High-Throughput Physiologically Based Pharmacokinetic Model for Rodent Pharmacokinetics Prediction Using Machine Learning-Predicted Inputs and a Large In Vivo Pharmacokinetics Data Set

Abstract

Accurate prediction of the pharmacokinetic (PK) properties of small-molecule drug candidates is a critical aspect of pharmaceutical research. Fast and reliable PK predictions can accelerate compound optimization cycles, reduce animal testing, and enhance the quality of molecules advancing to human studies. Although physiologically based PK (PBPK) models are well-established for compound selection, their application in early discovery faces limitations due to low throughput and the requirement for substantial in vitro data. Recently, high-throughput PBPK (HT-PBPK) methods have become possible, offering scalable, parallel PBPK simulations that can be executed on thousands of compounds within minutes. Additionally, advancements in machine learning (ML) have enabled the substitution of in vitro data by high-quality in silico predictions that are based solely on chemical structures. In this study, the performance of a corporate HT-PBPK application, called SwiftPK, that leverages the HTPK simulation module included in a commercial software package was evaluated for predicting ten primary and secondary PK endpoints for a large (>9000 compounds) set of rodent PK data. Utilizing a corporate ML pipeline, all in vitro parameter inputs were replaced with in silico predictions. This approach is particularly relevant for early stage project phases, such as lead identification, as well as for external collaborations where experimental data are unavailable. The findings demonstrate the highly predictive performance of the HT-PBPK approach, with most endpoints predicted within a three- to four-fold error. Performance improves after filtering for compounds that are predicted, based on structure alone, to be cleared by hepatic metabolism (Extended Clearance Classification System class 2) and when using ML inputs that demonstrate high confidence. The results highlight the key prerequisites for successful application in early phase projects: predicted primary elimination pathway accuracy and prediction quality. This study is expected to inspire more organizations to incorporate HT-PBPK into their discovery pipelines, expediting the development of safe and effective novel medicines for patients.

Introduction

Physiologically based pharmacokinetic (PBPK) modeling and simulation is a well-established technique designed to predict the time course of drug concentrations in tissues and blood following drug administration. These models simulate the processes of absorption, distribution, metabolism, and excretion (ADME), providing a detailed understanding of how a drug moves through and interacts with the body. PBPK models are distinguished from more empirical models by their ability to incorporate mathematical representations of physiological and anatomical features, enabling a mechanistic approach to predict and simulate pharmacokinetics. ,

PBPK models are applied during every stage of drug discovery and development as they are key components of model-informed drug development strategiesguiding drug design during lead identification (LI) and optimization (LO), supporting the simulation and prediction of human pharmacokinetics (PK) and dosing before first-in-human studies, and providing more sophisticated simulations during clinical development. High-throughput PBPK modeling (HT-PBPK), where simulations are conducted for a large number of molecules based on very limited data, is an increasingly common approach used during the early phases of drug discovery. A key strength of PBPK models lies in their ability to predict in vivo PK based on in vitro input data, often referred to as “bottom-up” modeling. HT-PBPK can reduce the number of in vivo studies needed during the discovery phase; therefore, it is aligned with the ethical 3Rs (replace, reduce, refine) principle and can decrease costs and cycle times involved in drug discovery. By rejecting compounds when their simulated in vivo behavior is poor, HT-PBPK provides an opportunity to prioritize discovery series and compounds for in vivo testing.

There is interest in further optimizing the use of experimental resources including, with the growth of machine learning (ML), the replacement of in vitro inputs with values predicted from chemical structures. , This opens the prospect of performing PBPK simulations using virtual structures, thereby saving time and resources by only synthesizing the most promising molecules. An alternative approach that has demonstrated success in several case studies is the use of ML that is based on chemical structures to directly predict PK, bypassing the use of any PBPK model. ,, However, a significant limitation of this technique is the lack of mechanistic insights into in vitro-in vivo correlations (IVIVC) and the associated translatability of learning across species. Such pure ML predictions are exclusively tied to chemical structures, without considering measurable properties that might be adjusted to enhance PK profiles. Furthermore, ML approaches rely heavily upon their training data set and upon their training process, which is in contrast to the deterministic equations that govern PBPK models and traditional IVIVC. The present HT-PBPK framework, using a more mechanistic approach, is believed to provide continuity with established IVIVC approaches enabling translatability and furnishing valuable information for further PK optimization.

The potential impact of HT-PBPK on early discovery projects comes with important considerations regarding prediction confidence. In contrast to ML approaches where the confidence in a prediction can be estimated using various methods (e.g., ensemble modeling, bootstrapping, distance-based approaches), PBPK relies on the comparison of simulations using in vivo PK study data to build confidence in a specific chemical series. Ideally, a limited number of in vivo rodent studies should be conducted to validate HT-PBPK predictions for a series before prospective application. Discrepancies between predicted and observed PK can reveal gaps in the model, such as missing processes or inaccuracies in drug-specific properties, and highlight areas requiring refinement. Addressing these gaps typically involves running additional in vitro experiments. Such iterative refinement enhances the predictive performance of the model and deepens the understanding of the compound’s behavior. In the earliest stages of drug discovery, simulations often rely heavily on ML for predicting input parameters. In such cases, appropriate uncertainty management strategies must be implemented to ensure reliable interpretation of the simulation results.

