds, resulting in the underprediction of plasma concentrations after oral administration [16]. To address this, the New GI model was developed through an extensive meta‐analysis of lengths and diameters of the human small intestine and colon, blood flow, transit time and surface area. Simcyp advises using the New GI model along with the Advanced Dissolution, Absorption and Metabolism (ADAM) model and unstirred boundary layer (UBL) fluid volume in the development of PBPK models for low solubility compounds. However, a systematic analysis of the performance of the New GI model compared to the Original model has not been reported.

In the present study, we describe the bottom‐up PBPK modeling approaches employed at AbbVie. over the past decade and evaluate the impact of different approaches on PK prediction performance through a retrospective analysis of internal clinical assets. Specifically, the effects of three system parameters, GI physiology, P‐glycoprotein (P‐gp) relative expression factor (REF), and recombinant CYP enzyme (rCYP) intersystem extrapolation factor (ISEF), were explored on oral absorption and elimination predictions. An optimized strategy for bottom‐up PBPK model development is proposed through the retrospective analysis.

2. Methods

2.1. Bottom‐Up PBPK Model Development

Bottom‐up PBPK models were developed in the Simcyp Simulator, Version 21 Release 1 (Certara USA Inc., Princeton, NJ) where only measured in vitro drug parameters and preclinical PK data were used. The compound properties are summarized in Table S1. All simulations were run for 100 healthy North‐European Caucasian individuals (10 trials of 10 individuals) using either the Original or New GI physiology selected. Other conditions (e.g., trial duration, fed vs. fasted) were selected to match the respective clinical trial designs for each compound (Table S2). The fasted condition was used in the simulations when no significant food effect was observed in the clinic.

2.2. Absorption

ADAM model was used to predict oral absorption [16]. The UBL fluid volume parameter was applied for ABBV‐X1. Dissolution from immediate release (IR) formulation was modeled using the diffusion layer model for selected assets (Table S2). The intrinsic solubility was calculated from the drug solubility measured between pH 1 and pH 7.4 with SIVA tool kit [17]. Bile‐micelle mediated enhancement of solubility was accounted for by fitting a logarithm of bile micelle: buffer partition coefficient (LogKm:w) using the measured fasted and fed state simulated intestinal fluid (FaSSIF, FeSSIF) solubility data [17]. In the absence of measured values, supersaturation and precipitation were assumed to be minimal (critical supersaturation ratio (CSR) = 1; precipitation rate constant (PRC) = 0.01 h⁻¹). The mechanistic passive regional permeability predictor (MechPeff) model was used to predict intestinal passive permeability [18, 19] with an optimized intrinsic transcellular permeability (P _trans,0) value that matches P _eff,man J1 to the value predicted with the apparent passive permeability determined in MDCKII‐WT_ZFN cells [20]. P _eff,man was calibrated using an appropriate combination of permeability reference compounds (atenolol, cimetidine, midazolam, metoprolol, propranolol and verapamil) for each test compound. Fraction unbound in the gut (f _u,gut) was set to a default of 1 as this value is not measured. A sensitivity analysis on f _u,gut indicated a minimal impact of this parameter on fa and Fg, where the decrease in f _u,gut from 1 to 0.1 resulted in < 2‐fold increase in the predicted fa and Fg (Figure S1). Efflux transporter effect in the gut was extrapolated from in vitro kinetics values (K _m and J _max) determined in MDCKII‐MDR1_ZFN and MDCKII‐BCRP_ZNF cells [20, 21]. P‐gp transporter kinetics were scaled using a range of P‐gp REF values between the default value in Simcyp Simulator for MDCK cell lines (REF: 1.5) and an in‐house value obtained from historical data, that indicated that a 3‐fold reduction in REF is commonly required during middle‐out model adjustment (REF: 0.5). BCRP transport kinetics were scaled with the default value for MDCK cell lines (REF: 1).

2.3. Distribution

Full PBPK disposition model with method 2 was used to predict volume of distribution (V _ss) with tissue: plasma partition coefficient (K _p ) scalars to match the V _ss predicted from preclinical species. Details of the preclinical PK studies are available in Methods S1. The fraction unbound in plasma (f _u,p) was determined as previously described using a conventional equilibrium dialysis for ABBV‐X1, ABBV‐X2, ABBV‐X4, ABBV‐X5, ABBV‐X6, ABBV‐X7 and ABBV‐X8, and flux dialysis for ABBV‐X3 [22, 23]. Blood to plasma ratio (R _BP) was determined as previously described [23].

2.4. Elimination

Human plasma clearance was predicted using the perfusion‐limited liver model. The metabolic clearance was determined in recombinant CYPs (rCYP; Gentest Supersomes, Discovery Life Sciences, Woburn, MA, USA) as described in Methods S1. Those were incorporated into the PBPK model using enzyme kinetics (K _m and V _max) for major enzymes and intrinsic clearance (CL_int) for additional metabolic pathways either with the default Simcyp ISEF values for Gentest Supersomes or in‐house adjusted ISEF values as described below. Fraction unbound in the incubation (f _u,inc) in the rCYP assay was determined from the unbound fraction in microsomes (f _u,mic) scaled for the protein content [24]. Non‐CYP mediated metabolic clearance (CL_{int,u,non‐CYP}) was included in the model as additional hepatic elimination as needed using Equation (1) below.

$${\mathrm{CL}}_{\mathrm{int},\mathrm{u},\text{non}-\mathrm{CYP}}=\frac{\sum {\mathrm{CL}}_{\mathrm{int},\mathrm{u},\mathrm{CYP}}}{f_{\mathrm{m},\mathrm{CYP}}}\times f_{\mathrm{m},\text{additional}}$$ (1)

where CL_int,u,CYP is the unbound intrinsic clearance in the rCYP assay f _m,CYP and f _m,additional are contributions of CYP‐mediated and non‐CYP‐mediated metabolic clearance to the total hepatic metabolism, respectively, as determined in human hepatocytes with or without a pan‐CYP inhibition regimen. Active hepatic scalar was measured as the ratio of uptake in plated human hepatocytes with or without transporter chemical inhibition (1 mM rifamycin SV and 0.1 mM pyrimethamine) and used in the PBPK model to account for the impact of active hepatic uptake on elimination. In the absence of kinetic studies, active hepatic uptake was assumed to be non‐saturable. Biliary clearance observed after intravenous dosing to bile‐duct cannulated rats was incorporated into the model as necessary assuming no species differences. In brief, the percentage of parent compound excreted in the bile after intravenous administration to bile‐duct cannulated rats was used along with the predicted human clearance in order to predict the biliary clearance in humans, which was then incorporated as CL_int (bile) in a PBPK model after converting the unit to μL/min/millions of cells using physiological scalers. Additional human renal clearance was predicted using allometric scaling of rat renal clearance with an exponent of 0.75 for ABBV‐X8 [25].

2.5. Adjustment of Intersystem Extrapolation Factor (ISEF)

The adjusted ISEF (ISEF_Adjusted) was calculated by comparing the unbound intrinsic clearance in rCYP (CL_int,u,CYP) to that in human hepatocytes (CL_int,u,hep) as described below. The fraction metabolized by each rCYP (f _m,rCYP) within total CYP‐mediated clearance is calculated using Equation (2). f _m,rCYP is converted to the respective fraction metabolized by the same enzyme in hepatocytes (f _m,met) by correcting for the fraction metabolized by non‐CYP enzymes (f _m,additional; Equation (3)), which is the fraction of metabolism remaining in the presence of the pan‐CYP inhibition regimen.

$$f_{\mathrm{m},\text{rCYP}}=\frac{{\mathrm{CL}}_{\mathrm{int},\mathrm{u},\mathrm{CYP}}}{\sum {\mathrm{CL}}_{\mathrm{int},\mathrm{u},\mathrm{CYP}}}$$ (2)

$$f_{\mathrm{m},\mathrm{met}}=f_{\mathrm{m},\text{rCYP}}\times \left(1-f_{\mathrm{m},\text{additional}}\right)$$ (3)

The matched CL_int,u for each CYP (CL_{int,u,CYP,matched}) is calculated from the CL_int,u,hep and f _m,met using Equation (4).