In this study, the degree of uncertainty in HT-PBPK predictions for different chemical series was explored. One crucial process, known to be needed for accurate prediction of the PK profile, is plasma clearance (CL). The HT-PBPK approach relies on the scaling of intrinsic clearance (CL_int) values to predict the in vivo plasma clearance; the CL_int values are based on in vitro measurements in which compounds are incubated with liver microsomes or hepatocytes. Accordingly, the clearance of the investigated compounds should occur mainly via hepatic metabolism to be correctly predicted using HT-PBPK; the more the plasma CL is influenced by nonmetabolic or extra-hepatic processes, the higher the probability that the model will poorly predict plasma CL and other parameters dependent on CL, such as half-life and area under the curve (AUC). To recognize compounds that are likely to be primarily metabolized in the liver, the extended clearance classification system (ECCS) framework may be employed.

The ECCS system, first introduced by Varma et al. in 2015, serves as a useful framework for categorizing compounds based on their predominant CL pathways in humans. This system classifies small molecules into six distinct groups: 1A, 1B, 2, 3A, 3B, and 4. The classification involves three key parameters: ionization class, molecular weight, and passive permeability. The specific cutoffs used for assigning compounds to these classes are detailed in Supporting Table S1. Classes 1A and 2 are primarily associated with hepatic metabolism, whereas classes 1B and 3B are linked to transporter-mediated hepatic uptake. Ultimately, classes 3A and 4 are characterized by renal clearance as their main metabolic pathway.

As reported in the original publication, this model was originally validated using clinical data for 307 compounds, each having a single mechanism contributing to at least 70% of its systemic clearance. The ECCS framework successfully predicted the predominant clearance pathway in humans in 92% of cases.

Materials and Methods

Data Exchange and Preparation

The data set used in this study consisted of rodent in vivo data for diverse compounds generated in a separate part of the Roche group. Two parts of the Roche group (pRED and gRED) work separately on discovery and early development projects, and the data set used in this study was shared by gRED. This provided an unprecedented opportunity to evaluate a large, well-curated, and chemically diverse test set of structures not previously explored using internal ML models. Importantly, the chemical space covered by the gRED data set is different from the internal (pRED) one; Napoli et al. recently published a paper highlighting this difference for the extensive data sets. The PK profiles for the compounds investigated in the present study were from single-dose PK studies conducted via intravenous (IV) and oral (PO) administrations. For this evaluation to reflect typical small molecule early discovery pharmacokinetics, data from PK experiments with doses exceeding 100 mg/kg and compounds with a molecular weight over 1500 g/mol were excluded. AUC values, extrapolated to infinity (AUC_inf), were dose-normalized and, in cases where a single compound was measured multiple times for the same PK endpoint, the median values were used. The final aggregated data set consisted of 9098 diverse compounds. The PK endpoints that were associated with the molecules, together with the number of entries associated with each specific property, are listed in Table .

1. Overview of the Compounds Considered in this Study for each Rodent PK Endpoint Analyzed,

PK property (route of administration)	species	units	no. of compounds
AUCinf (IV)	mouse	h·ng/mL/mg/kg	3760
	rat		6236
CL (IV)	mouse	mL/min/kg	3773
	rat		6218
V ss (IV)	mouse	L/kg	3772
	rat		6230
T half (IV)	mouse	h	3753
	rat		6233
F (PO)	mouse	%	2490
	rat		4303

^aAUC_inf = area under the curve extrapolated to infinity; CL = clearance; F = bioavailability; IV = intravenous; PK = pharmacokinetic; PO = oral; T _half = half-life; V_ss = volume of distribution at steady state.

^bThe CL and AUC_inf values were dose normalized.

SwiftPK Architecture

The SwiftPK architecture was built on commercial software platforms, with Spotfire Analyst (Version 10.3.2; Cloud Software Group, Fort Lauderdale, FL) serving as the primary user interface. The Spotfire template is connected to an internal database (managed via D360, Version 23.0.24; Certara, Radnor, PA) that allows retrieval of input data via application programming interface (API) calls; additional functionality was implemented to allow users to load external input data sets. Within Spotfire, users can configure modeling parameters for the simulations; these configurations are sent via an API to the ADMET Predictor software (SimulationsPlus, Lancaster, CA) to execute high-throughput PBPK simulations. Similarly, simulation results are returned from ADMET Predictor and can be analyzed directly in Spotfire using a set of preconfigured, customizable plots. A schematic representation of the SwiftPK architecture is provided in Figure .

1.

Graphical representation of the SwiftPK architecture. ADMET = absorption, distribution, metabolism, excretion, and toxicity.

A key feature of the platform is its support for fallback parameter values. For example, if a preferred (experimental) measurement is unavailable for a compound, the system can automatically substitute it with a value from an alternative assay or use an ML-based prediction. SwiftPK is integrated with an internal ML pipeline for property predictions, allowing automated retrieval of predicted values from the same internal database (see next paragraph). For the purpose of this study, only ML-predicted values were used as inputs.

Machine Learning Property Prediction (MLPP) Pipeline

The internal machine learning pipeline (MLPP) consists of a multitask graph neural network model based on the Deep Graph Library in Python. The model, an Attentive Fingerprint (AttentiveFP) network, was trained using the Roche in vitro absorption, distribution, metabolism, and excretion (ADME) data set with an early stopping criterion to prevent overfitting. ,

The MLPP automatically retrains the model regularly, incorporating newly measured data. Following each training session, the model undergoes a 5-fold scaffold split cross-validation. This type of cross validation offers a more realistic performance assessment by predicting compounds with novel scaffolds compared to random cross-validation, which tends to overestimate model performance.

Before each retraining, the model predicts outcomes for newly measured molecules (creating a so-called “time-split cross-validation” data set), providing insights into the model’s performance over time.

The MLPP also incorporates a robust uncertainty quantification method, based on a combination of ensemble models and mean variance estimation (MVE). This allows the predictions to be coupled with predicted errors that are then binned on a scale from 1 to 10 to produce “predicted goodness” (Pred Gdns) values. These are more understandable for the user and translate the model confidence into a simple intuitive scale that ranges from 1 (maximum prediction uncertainty) to 10 (minimum prediction uncertainty).

ECCS Framework Application

Suspecting that better SwiftPK predictivity might be observed for compounds primarily metabolized by the liver when using in vitro intrinsic hepatic clearance to estimate in vivo clearance, molecules were grouped using the ECCS. Classes 1A and 2, where the clearance should be mainly hepatic metabolism, were assumed to yield better predictions than other classes.

To do this, the ECCS class for each compound was predicted using its molecular weight, ionization category (based on the predicted acid dissociation constant [pK _a], using ADMET Predictor, Version 11.0), and predicted passive cellular permeability in LLC-PK1 (porcine kidney epithelial cell line LLC-PK1-Mdr1a overexpressing mouse P-glycoprotein [P-gp]) cells in the presence of the P-gp inhibitor zosuquidar using the MLPP model. Although the permeability measure is different from the value used in the original publication (using Madin-Darby canine kidney cells), the permeability estimate is considered to also capture the true passive permeability, excluding efflux. , An overview of the ECCS-based classification of the investigated compounds is shown in Figure .

2.

ECCS classes of compounds investigated in this study. ECCS = extended clearance classification system. Note: Supporting Table 1 explains the characteristics and thresholds used to categorize compounds into each class.

SwiftPK Simulations

This study aimed to represent the real-world application of SwiftPK using a project in late LI or early LO phase, where simulations are run at a single preliminary dose level for all compounds. For the IV route of administration, the chosen dose was 1 mg/kg and for the PO route of administration, 10 mg/kg was selected; these dose levels are typically used for early PK studies in rodents. As the platform was based on the HTPK module of the ADMET Predictor software (Version 11.0), the only available dosage forms were IV bolus and immediate-release (IR) tablet. The IR tablet model was deemed to be appropriate for the suspension and solution PO doses administered to rodents. An additional solution dosage form has been added to more recent versions of the HTPK module, but it was not evaluated in the current study.

The HTPK module leverages the Advanced Compartmental Absorption and Transit (ACAT) model, − which is part of the GastroPlus (SimulationsPlus, Lancaster, CA) PBPK platform, to simulate dissolution, permeation, and transit within the gastrointestinal tract; it also considers liver metabolism to estimate the first-pass effect and oral bioavailability. To enhance simulation efficiency, the general circulation and other organs and tissues are consolidated into a single “central” compartment. However, tissue composition variations are accounted for when predicting the volume of this central compartment, referred to as the mechanistic volume of distribution (V).

Consistent with an effort to mimic the early adoption of SwiftPK, the other parameters (outlined below) for the simulations remained as the default values. Specifically, in the present study, the SwiftPK simulation results presented were obtained for the described rodent PK data set using the following conditions:

The in vivo clearance was calculated using the well-stirred liver model starting with the MLPP-predicted in vitro intrinsic clearance in hepatocytes (CL_int,hep) and accounting for the fractions unbound in plasma and in vitro. The dilution scaling method was used to estimate in vitro binding. , Our internal assay for CL_int,hep uses hepatocyte suspension cultures and has a lower limit of quantification (LLOQ) of 1 μL/min/10⁶ cells. Among the compounds included in this study, only one had an ML-predicted CL_int,hep in rat below the LLOQ, and none were predicted below the LLOQ for mouse CL_int,hep.