$${\mathrm{CL}}_{\mathrm{int},\mathrm{u},\mathrm{CYP},\text{matched}}={\mathrm{CL}}_{\mathrm{int},\mathrm{u},\mathrm{hep}}\times f_{\mathrm{m},\mathrm{met}}$$ (4)

The adjusted ISEF (ISEF_Adjusted) (Equation (6)) is calculated from the default ISEF (ISEF_Default) and the ISEF Scalar, which is the ratio between CL_{int,u,CYP,matched} and CL_int,u,CYP (Equation (5)).

$$\text{ISEF Scalar}=\frac{{\mathrm{CL}}_{\mathrm{int},\mathrm{u},\mathrm{CYP},\text{matched}}}{{\mathrm{CL}}_{\mathrm{int},\mathrm{u},\mathrm{CYP}}}$$ (5)

$${\text{ISEF}}_{\text{Adjusted}}={\text{ISEF}}_{\text{Default}}\times \text{ISEF Scalar}$$ (6)

2.6. FIH Clinical Trial Protocols

Included FIH studies were randomized, double‐blinded, placebo‐controlled studies performed to evaluate the safety, tolerability, and pharmacokinetics of internal investigational drugs in healthy adult participants. All studies were conducted in accordance with Good Clinical Practice guidelines and the ethical principles that have their origin in the Declaration of Helsinki. The protocols and informed consent forms were approved by the respective institutional review boards and participants provided written informed consent before any study‐related procedures were performed. Participants were assigned to ascending dose groups and randomized to receive a single dose of either the study drug or placebo. Study drugs were administered orally with 240 mL water under fasting conditions after a minimum of 10 h fast and approximately 4 h before lunch or under non‐fasted or fed conditions approximately 30 min after a meal (moderate fat diet). Participants were confined to the study site and supervised for the entire study duration. Blood samples for pharmacokinetic assessment were collected prior to dosing and at appropriate timepoints after drug administration. The respective clinical trial protocols are summarized in Table S2.

2.7. Model Performance

Model prediction performance for each compound was assessed by comparing the predicted and observed human PK data. Exposure (area under the concentration‐time curve to infinity [AUC_INF] and maximum plasma concentration [C_max]) prediction accuracy was assessed using the absolute fold error (AFE, Equation (7)).

$$\mathrm{AFE}={10}^{\left|{\log }_{10}\left(\raisebox{1ex}{$\text{Observed}$}\!\left/ \!\raisebox{-1ex}{$\text{Predicted}$}\right.\right)\right|}$$ (7)

Predictions were considered accurate if the AFE is ≤ 2 as is commonly assumed in bottom‐up PBPK modeling, while those with AFE between 2 and 3‐fold and > 3‐fold were considered acceptable and failed predictions, respectively [2, 4, 7, 10]. Over‐ and underpredictions were distinguished using the ratio between the observed and predicted AUC_INF and C_max values (Equation (8)).

$$R=\frac{\text{Observed}}{\text{Predicted}}$$ (8)

The overall impact of each PBPK model parameter on the prediction of AUC_INF and C_max was assessed with the median AFE (MAFE; Equation (9)).

$$\text{MAFE}=X_{\mathrm{AFE}}\left[\frac{n+1}{2}\right]$$ (9)

Additionally, the predicted PK profile shape was assessed against the observations both visually and with the reduced Χ ² statistic (Equation (10)).

$${\chi }^2=\frac{1}{N}\sum _{i=1}^N\left(\frac{{\left(C_{\mathrm{Obs}}-C_{\text{Pred}}\right)}^2}{{\sigma }^2}\right)$$ (10)

where Χ ² is calculated from the number of observed clinical plasma samples (N), the observed (C _Obs) and predicted plasma concentration (C _Pred) and standard deviation (σ) at each time point. The first time point for each compound (0.25 or 0.5 h) was removed from the calculation since some of the measurements were lower than or close to the lower limit of quantification, and thus the estimate of σ was considered unreliable. A reduced Χ ² statistic closer to 1 indicates a better fit of the predicted PK curve to the observed data [10, 11].

3. Results

3.1. Overall Prediction Performance Using Different Combinations of PBPK Parameters

The prediction performance of AUC_INF and C_max was compared using different combinations of model parameters, that is, GI anatomy (Original vs. New), P‐gp REF (1.5 vs. 0.5) and CYP ISEF (adjusted vs. default) (Figure 1). The overall PBPK model performance was poor using Original GI physiology, P‐gp REF of 1.5 and adjusted ISEF with an average AUC_INF and C_max ratio of 22 and 13 across all compounds, respectively (Table 1). The prediction performance was significantly improved by switching to New GI physiology, P‐gp REF of 0.5 and the default ISEF values; the average AUC_INF and C_max ratio were 1.5 and 1.8, respectively (Table 1). Model performance was marginally improved when changing only one or two of the parameters. Using the Original GI, P‐gp REF of 1.5 and adjusted ISEF, the MAFE was 3.6 for AUC_INF and 7.9 for C_max (Table 1). With this combination of model parameters, only 29% and 43% of the predicted AUC_INF and C_max values were within 2‐fold of the observed values, respectively (Figure 2). In contrast, the combination of New GI, P‐gp REF of 0.5 and default ISEF significantly improved the prediction accuracy for both AUC_INF and C_max with MAFE of 1.5 and 1.7, respectively (Table 1). Consistently, 76% of AUC_INF and 71% of C_max values were predicted within 2‐fold of the observed values (Figure 2).

FIGURE 1.

Overall PBPK model performance. The ratio between the observed and predicted AUC_INF (AUC_INFR) (A) and C_max (C_maxR) (B) were calculated using eight different combinations of model parameters, that is, GI physiology (Original vs. New), P‐gp REF (1.5 vs. 0.5) and CYP ISEF (adjusted (Adj) vs. default (Def)). The data for multiple compounds (ABBV‐X1‐X8) at up to 5 dose levels (Table S2) were aggregated as the average and range for each combination. The dashed vertical lines and shaded areas represent the ratio of 1 and within 3‐fold predictions, respectively.

TABLE 1.

Overall evaluation of PK parameter estimates.

	GI Physiology	Original				New
	P‐gp REF	1.5		0.5		1.5		0.5
	ISEF	Adjusted	Default	Adjusted	Default	Adjusted	Default	Adjusted	Default
Average ratio (observed/predicted)	AUCINF	22	4.7	17	2.5	7.0	1.9	4.3	1.5
	Cmax	13	9.0	8.7	4.5	3.3	2.5	2.3	1.8
MAFE	AUCINF	3.6	3.4	3.3	2.3	2.4	1.6	1.9	1.5
	Cmax	7.9	6.3	5.2	4.4	1.6	1.8	2.1	1.7
Geomean Χ 2 statistic		8.15	4.17	8.90	4.31	7.47	3.01	7.38	2.65

Note: The prediction accuracy of AUC_INF and C_max was assessed with the average of the observed/predicted ratio and MAFE for all the compounds evaluated as described in Methods. The predicted and observed PK profiles were assessed with the reduced Χ ² statistic for all the compounds evaluated except ABBV‐X8.

FIGURE 2.

AUC_INF and C_max prediction accuracy. The prediction accuracy of AUC_INF (A) and C_max (B) for each compound and model was assessed with the AFE (Equation (7)) using eight different combinations of model parameters, that is, GI physiology (Original vs. New), P‐gp REF (1.5 vs. 0.5) and CYP ISEF (adjusted vs. default). The number of predictions with an AFE within 2‐fold (green), between 2‐ and 3‐fold (yellow) and outside 3‐fold (red) of the observed parameter were counted and aggregated as percentages in pie charts.