The in vivo volume of distribution at steady-state (V _ss) was determined using the Lukacova extension of the Rodgers approach. ,

The regional in vivo human intestinal effective permeability (P _eff) was calculated using the MLPP-predicted in vitro apparent permeability (P _app); for the present data set, this represents the apical-to-basolateral (AB) permeability in LLC-PK1-Mdr1a cells (porcine kidney epithelial cell line overexpressing mouse P-gp). Our internal MLPP model for P _app prediction is trained on our proprietary cellular permeability data set and demonstrates robust performance across the full dynamic range of the assay, providing more reliable predictions at higher P_app values, with confidence becoming increasingly moderate at lower values. The P _eff value is derived using a correlation equation that we developed in-house based on Roche proprietary data.

The plasma protein binding, CL_int,hep, log D (at pH 7.4), aqueous kinetic solubility (derived from the lyophilization solubility assay assay, at pH 6.5), the apical efflux ratio (AP-ER), and the passive cellular permeability (the aforementioned AB permeability in LLC-PK1-Mdr1a cells in the presence of a P-gp inhibitor) were calculated using the MLPP pipeline.

The pK _a values, blood-to-plasma ratios, fasted state simulated intestinal fluid (FaSSIF) and fed state simulated intestinal fluid (FeSSIF) solubilities, and diffusion coefficients were calculated using ADMET Predictor (Version 11).

The aqueous solubilities at different pHs were predicted based on the MLPP-predicted aqueous solubility at pH 6.5 and the predicted pK _a values. Our in-house MLPP solubility model, built on our proprietary data set, is affected by the imbalanced distribution of experimental solubility values: it provides more confident estimates at the low and high ends of the assay range, whereas predictions for intermediate solubility values exhibit moderately higher prediction uncertainty.

The in vivo solubility was handled by estimating a bile salt solubilization ratio for enhanced solubility in the gastrointestinal tract. This was done using the predicted aqueous kinetic solubility and predicted solubility in biorelevant FaSSIF and FeSSIF media, as done in the full GastroPlus ACAT model.

Several GastroPlus default parameters were used in all simulations; the particle density was kept at 1.2 g/mL, particle radius at 10 μM, and body weights were kept at 25 g for mice and 250 g for rats.

Using the aforementioned conditions and inputs, SwiftPK simulations were performed to get single-dose PK parameters for all 9098 compounds in the data set.

Evaluation Metrics

Several performance metrics were used to evaluate prediction bias, precision (i.e., dispersion), and correlations between observations and predictions. Prediction bias was captured using the average fold error (AFE), where AFE values of <1 and >1 indicate bias underprediction and overprediction, respectively.

$$\mathrm{A}\mathrm{F}\mathrm{E}={10}^{[\sum {\log }_{10}\left(\frac{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}{\mathrm{o}\mathrm{b}\mathrm{s}\mathrm{e}\mathrm{r}\mathrm{v}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}\right)/\mathrm{n}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r}\mathrm{o}\mathrm{f}\mathrm{o}\mathrm{b}\mathrm{s}\mathrm{e}\mathrm{r}\mathrm{v}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{s}]}$$

Prediction accuracy was assessed using the absolute average fold error (AAFE), also known as geometric mean fold error. The AAFE is the geometric mean of individual absolute fold errors, where a value of 1 indicates a perfect prediction.

$$\mathrm{AAFE}={10}^{[\sum \mathrm{ABS}{(\log )}_{10}\left(\frac{\mathrm{prediction}}{\mathrm{observation}}\right)/\mathrm{number}\mathrm{of}\mathrm{observations}]}$$

The correlations between observations and predictions were described by the Pearson’s correlation coefficient (r), which quantifies the strength and direction of the linear association between two continuous variables, ranging from −1 (perfect negative linear relationship), through 0 (no linear relationship), to +1 (perfect positive linear relationship). To make the correlation clearer, the experimental and predicted values were transformed to the log₁₀ scale for use in the Pearson’s correlation calculation. As bioavailability was the only PK property that ranged between 0% and 100%, these values were transformed to the logit scale instead of the log₁₀ scale.

$$r=\frac{\sum ({\mathrm{observation}}_i-\stackrel{\circledR }{\mathrm{observation}})({\mathrm{prediction}}_i-\stackrel{\circledR }{\mathrm{prediction}})}{\sqrt{\sum {({\mathrm{observation}}_i-\stackrel{\circledR }{\mathrm{observation}})}^2\sqrt{\sum {({\mathrm{prediction}}_i-\stackrel{\circledR }{\mathrm{prediction}})}^2}}}$$

Results

SwiftPK Performance

Analysis Using All Compounds with MLPP-Predicted Input Parameters

Figure illustrates the Pearson’s correlations, AFEs, and AAFEs as measures of the overall performance of SwiftPK using the rodent PK data set. The majority of the investigated PK parameters were predicted with errors between 3- and 4-fold, as shown by the AAFE. The AFE indicates a trend toward overprediction for IV AUC_inf values and PO bioavailability predictions in both species, whereas the IV V _ss and CL values tended to be underpredicted. The Pearson r statistic shows correlations of 0.13–0.41 when comparing predicted and observed PK parameters, although bioavailability values showed lower correlations in both species.