3.2. Impact of Individual PBPK Parameters on Prediction Performance

The number of AUC_INF predictions within 2‐fold of the observed values is higher when using New compared to Original GI physiology, P‐gp REF 0.5 compared to 1.5, and default compared to adjusted ISEF regardless of the other model parameters. In parallel, there was a decrease in the number of AUC_INF predictions varying by more than 3‐fold by individually changing these model parameters (Figure 2A). The improvement in C_max predictions was mainly driven by the New GI physiology with an additional improvement by changing P‐gp REF from 1.5 to 0.5 when New GI physiology was used in the model development (Figure 2B). The change in the CYP ISEF value (default vs. adjusted) had a modest impact on C_max predictions (Figure 2B). The impact of each model parameter on the prediction accuracy varied across compounds (Figure 3). The prediction performance was high in almost all conditions for ABBV‐X2, while it was not optimal for ABBV‐X3 in any conditions. The improvement in prediction performance was mainly driven by the New GI physiology for ABBV‐X1, ‐X4 and ‐X6. Lastly, the prediction performance of ABBV‐X5 was dependent on multiple parameters, where improved performance was observed only when all three parameters were optimized (Figure 3).

FIGURE 3.

AUC_INF and C_max prediction accuracy by compound. Heatmap of the prediction accuracy for AUC_INF (A) and C_max (B) across eight different combinations of model parameters, that is, GI physiology (Original vs. New), P‐gp REF (1.5 vs. 0.5) and CYP ISEF (adjusted (Adj) vs. default (Def)). Each row represents a different dose for individual compounds with the arrow showing the direction of increasing doses (low to high). Prediction accuracy is color coded by AFE (Equation (7)) (within 2‐fold (green), 2‐ to 3‐fold (yellow), 3‐ to 5‐fold (orange) and outside 5‐fold (red) of the observed parameter). ABBV‐X7 is not shown due to IP related matters.

3.3. Assessment of Predicted PK Profiles

The predicted and observed PK profiles were compared using the reduced Χ ² statistic (Table 2). The combinations of New GI physiology, P‐gp REF of 0.5 and default ISEF values for each enzyme provided the best Χ ² statistic for 5 out of 7 compounds evaluated (Table 2). However, the impact of each model parameter on the prediction of PK profile varied across compounds as will be exemplified with the following two cases.

TABLE 2.

Evaluation of PK profile prediction for individual compounds.

GI physiology	P‐gp REF	ISEF	ABBV‐X1	ABBV‐X2	ABBV‐X3	ABBV‐X4	ABBV‐X5	ABBV‐X6	ABBV‐X7	Geo‐mean
Original	1.5	Adjusted	6.17	1.63	10.8	68.5	3.02	9.59	63.3	8.15
		Default	5.00	0.59	10.2	42.4	2.46	11.6	26.6	4.17
	0.5	Adjusted	4.18	1.63	10.5	144	3.80	10.1	63.3	8.90
		Default	3.05	0.59	9.25	86.3	3.01	7.65	26.6	4.31
New	1.5	Adjusted	3.36	2.20	10.3	83.8	1.37	3.19	59.6	7.47
		Default	2.07	0.91	8.10	45.4	1.39	5.40	2.30	3.01
	0.5	Adjusted	2.22	2.20	10.0	103	1.16	5.40	59.6	7.38
		Default	1.23	0.91	6.68	54.4	1.39	2.04	2.30	2.65

Note: The predicted and observed PK profiles were assessed with the reduced Χ ² statistic. The Χ ² statistic closest to 1 for each drug is bolded. Profile fit for ABBV‐X8 could not be assessed quantitatively since only one participant was dosed.

3.3.1. Example 1: ABBV‐X1

ABBV‐X1 is a large monoprotic acid (> 500 g/mol) with low logP (< 5), moderate solubility (0.01–0.05 mg/mL) and high passive permeability (P _app > 5 × 10⁻⁶ cm/s; Table S1). ABBV‐X1 is a P‐gp substrate with low predicted clearance (hepatic extraction ratio, ER < 0.3) mainly driven by CYP3A4‐mediated metabolism followed by CYP1A2, 2B6, 2C8 and 3A5. ABBV‐X1 does not undergo active hepatic uptake, biliary or renal elimination (Table S1). ABBV‐X1 has a high f _u,p (> 0.1) and moderate predicted V _SS (0.6–2 L/kg). The prediction accuracy of AUC_INF, C_max and PK profile for ABBV‐X1 varied widely across the model settings (Figure S2). Absorption, as indicated by C_max, was underpredicted by > 5‐fold using the Original GI physiology and P‐gp REF of 1.5 regardless of ISEF values across the three dose groups (Figures 3B, 4A, and S2). Optimizing P‐gp REF to 0.5 alone resulted in a marginal improvement in C_max predictions (3 to 5‐fold underprediction when Original GI is used) (Figures 3B and S2). In contrast, C_max predictions were significantly improved (AFE ≤ 3) by the application of New GI physiology and further improved when combined with the optimized P‐gp REF of 0.5 and default ISEF (AFE ≤ 2; Figures 3B and 4). The default ISEF values significantly improved the AUC_INF predictions compared to the adjusted values across all doses (Figures 3A, 4C and S2). The highest prediction accuracy across all endpoints (C_maxR, AUC_INFR and Χ ² statistic) was achieved when combining the New GI physiology, P‐gp REF of 0.5 and default ISEF for CYP3A4 (AFE < 2; Χ ² statistic 1.23; Figure 4C, Table 2). Moreover, ABBV‐X1 showed nonlinear PK clinically where the increase in plasma exposures (C_max and AUC_INF) was less than dose proportional. The performance of the bottom‐up PBPK model was comparable between the three doses for this compound (Figure 3).

FIGURE 4.

Comparison of observed and predicted plasma concentration–time profile for ABBV‐X1. ABBV‐X1 plasma concentration–time profile comparisons of (A) original (red) and New GI physiology (yellow), using P‐gp REF 1.5 and adjusted ISEF, (B) P‐gp REF 1.5 (yellow) and 0.5 (blue), using New GI physiology and adjusted ISEF, and (C) adjusted (blue) and default ISEF (green), using New GI and P‐gp REF 0.5. Dots and solid lines represent the observed exposure (mean ± standard deviation) and geomean of simulated exposures, respectively. Comparison between the observed and predicted exposure–time profile for all models are available in Figure S2.

3.3.2. Example 2: ABBV‐X2

Like ABBV‐X1, ABBV‐X2 is a large monoprotic acid (> 500 g/mol), moderate solubility (0.01–0.05 mg/mL) and high passive permeability (P _app > 5 × 10⁻⁶ cm/s) compound with low predicted clearance (ER < 0.3) predominantly by CYP3A4‐mediated metabolism without any contribution of active hepatic uptake or renal elimination. On the other hand, it undergoes biliary excretion, has high logP (> 5) and is not a P‐gp substrate (Table S1). The C_max and AUC_INF of ABBV‐X2 are well predicted under all conditions across the five dose levels (AFE < 2) except the AUC_INF at the second low dose (AFE 2–3) when using New GI and adjusted ISEF (Figure 3). Unlike ABBV‐X1, the change in the GI physiology had a marginal impact on the predicted exposures (Figure 5A). The PK profile was best predicted using the New GI physiology and default ISEF value for CYP3A4 (AFE < 2; Χ ² statistic 0.91; Figure 5B, Table 2). P‐gp REF values had no impact on the prediction, consistent with ABBV‐X2 being not a substrate for this transporter.

FIGURE 5.

Comparison of observed and predicted plasma concentration–time profile for ABBV‐X2. ABBV‐X2 plasma concentration–time profile comparisons of (A) Original (red) and New GI Physiology (dotted blue), using adjusted ISEF, and (B) adjusted (blue) and default ISEF (green), using New GI physiology. Dots and dotted or solid lines represent the observed exposure (mean ± standard deviation) and geomean of simulated exposures, respectively. Other comparisons are summarized in Figure S3.

4. Discussion

Despite many publications discussing the performance of the PBPK model in human PK predictions, the majority of them do not consider the changes in software or improved understanding of system parameters [3, 5, 10, 11, 26]. One exception is a study that covers 116 clinical assets developed over 20 years [5]. An improvement in the prediction performance was observed since a fully mechanistic approach in PBPK model development was implemented in 2011. However, a direct comparison of prediction performance between different approaches was not reported on a compound‐by‐compound basis.