3.

SwiftPK performance using 10 rodent PK endpoints for the investigated compounds. AFE = average fold error; AAFE = absolute average fold error; AUC = area under the curve; CL = clearance; DN = dose-normalized; F = bioavailability; IV = intravenous; PK = pharmacokinetic; PO = oral; Thalf = half-life; V _ss = volume of distribution at steady state. Notes: SwiftPK performance was examined based on the correlation (measured using log₁₀ Pearson coefficients).

To investigate if a compound’s ECCS class influenced SwiftPK performance, the data set was categorized by ECCS class and the same analysis was performed (Figure ).

4.

SwiftPK performance using 10 rodent PK endpoints, according to ECCS class. AFE = average fold error; AAFE = absolute average fold error; AUC = area under the curve; CL = clearance; DN = dose-normalized; ECCS = extended clearance classification system; F = bioavailability; IV = intravenous; PK = pharmacokinetic; PO = oral; Thalf = half-life; V _ss = volume of distribution at steady state. Notes: The plots show the performance in terms of correlation (measured using the log₁₀ Pearson r), accuracy (quantified using AAFE), and bias (indicated using the AFE). The ECCS classes are described in Supporting Table 1.

Analysis Using MLPP Pred Gdns Values of 6 and Above

This analysis was conducted using the MLPP predicted properties as inputs, irrespective of the confidence reported by the models for those predictions. To investigate the impact of the quality of the MLPP-predicted properties on the performance of SwiftPK, the analysis was performed again using only compounds with in vitro inputs that were predicted with a reasonable level of confidence, arbitrarily defined as a Pred Gdns value of 6 or higher (as stated previously, Pred Gdns values range from 1 to 10; higher values represent the higher levels of confidence that the model assigns to the prediction) for all ML-predicted input properties. The parameters for which the Pred Gdns values were considered were plasma protein binding, CL_int,hep, logD (at pH 7.4), aqueous kinetic solubility, AP-ER, and passive cellular permeability. The Pred Gdns filtration criteria were met by 301 compounds in mice and 296 in rats, covering a total of 406 diverse molecules. Figure reports the results for mouse and rat predictions and shows that the SwiftPK performance with these compounds is significantly better than average, with all the mouse endpoints predicted to have less than a 2.73-fold error, and only one rat endpoint that was slightly above a 3-fold error.

5.

SwiftPK performance obtained using 10 rodent PK endpoints for compounds with Pred Gdns scores ≥ 6. AAFE = absolute average fold error; AFE = average fold error; AP-ER = apical efflux ratio; AUC = area under the curve; CL = clearance; DN = dose-normalized; F = bioavailability; IV = intravenous; PK = pharmacokinetic; PO = oral; Pred Gdns = predicted goodness; Thalf = half-life; V _ss = volume of distribution at steady state. Notes: This study considering only molecules for which all machine learning inputs (mouse plasma protein binding, mouse CL_int,hep, log D at pH 7.4, aqueous kinetic solubility, AP-ER, and passive cellular permeability) were associated with a goodness score of ≥6. The plots show the performance for correlation (measured using the log₁₀ Pearson r), accuracy (quantified using the AAFE), and bias (indicated using the AFE).

A further examination of SwiftPK performance for compounds with confident property predictions based on their ECCS classification was not meaningful. For example, for a subset of 301 compounds in mice, 296 belonged to ECCS class 2, with only 1 each from classes 4, 1A, and 3; no compounds were classified as being from classes 3A or 3B. Similarly, for the rat subset, 295 out of 296 molecules were classified as ECCS class 2.

Discussion

This study investigated the performance of SwiftPK using a large database of rodent PK endpoints. The overall predictive power of the process, across the entire compound database, was assessed first. Subsequently, a granular analysis was conducted, stratifying the results based on ECCS categories, to evaluate SwiftPK’s performance across distinct elimination mechanisms. Finally, the influence of ML input confidence on the quality of PK predictions was investigated, elucidating how the reliability of in silico input parameters impacted the overall success of this HT-PBPK approach.

Performance Analysis

As shown in Figure , SwiftPK predicted rodent PK parameters for the full compound database with AAFEs ranging from 2.90 to 4.15. The ECCS was developed based on human data and has been successfully applied to strategies for human clearance predictions. It has also been reported to be a valuable tool for assisting with nonclinical PK clearance scaling approaches. , We acknowledge that a validation of the cutoff values for MW and permeability has not been performed specifically in rodents. However, we have shown that our internal assay performs well when using the permeability cutoff of 50 nm/s for human and experience leads us to believe that use of the same cutoff values is reasonable when applied for rodents.

Therefore, an analysis of predictions based on compound ECCS classifications was performed, as shown in Figure .

ECCS class 2, which includes neutral or basic highly permeable compounds with hepatic metabolic clearance as the main elimination pathway, is the most highly represented category in the present data set. As expected, the HT-PBPK approach performed best for this class of compounds, showing improved predictions across all analyzed PK endpoints. The AAFE values indicated an error between 2- and 3-fold for all the endpoints related in the mouse, apart from the oral bioavailability (AAFE = 3.45). For the rat IV AUC_inf, CL, and V _ss, the AAFE values indicated errors between 3- and 3.5-fold. The AFE values indicated a strong tendency to underpredict IV CL and V _ss for both mouse and rat; the IV AUC_inf and oral bioavailability were overpredicted. This correlates with reported findings suggesting that the hepatic metabolic clearance of compounds tends to be underpredicted based on hepatocyte CL_int.

The second-best performing ECCS class was 1A, for which liver metabolism is still considered the principal elimination mechanism, but where the protein binding and distribution characteristics of these small acids or zwitterions may lead to under prediction of clearance by in vitro systems. For this class, several PK endpoints were predicted within a 3-fold error for the mouse; a notable exception was the IV Thalf that had an AAFE of 5.83. For the rat, however, the IV AUC_inf, CL, and Thalf values were predicted with an error significantly above 4-fold. Interestingly, the log₁₀ Pearson correlation for the IV V _ss was above 0.5 for both the mouse and rat, indicating that the underprediction indicated by the AFE may reflect a consistent bias.

For molecules in class 1B, transporter-mediated hepatic uptake is expected to play a major role in clearance. As depicted in Figure , an overall prediction profile was observed that was similar to class 1A, but with tendentially higher AAFE values. In this class, apart from the oral bioavailability in both species and the mouse IV V _ss, none of the PK endpoints were predicted with less than a 4-fold error. This suggests that these compounds tend to have their PK more strongly influenced by transporter-mediated processes than by passive diffusion.

Nonhepatic metabolic clearance pathways and transporter-driven PK are also expected to play significant roles for compounds from classes 3A, 3B, and 4. This was reflected by the poor SwiftPK predictive performance across these classes.

For the few compounds assigned to ECCS class 3A, a general tendency was observed to predict mouse PK endpoints within a 3.5-fold error. The IV Thalf was an exception, demonstrating a 4.62-fold error. The rat data presented a less favorable picture, with the IV AUC_inf and CL predictions having greater than 12-fold error values; the IV V _ss and Thalf predictions also had errors above 5-fold. Despite this, the log₁₀ Pearson r values remained reasonable for the oral bioavailability and IV V _ss predictions, suggesting some rank-ordering consistency even when the absolute values are prone to having substantial errors.

For Class-3B drugs, the model showed poor predictive accuracy, with almost no properties falling within a 4.5-fold error; the only exceptions were the IV AUC_inf and CL predictions in the mouse. The log₁₀ Pearson correlations were generally low across the parameters, with the exception of rat oral bioavailability (r = 0.68). However, this value likely reflects a systematic overprediction, as indicated by the AFE, similar to the trend previously observed for the IV V _ss predictions for Class 3A compounds.

Class 4 compounds, which were the second most represented class in the data set, displayed the worst overall performance, in terms of AAFE. Rat IV AUC_inf and CL values were predicted with errors exceeding 7-fold; oral bioavailability predictions were also highly inaccurate. These results align with the expectation that Class 4 compounds are primarily eliminated via renal clearance, a mechanism not well captured by SwiftPK, which assumes a prevailing liver-driven metabolism. The AFE values support this interpretation, showing moderate to massive overprediction for most parameters, except for IV CL in both species, which is strongly underpredicted, and rat IV V _ss, which is moderately underpredicted.

This analysis shows the relevance of considering the ECCS class as a check on the confidence associated with SwiftPK predictions in a pure in silico mode where all input parameters are derived from ML. The use of the predicted ECCS class can be an effective first check on the likely accuracy of SwiftPK when used in this way.

The next step of the study was consideration of the influence of the estimated confidence in ML-predicted inputs on overall PK prediction success. The Pred Gdns scores could range from 1 (lowest confidence) to 10 (highest confidence); in this study, high overall confidence for a compound was arbitrarily defined to be when the Pred Gdns values were 6 or higher for every ML-predicted input (plasma protein binding, CL_int,hep, logD at pH 7.4, aqueous kinetic solubility, AP-ER, and passive cellular permeability).

Overall, we found that compounds associated with ML inputs of higher confidence showed better performance than those with low confidence, with all mouse PK endpoints predicted well below the 3-fold error threshold and eventually below the 2.5-fold error threshold. Despite this, the AFE values still indicated overprediction of IV AUC_inf and oral bioavailability values, along with underprediction of IV CL and V _ss values. This is not surprising since, even when using in vitro-measured inputs, underprediction of CL is often reported. Pearson’s correlation demonstrated good results for IV CL, Thalf, and V _ss, but revealed a low correlation for oral bioavailability.

The Pearson’s correlations were tendentially low across all endpoints for the rat PK endpoints, although they were better than those for the unfiltered overall data set for most endpoints. The AAFE values remained below the 3-fold error threshold, except for IV V _ss (3.06). The bias scenario, depicted by the AFE, was similar to that observed in the mouse subset.