In the present study, the prediction performance of bottom‐up PBPK models using a combination of three different system parameters (GI physiology, P‐gp REF and CYP ISEF values) was evaluated through a retrospective analysis of AbbVie clinical assets. Overall, the prediction performance was significantly improved by using the New GI physiology, P‐gp REF of 0.5 and default ISEF for Gentest Supersomes compared with other combinations (Figure 1). With this combination, AUC_INF and C_max were within 2‐fold of the observed values for 76% and 71% of the predictions, respectively, with 81% and 86% within 3‐fold (Figure 2). However, the impact of each parameter on model performance was compound‐dependent (Figure 3).

To date, the performance of the New GI model compared to the Original model in the prediction of oral absorption has not been externally validated. Here, we show for the first time that the New GI physiology performs better than the Original model in predicting AUC_INF and C_max regardless of the other model parameters (Figures 1 and 2). It is of note that the improvement in C_max predictions was mainly driven by the New GI physiology between the three system parameters evaluated, reflecting the impact of GI physiology on the initial oral absorption process. The major difference between the Original and New GI models is the radius of the small intestine, where it is approximately 2‐fold shorter in the new model throughout the intestinal segments (0.92–2.1 cm) compared to the original model (1.6–3.2 cm). The shorter radius increases the absorption rate constant (K _a ) and accordingly the fraction absorbed (f _a ) in Simcyp Simulator where those are calculated with the equations below:

$$K_a=\frac{2\times P_{\mathrm{eff},\mathrm{man}}}{r}$$

$$f_a=1-{\left(1+\frac{K_a}{K_t}\right)}^{-7}$$

where P _eff,man is human intestinal passive permeability, r is small intestine radius and K _t is the small intestine transit time. This is consistent with the observation in our study, where there is generally an increase in the predicted C_max, resulting in an improvement in the C_max ratio (observed/predicted) across different compounds by switching to New GI physiology (Figure 1, Table 1).

Considering a wide range of intrinsic solubility from low (< 0.01 mg/mL) to high (> 0.05 mg/mL) within our data set, the New GI physiology may be suitable in model development for orally administered compounds with not only low but also high solubility. Additional investigations with a larger set of compounds are warranted to evaluate the general performance of the New GI model on oral absorption predictions.

The application of REF has been demonstrated to improve the IVIVE of transporter kinetics [27]. Quantitative proteomics is often used to determine the abundance of transporter proteins in the in vitro systems and in vivo tissues and estimate REFs [28]. However, such data is not readily available from all human tissues including the small intestine. Transporter activity also requires localization to the plasma membrane and thus may not always correlate with total protein expression [29]. Furthermore, physiological scalers (e.g., cellularity) have not been well established for small intestine and commonly used in vitro transporter models do not necessarily reflect intestinal physiology (e.g., lack of villi) and have high inter‐laboratory variability in transporter expression [30, 31, 32, 33]. As an alternative approach, REFs can be estimated by a retrospective comparison of observed human PK and in vitro transporter kinetics data obtained in the same cell lines across multiple compounds. A significant improvement in the absorption predictions for P‐gp substrates was observed by decreasing P‐gp REF to 0.5 from the default value for MDCK cells in Simcyp Simulator (1.5) (Figures 1B and 3B). This may infer approximately 3‐fold higher P‐gp activity in the MDCK cell line used in our study compared to the ones used to derive the default value. These observations emphasize the importance of determining REFs for individual in vitro systems either experimentally or empirically for a successful translation of transporter kinetics data.

ISEF is defined as the ratio of intrinsic CYP activity per unit of enzyme between human liver microsome and recombinant systems and incorporated in the IVIVE of rCYP clearance [34, 35]. During bottom‐up PBPK model development, the ISEF values can be adjusted to match the predicted clearance from rCYPs to those by other experimental systems (e.g., hepatocytes). Within our data set where the in vitro clearance was measured in Gentest Supersomes, AUC_INF was predicted more accurately using default ISEF values compared to adjusted values (Figures 1A, 2A, 3A). This may reflect an uncertainty in clearance prediction using human hepatocytes and accompanying IVIVE scalers obtained from preclinical species. Additionally, there was no correlation between the prediction performance using either default or adjusted ISEF values and the fraction metabolized by CYPs, ranging from 38% to 96% (Table S1). A broader and more systematic analysis is warranted to further explore the appropriate use of ISEF in the IVIVE of rCYP clearance.

While the combination of optimized system parameters significantly improved PK predictions in general, several models leave room for improvements. For example, C_max predictions of ABBV‐X3 deviated by more than 3‐fold even with the optimized approach (Figure 3B). For compounds with low solubility, such as ABBV‐X3, improved modeling of the impact of solubility and precipitation on oral absorption as informed by in vitro data could improve C_max predictions [17]. The selection of dosage form in the prediction may also affect the model performance. Solution formulation was used in the human PK prediction of ABBV‐X4 and ‐X5 assuming complete dissolution and no precipitation (Table S2). This may have affected the C_max prediction accuracy when using the Original GI physiology. Clinically, ABBV‐X4 and ‐X7 showed biphasic PK profiles, which were not captured accurately in the current PBPK models as indicated by relatively large Χ ² statistic values compared to other compounds (Table 2). This may be ascribed to the application of the Rodgers and Rowland method alone for V _ss prediction, which does not account for transporter‐mediated tissue distribution [36]. Permeability‐limited organ models using transporter kinetics in combination with other approaches, including the Dedrick or Wajima methods, may better predict the impact of transporters on tissue distribution and PK profiles [3, 9]. Finally, none of the PBPK parameter combinations could predict the observed AUC_INF of ABBV‐X8, where clinical PK data were available from only one subject. This may represent a limitation of the present study. Nevertheless, the general trend of the overall model performance is likely unchanged even if data from additional subjects were available for this compound.

Although the utility of PBPK modeling for PK assessments has been widespread, different strategies were adopted in the prospective prediction of human PK parameters, resulting in some limitations in the model application. For example, allometric scaling was employed to prospectively predict human clearance and V _ss instead of a fully mechanistic PBPK model [3]. Albeit a reasonably well prediction performance, where 65 to 80% of AUC and C_max predictions were within 2‐fold of observed values, such models cannot predict potential nonlinear PK due to a saturation of clearance pathways or be used for DDI predictions. Human PK was successfully predicted for 18 diverse compounds by a PBPK‐IVIVE approach, which involved the development of animal PBPK models for each compound to inform design and parameter estimation in the human models [10]. In contrast, only the contribution of extrahepatic elimination and V _ss is informed by preclinical in vivo data in our approach. While animal PBPK models can be informative, the streamlined process presented here is less data intensive and may accelerate decision making in early drug development without losing prediction accuracy.

In conclusion, the present study proposes an optimized strategy for bottom‐up PBPK model development in Simcyp Simulator using New GI physiology, P‐gp REF of 0.5 and default rCYP ISEF values. Caution is warranted for the P‐gp REF presented here as it is expected to vary between different in vitro systems. Retrospective comparison of bottom‐up PBPK predictions with observed clinical PK data is vital for model verification as well as further improvement in prospective predictions for future drug candidates.

Author Contributions

All authors wrote the manuscript. J.A.S.P., A.F.L., and R.K. designed the research. All authors performed the research. J.A.S.P. and A.F.L. analyzed the data.

Conflicts of Interest

The authors declared no competing interests for this work. J.A.S.P., A.F.L., E.A.C., Y.Q., L.R., C.M.L., M.B., R.K. are or were employees of AbbVie and may own AbbVie stock. AbbVie contributed to the design; participated in the collection, analysis, and interpretation of data, and in writing, reviewing, and approval of the final publication. J.A.S.P. and L.R. former AbbVie employees, now employees of Amgen and Merck, respectively, have no conflicts of interest.

Supporting information

Figure S1: Impact of f _u,gut on oral absorption prediction.

Figure S2: ABBV‐X1 model fit—all comparisons.

Figure S3: ABBV‐X2 model fit—all comparisons.

Data S1: Supplemental materials.

Table S1: Overview of compound properties in bottom‐Up PBPK model development.

Table S2: Clinical trial protocols.

Schulz Pauly J. A., Leblanc A. F., Chowdhury E. A., et al., “Optimization of Bottom‐Up PBPK Model Development in SIMCYP via Retrospective Analysis of Clinical Human PK Data,” Clinical and Translational Science 18, no. 11 (2025): e70417, 10.1111/cts.70417.