These findings suggest that filtering compounds based on their ML-predicted confidence for PBPK inputs can significantly enhance the performance of SwiftPK. Comparing this scenario with the study conducted by Naga et al., where SwiftPK was applied to a data set of compounds with numerous experimentally measured in vitro inputs, the performance obtained in the current setup was modestly inferior. This difference was expected, as the experimental values provide more stable and solid starting points for HT-PBPK simulations and are thus recommended for LO and late discovery phases. Nonetheless, this study demonstrates that SwiftPK performance remains useful when the input data are predicted entirely in silico, provided that the prediction tools used are validated and perform well. Consequently, SwiftPK can be reliably utilized during the LI phase of drug discovery for compound ranking and early rodent PK predictions. It is noted that, as mentioned previously, the pRED data used to train the MLPP ML models and the gRED data for which the simulations were performed cover different regions of the chemical space. Therefore, we expect a rather better than lower performance when applying the pipeline in internal pRED projects. Moreover, greater reliability in its application is expected for ECCS class 2 compounds and molecules with inputs that are predicted with high confidence by the ML models.

Industrial Impact of These Findings on LI/LO Discovery Phases

This study demonstrated the feasibility and value of using ML-predicted inputs in an HT-PBPK model for early in vivo PK predictions. The SwiftPK application is believed to be able to support the prioritization of lead series during the LI phase and has the potential to reduce the number of in vivo PK or PK/pharmacodynamic (PD) studies needed. This can accelerate timelines in the LO phase of a small molecule discovery project. In an industrial setting, this translates into earlier differentiation between series that are likely to meet project PK/PD requirements and those that are unlikely to do so, enabling project teams to concentrate synthetic and in vivo resources on a smaller subset of candidates.

SwiftPK, used in the LI phase, should be able to facilitate early comparisons of compound series based on their simulated PK and PD profiles against target profiles. If ML models can reliably predict input parameters and these predictions are validated with in vitro data for a subset of compounds, the SwiftPK inputs can be predicted for numerous virtual analogs. This enables the simulation of in vivo profiles prior to synthesis. Thereafter, compound series with favorable simulated PK can be prioritized for synthesis and subsequent in vitro and in vivo validation. The selection of the lead series can be informed by the likelihood that simulated or observed PK/PD profiles meet the required criteria. Compounds or series for which SwiftPK consistently predicts PK/PD profiles that fall outside the predefined acceptance windows, even under optimistic assumptions, can be deprioritized or stopped from further exploration before resource-intensive optimization. The level of confidence in a stop decision would therefore be linked to the demonstrated predictive performance of the underlying predictive models and would be periodically re-evaluated as more prospective data accumulate.

During the LO phase, the SwiftPK simulation engine can help reduce the number of in vivo PK and PK/PD studies by focusing experimental efforts on compounds with the highest predicted potential. Following the experimental validation of initial PK/PD hypotheses using early lead compounds, virtual analogs can be designed and their PK/PD profiles simulated using the SwiftPK platform. These simulation results can then guide compound ranking and prioritization for subsequent in vivo testing. By limiting in vivo studies to top-ranked candidates, this approach will enable more design cycles within a given time frame and support the identification of clinical candidates with improved properties, ultimately accelerating the progression of a novel drug to the clinic.

Taken together, we envisage that in LI the SwiftPK platform will mainly support early ranking and, in clear-cut cases, early stop decisions when predicted profiles are incompatible with project objectives. In LO, its main impact will be on reducing in vivo experimentation by prospectively shaping study design and focusing resources on those compounds with the highest modeled probability of success. Quantitative decision thresholds (e.g., acceptable deviation of predicted from target exposure or dose, or minimal predicted probability of meeting key PK/PD criteria) would be defined at the project level.

Relevance of the Findings for Human PK and Dose Predictions

This study investigated the use of an HT-PBPK simulation engine, employing only structure-based in silico property predictions as inputs. The evaluation was restricted to the prediction of PK in mice and rats. Although a key interest in drug discovery is the optimization of PK in humans, this assessment remains valuable for several reasons. First, there is the availability of a very large and structurally diverse set of compounds with in vivo rodent PK data, which contrasts with the much more limited human PK data sets available (e.g., a recent evaluation of clinical PK predictions was limited to 12 compounds). Second, assessing the success of in vitro-in vivo extrapolation and PBPK in rodents facilitates the establishment of a level of confidence in human PK predictions. Here, the aim was not to directly translate rodent PK to human PK, but to assess the confidence in PBPK methods using in vivo data before any clinical PK data are available. Third, the use of PBPK predictions of rodent PK can be very valuable for reducing the number of costly and time-consuming in vivo PK studies. Furthermore, these PBPK simulations, if based on in silico predictions, can reduce the need for in vitro studies and can even be applied to structures not yet synthesized.

The results of this evaluation confirmed several important assumptions underlying the use of HT-PBPK methods in drug discovery projects at the LI/LO stage. First, the methods for assessing the goodness of predicted input properties were demonstrated to translate to improved PK predictions. Therefore, if any key inputs for a compound are not predicted with a high level of goodness, caution is needed, and the generation of in vitro data to improve the PK prediction should be considered. Second, and importantly, the use of the ECCS was found to help guide the level of confidence in HT-PBPK. Although the ECCS was developed based on human PK data, it appears to be applicable for use in rodents, as well.

Besides predicting PK, SwiftPK also allows predictions of optimal doses. Key applications of SwiftPK include (i) estimation of doses to be used in in vivo animal efficacy models, (ii) estimation of doses which might trigger off-target safety concerns, (iii) simulation of efficacious doses for humans. In addition to the inputs for PK prediction, the database linked to SwiftPK also contains key PD measures relevant for animal species and humans. For example, IC₅₀ values obtained in cellular assays for the rodent and human target or off-target binding parameters (e.g., hERG IC₂₀) values are available for most projects and can be combined with PK predictions to result in dose estimates (e.g., to match steady state average, maximum or trough concentrations).

Current Limitations and Future Improvements

The high-throughput PBPK model framework used in this work simplifies the typical multicompartment PBPK structure to gain speed and scalability. Although this reduction sacrifices some details compared with full PBPK models, previous publications have shown that it can still achieve accurate predictions of plasma concentrations when supplied with reliable inputs. , Rather, it is the quality of the input parameters which is critical for the performance of HT-PBPK approaches. We have demonstrated that the ECCS classifications play a critical role in determining the predictive accuracy. Indeed, the HTPK model simplifies human physiology to a scenario in which the liver is considered the main metabolizing organ, and for this reason the ECCS 2 compounds are expected to be better predicted. In addition, some major gaps in the models are highlighted by the poorer predictions for ECCS 3 and ECCS 4 compounds. For such molecules the role of active processes is important in determining pharmacokinetics. Processes such as active hepatic uptake and active renal or biliary secretion are not covered in the model. This is not a weakness in the model structure but reflects a lack of input data to quantify these processes. Since bottom-up prediction of active transport from in vitro data remains a challenge for PBPK modeling in general, improved understanding in this area is necessary before we can expect to include such factors in HT-PBPK models used in early discovery.

Two approaches could further improve the impact and predictivity of this study. First, better predictivity of the HT-PBPK model is expected when considering the applicability domain of the ML model and only including PK properties of “within domain” compounds. Robust algorithms, such as conformal prediction, could be applied to properly define the applicability domain of the ML models and to enable more precise prediction of the PK properties of compounds that lie within the defined domain. Such an approach would allow project teams to quantify, for each compound, whether a prediction can be considered “reliable enough” for decision making (e.g., ranking, triage or dose selection) or whether experimental data should be generated instead.

For compounds falling outside the applicability domain, Alvarsson et al. suggested that synthesis and experimental measurements could be pursued based on their importance to the discovery project. An expanded applicability domain can be expected for a retrained predictive ML model by incorporating these “out-of-domain” measurements. Iteratively combining SwiftPK simulations with targeted experimental measurements in this way would progressively tighten prediction intervals and increase confidence in the use of model outputs for project-critical decisions. With more precise properties as inputs, the error margins for in vivo PK predictions using an HT-PBPK model are expected to be reduced. In addition, the statistical metrics will be more representative if they incorporate in vivo PK study variabilities. Daublain et al. suggested that 2–5-fold variabilities in PK parameters can be expected for in vivo studies. These study variabilities are likely to influence the statistical metrics calculation in this study. However, incorporating study variabilities is unlikely to affect the general findings and conclusions. Another limitation is that the results presented apply to a specific chemical space and project history. Consequently, the current work should be viewed as a detailed case study that establishes a framework and performance benchmark, and future efforts will focus on extending this evaluation across additional discovery projects and chemical spaces to better define when and how SwiftPK-type approaches can be most effectively deployed.

Conclusions

This study demonstrated the feasibility and value of using an integrated HT-PBPK modeling tool (SwiftPK) to generate inputs based on only ML predictions to provide early in vivo PK projections. This approach facilitates lead series prioritization during the LI phase and streamlines in vivo PK and PK/PD testing during LO phase, accelerating drug discovery timelines. Although the current assessment focused on rodent PK, it established confidence in the SwiftPK predictions that can, then, inform human PK extrapolations. Further improvements, such as refining the ML model applicability domains and incorporating in vivo study variabilities, could enhance prediction accuracy and reliability. Despite these limitations, the study supports the integration of HT-PBPK methods into early discovery workflows, ultimately aiding in the identification of clinical candidates with optimized properties.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.molpharmaceut.5c01317.

• Outlines the thresholds used for ECCS classification of the compounds (Table S1) (PDF)

The authors declare the following competing financial interest(s): All authors are employees of F. Hoffmann-La Roche Ltd. The authors have no other relevant affiliations or financial involvement with any organization or entity having a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript.