Leveraging In Silico and Artificial Intelligence Models to Advance Drug Disposition and Response Predictions Across the Lifespan

ABSTRACT

Incorporating inter‐individual differences in drug disposition and responses is essential for ensuring the safe and effective use of drugs in real‐world patients. Despite ongoing efforts, lower participation of children, older individuals, pregnant and breastfeeding women, postmenopausal women, and people with disease states and disabilities in drug clinical trials is frequent, and it requires multifaceted strategies and tools to evaluate drug exposure and responses in broad populations. The availability of modeling and simulation tools, such as physiologically based pharmacokinetic (PBPK) and quantitative systems pharmacology/toxicology (QSP/QST) modeling, enables the application of virtual populations that reflect the differences in drug disposition and responses for disease states and different stages of the lifespan. These models integrate clinical trial and real‐world data (RWD) to predict drug exposure, efficacy, and safety. Additionally, machine learning (ML) and artificial intelligence (AI) offer powerful tools for analyzing large datasets and identifying key physiological determinants of drug response across the lifespan. This review discusses the application of in silico and AI models to advance the prediction of drug exposure and responses across the lifespan, including examples of virtual populations in PBPK and QSP/QST models. A case study on QST modeling for drug‐induced liver injury (DILI) in postmenopausal women is presented, along with opportunities and challenges in applying AI for modeling physiological determinants of drug dosing in individuals ranging in age from 12 to > 80 years old in drug development.

1. Introduction

Evaluation of inter‐individual differences in drug disposition and responses in clinical trials is crucial to ensuring the safe and effective use of drugs in real‐world patients. Despite ongoing efforts, children, the elderly, pregnant or breastfeeding women, and people with disabilities are less likely to participate in clinical trials [1, 2, 3]. The lack of clinical data in broad populations compromises the generalizability of clinical research and increases the risk of treatment failure. Novel strategies and tools are needed to address and mitigate these challenges.

Advances in in silico tools and developments in artificial intelligence (AI) can help bridge the gap and translate existing clinical datasets to the real‐world setting. These tools can leverage existing datasets from clinical trials and real‐world data (RWD) to further inform and enhance drug development and utilization. In clinical pharmacology, modeling and simulation strategies are routinely applied to optimize clinical trial design and safe and effective regimens for therapeutics. For example, by applying physiologically and mechanistically based models, virtual populations can be constructed based on known physiological changes, clinical trial data, and RWD.

Virtual populations—simulated cohorts that reflect physiological and pathophysiological characteristics, including variability within specific patient subgroups—serve as critical inputs into physiologically based pharmacokinetic (PBPK) and other mechanistic models, ultimately informing drug optimization and regulatory decision‐making. Virtual populations can be used to predict drug exposure, efficacy, or safety in broader populations and then can be further validated and refined as more data become available. Machine learning (ML) and AI provide powerful tools to integrate and analyze large datasets and identify key physiological determinants of drug exposure and outcomes.

This review will discuss how clinical pharmacology considerations related to the variability of drug disposition and responses across the lifespan and disease states can be addressed with in silico tools (Figure 1). The scope of published virtual populations in physiologically based pharmacokinetic (PBPK) and quantitative systems pharmacology/toxicology (QSP/QST) models will be summarized to identify opportunities and gaps. A case study involving the application of QST modeling to evaluate drug‐induced liver injury (DILI) in postmenopausal women will be described. In addition, opportunities and challenges of the application of AI in drug development will be discussed with a case example, where generative adversarial networks (GAN) were leveraged to identify physiological determinants of drug dosing.

FIGURE 1.

In silico tools and clinical data, including real‐world data, can be used to enhance clinical pharmacology and drug development across lifespan. PBPK/QSP/QST models integrate preclinical and clinical data to predict drug exposure, efficacy, and safety. Clinical data from initial trials in healthy populations can be employed to validate and/or refine PBPK/QSP/QST models (A). Within validated models, virtual populations can be developed based on available population‐specific physiology data to prospectively predict drug exposure, efficacy, and/or safety in patient populations, which can be used to optimize the dosing regimen in planned trials in broader populations (B). Once executed, data from clinical trials can be employed to further validate and/or refine virtual populations (C). Next, validated models and virtual populations can be further expanded to represent patients with characteristics not included in trials to optimize dosing regimens in real‐world patients (d), predictions of which can be validated or further refined leveraging real‐world data (E). If clinical trials in certain disease states or age groups are not feasible, virtual populations initially validated using clinical data from available data can be expanded to represent intended patient groups to inform dosing regiments in real‐world patients (F). Machine learning (ML) and artificial intelligence (AI) offer powerful tools for analyzing large datasets and identifying key physiological determinants of drug exposure and responses, which can be used to predict exposure and drug outcomes in specific patients of interest.

2. Clinical Pharmacology Considerations

Optimal drug dosing requires consideration of factors such as genetics, age, and sex that can influence drug disposition and response. For instance, genetic polymorphisms can affect drug‐metabolizing enzymes, for example, cytochrome P450 2D6 (CYP2D6), leading to drug disposition and toxicity variations among different populations. Additionally, age and sex can significantly alter pharmacokinetics (PK) and pharmacodynamics (PD). Age‐dependent decreases in hepatic and renal function can impair drug metabolism and excretion and prolong drug half‐life in older adults [4]. Sex‐related differences in PK may occur because the differences in body composition between men and women can alter drug distribution. Impairment of renal function and hepatic metabolism can alter drug clearance [5]. Extrinsic factors, such as diet and environmental exposures, can significantly influence drug disposition and response [6, 7].

Modeling and simulation methods, such as population PK/PD, PBPK, and QSP/QST models, are frequently employed to analyze drug disposition data and predict drug effects [8]. These techniques evaluate intrinsic and extrinsic factors influencing drug behavior across different groups and create virtual populations that reflect the variability of drug disposition in real‐world settings. In silico modeling and simulation have emerged as powerful clinical pharmacology tools for understanding inter‐individual drug exposure and response variability.

Clinical pharmacology data are essential for integrating PK, PD, and patient‐specific factors to assess inter‐individual variability in drug metabolism, response, and safety. RWD collected from electronic health records, patient registries, and patient‐reported outcomes is representative of the variations in comorbidities, concomitant medications, treatment adherence, etc., that occur in everyday clinical practice. RWD has become an invaluable complementary resource for assessing inter‐individual drug exposure and response variability [9, 10, 11].

3. PBPK Modeling

PBPK modeling uses differential equation systems to predict how drugs are absorbed, distributed, metabolized, and excreted in humans and animals. PBPK models combine drug and physiology information to estimate the concentration of a drug in plasma and tissues over time. As such, PBPK models can simulate the impact of intrinsic (e.g., genetic, disease, organ impairment) and extrinsic (e.g., formulation, food effects, drug–drug interactions) factors on drug PK profiles. This makes PBPK models well suited to extrapolate PK data from healthy adults to broader populations, which can inform clinical trials (e.g., the first‐in‐human dose) or reduce unnecessary clinical trials in disease populations. Table 1 summarizes virtual populations for different clinical settings.

TABLE 1.

Selected virtual populations representing special populations (other than healthy adults) in human PBPK modeling.

Category	Population	Study	Age	Gender	Race/ethnicity	Drug
Pediatrics	Term neonates, infants, children, and adolescents	Johnson et al. (2006) [12]	Birth to 18 years	M/F	Not reported	Midazolam, caffeine, carbamazepine, cisapride, diclofenac, S‐warfarin, omeprazole, theophylline, phenytoin, gentamicin, and vancomycin
Pediatrics	Premature and term neonates, infants, children, and adolescents	Edginton et al. (2006) [13]	Birth to 18 years	M/F	Not reported	Paracetamol, alfentanil, morphine, theophylline, and levofloxacin
Pediatrics	Preterm neonates	Claassen et al. (2015) [14]	24 to 40 weeks gestational age and up to 30 days postnatal age	M/F	Not reported	Amikacin and paracetamol
Geriatrics	Elderly individuals	Schlender et al. (2016) [15]	65 to 100 years	M/F	Not reported	Morphine and furosemide
	Elderly individuals	Cui et al. (2021) [16]	65–75 years and > 75 years	M/F	Chinese	Simvastatin, midazolam, theophylline, ceftazidime, gentamicin, and vancomycin
Pregnancy	Pregnant women	Dallmann et al. (2017) [17]	Different stages of pregnancy	F	Not reported	Cefazolin, cefuroxime, and cefradine
	Pregnant women and fetus	Abduljalil et al. (2012) [18]	Different stages of pregnant	F	Predominantly Caucasian	Not applicable
	Pregnant women and fetus	Szeto et al. (2021) [19]	Different stages of pregnancy	F	Not reported	Cefuroxime, cefazolin
Organ impairment	Renal impairment (RI—mild, moderate, severe) a	Heimbach et al. (2021) [20]	18–85	M/F	Not reported	25 blinded compounds
	RI (mild, moderate, severe)	Hsueh et al. (2018) [21]	Average age: 25–63	M/F	Not reported	Adefovir, avibactam, entecavir, famotidine, ganciclovir, oseltamivir carboxylate, sitagliptin
	RI (severe)	Tan et al. (2019) [22]	Average age: 48–54	M/F	Not reported	Rosiglitazone, pioglitazone, pitavastatin, repaglinide
	RI (mild, moderate, severe)	Yee et al. (2018) [23]	26–84	M/F	Not reported	4 blinded compounds, Oseltamivir carboxylate, cidofovir, cefuroxime
	RI (mild, severe)	Khalid et al. (2023) [24]	3.5–20	Not reported	Not reported	Captopril
	Hepatic impairment (HI—mild, moderate, severe) b	Heimbach et al. (2021) [20]	23–76	M/F	Not reported	27 blinded compounds
	HI (mild, moderate)	Li et al. (2015) [25]	Not reported	Not reported	Not reported	Bosentan, olmesartan, repaglinide, and valsartan
Disease	Obesity c (adults)	Berton et al. (2023) [26]	20–50	M/F	Not reported	Midazolam, triazolam, caffeine, chlorzoxazone, acetaminophen, lorazepam, propranolol, amikacin, tobramycin, and glimepiride
	Obesity (adults)	Berton et al. (2023) [27]	20–50	M/F	Not reported	Drug–drug interaction between dolutegravir and rifampicin
	Obesity (children)	Gerhart et al. (2022) [28]	2–20	M/F	Asian‐American, Black‐American, Mexican‐American, and White‐American children	Clindamycin, trimethoprim/sulfamethoxazole
	Obesity (children)	Gerhart et al. (2022) [29]	2–18	M/F	White, Black, Asian Americans	Enoxaparin
	Obesity (children)	Ford et al. (2022) [30]	10–18	M/F	White, Black, Asian Americans	Metformin
	NAFLD/NASH (simple steatosis and non‐cirrhotic NASH)	Adiwidjaja et al. (2024) [31]	Mean age 29–56	M/F	Not reported	Chlorzoxazone, caffeine, midazolam, pioglitazone, rosuvastatin, 11C‐metformin, morphine and the glucuronide metabolite of morphine
	Non‐cirrhotic NASH)	Sjöstedt et al. (2021) [32]	33–63	M/F	Not reported	Morphine and morphine‐3‐glucuronide
	NAFLD	Newman et al. (2022) [33]	20–65	M/F	Not reported	Caffeine, clozapine, omeprazole, metoprolol, dextromethorphan, midazolam, s‐warfarin, rosiglitazone
	Cancer (all cancer types)	Cheeti et al. (2013) [34]	23–92	M/F	Not reported	Saquinavir, midazolam
	Cancer (pregnant and non‐pregnant patients)	Yang et al. (2023) [35]	18–45	F	Not reported	Paclitaxel, docetaxel, acalabrutinib
	Cancer (all cancer types)	Salgado et al. (2022) [36]	Not reported	Not reported	Not reported	89Zr‐fresolimumab, 89Zr‐bevacizumab, 89Zrbevacizumab, 89Zr‐bevacizumab, 64Cu‐DOTA‐trastuzumab, 89Zr‐trastuzumab, 89Zr‐MMOT0530A, 89Zr‐MMOT0530A, 89Zratezolizumab, 89Zr‐atezolizumab, 89Zr‐atezolizumab

^a Mild renal impairment (RI, GFR 60–90 mL/min per 1.73 m²), moderate RI (GFR 30–60 mL/min per 1.73 m²), severe RI (GFR 15–30 mL/min per 1.73 m²).

^b Mild hepatic impairment (HI,Child‐Pugh score 5,6), moderate HI (Child‐Pugh score 7–9), severe HI (Child‐Pugh score 10–15).

^c Obesity in adults (BMI > 30 kg/m²), obesity in pediatrics (BMI > 95th percentile).

3.1. Pediatrics

PBPK modeling provides a robust framework for regulatory submissions and supports pediatric drug development [37]. Pediatric virtual populations developed for PBPK modeling account for the dynamic physiological changes from birth through adolescence, including age‐specific anatomical and physiological parameters such as organ size, tissue composition, blood flow rates, and enzymatic activity. Key assumptions in these models include the scaling of organ weights and blood flow rates and the maturation profiles of drug‐metabolizing enzymes and transporters [12, 38, 39].

PBPK modeling software platforms have developed pediatric virtual populations that simulate drug absorption, distribution, metabolism, and excretion in children of different ages. Johnston et al. developed a PBPK model incorporating age‐dependent changes from birth to 18 years in cardiac output, liver blood flow, renal function, plasma protein concentrations, and drug‐metabolizing enzymes and evaluating its clearance predictions for 11 drugs (Table 1) [12]. The work by Edginton et al. described the development and evaluation of a virtual population that accounts for age‐dependent changes (birth to 18 years) in cardiac output, organ blood flows, tissue composition, renal function, and the ontogeny of drug‐metabolizing enzymes and evaluated clearance prediction for six drugs [13, 38]. These studies demonstrated the utility of PBPK modeling for simulating PK parameters across pediatric age groups. Despite significant advances, knowledge gaps related to physiological processes affecting pediatric drug absorption, maturation in drug transporters, developmental changes in the blood–brain barrier, and the impact of age on physiological processes altering the PK of therapeutic proteins remain [40].

3.2. Geriatrics

Geriatric virtual populations can incorporate age‐related alterations such as decreased renal and hepatic function, changes in body composition, and altered protein binding [41]. Schlender et al. developed a PBPK model for adults 65 to 100 years old that accounted for age‐dependent changes in muscle and fat mass and total body water, reduced kidney weight, blood flow and function, liver weight, and decreased cardiac output [41]. The performance of the virtual population from the model was evaluated using morphine and furosemide PK data [41].

3.3. Maternal and Fetal Exposure During Pregnancy

Virtual populations for pregnant women simulate the alterations in maternal physiology, such as increased cardiac output, expanded blood volume, and enhanced renal and hepatic clearance. Key assumptions include the trimester‐specific changes in physiology, for example, increased renal blood flow and glomerular filtration rate (GFR), as well as altered activity of drug‐metabolizing enzymes [42]. Hormonal factors regulate these changes, which vary with gestational age [42].

3.4. Organ Impairment

Renal and/or hepatic impairments impact drug exposure because the liver and kidneys are the major organs of drug metabolism and elimination. Table 1 summarizes selected publications on PBPK modeling in patients with renal or hepatic impairments.

3.5. Renal Impairment

Renal impairment (RI) is characterized by reduced GFR and/or tubular secretion. The severity of renal impairment is classified based on the estimated GFR: mild (60–90 mL/min per 1.73 m²), moderate (30–60 mL/min per 1.73 m²), severe (15–30 mL/min per 1.73 m²). Published virtual renal impairment populations covered all severity levels (i.e., mild, moderate, and severe); however, patients with end‐stage renal disease were not explicitly represented due to potential confounding factors such as dialysis and kidney transplants. Most virtual populations with renal impairment represented reduced GFR, increased serum creatinine, reduced hematocrit, and reduced gastric emptying rate based on physiological data obtained from patients with respective disease severity [20, 43, 44]. The representation of the altered expression of enzymes/transporters and the impact of uremic toxic substances (e.g., renal transporter inhibition) was case‐by‐case, depending on the primary elimination route of the drug [44]. Most published virtual populations of renal impairment represent adults. Khalid et al. simulated captopril exposure in pediatric patients with mild and severe renal impairment [21].

3.6. Hepatic Impairment

Hepatic impairment (HI) is caused by chronic diseases that lead to hepatic and biliary cirrhosis. Liver cirrhosis is characterized by reduced hepatic blood flow, hematocrit, alpha‐1‐acid glycoprotein (AAG or orosomucoid), albumin, functional hepatocytes, and biliary excretion. Also, the expression and activity of hepatic/intestinal enzymes and transporters may be altered in cirrhotic patients. Virtual populations representing mild, modest, and severe hepatic impairment on the Child‐Pugh (CP) classification scale [22] have been reported. Hepatic impairment PBPK models implement a progressive decrease in hepatic blood flow, hematocrit, AAG, and albumin, as well as altered expression of specific CYP enzymes and transporters depending on the drug's routes of elimination. The reported virtual populations of hepatic impairment are for adults, and there are no published examples focused on children with hepatic impairment.

3.7. Disease States

It is critical to evaluate drug exposures in patients with the indication or disease of interest. PBPK models for several disease states, including obesity, nonalcoholic fatty liver disease (NAFLD, also known as metabolic dysfunction‐associated steatotic liver disease, MASLD [24]), nonalcoholic steatohepatitis (NASH, or metabolic dysfunction‐associated steatohepatitis, MASH [24]), cancer, and chronic heart failure, have been reported.

3.8. Obesity

The physiological changes in obese adults (BMI ≥ 30 kg/m²), such as increased cardiac output and blood flow, increased organ weight, increased GFR and AAG, and altered expression of metabolizing enzymes, were comprehensively characterized and used for PBPK modeling [45]. PBPK modeling reasonably reproduced clinical PK data for 10 exemplar compounds in obese patients [46] and the clinically observed DDI between dolutegravir and rifampicin [26]. A virtual population based on physiological data from children with obesity adequately predicted clinical PK of clindamycin, trimethoprim/sulfamethoxazole, enoxaparin, and metformin and informed potential dose adjustments in children with obesity [27, 28, 29].

3.9. Liver Diseases

NAFLD encompasses a spectrum of chronic liver diseases ranging from simple steatosis or nonalcoholic fatty liver to NASH, the more advanced stage with hepatic inflammation [30]. Earlier publications of NAFLD virtual populations focused on specific physiological changes in NAFLD patients, such as altered abundance of CYP enzymes [47] or transporter function [48]. More recently, a comprehensive, quantitative analysis of NAFLD‐associated changes in physiological characteristics and the abundance of various metabolizing enzymes and transporters involved in drug disposition has been published [33]. Based on this information, virtual populations for different stages of non‐cirrhotic NAFLD (i.e., NAFL and NASH) were developed, accounting for pathophysiologic changes, and proteomics‐informed alterations in the abundance of metabolizing enzymes and transporters pertinent to drug disposition [32]. These virtual populations were verified using exemplar drugs whose elimination is mediated by CYP enzymes or transporters.

3.10. Cancer

Cancer patients frequently differ from healthy volunteers on physiologic factors (e.g., age, height, body weight, plasma protein concentrations), and various anti‐cancer drugs have lower clearance in cancer patients compared to healthy subjects. Examples of PBPK modeling in cancer patients from the published literature are summarized in Table 1. Yang et al. [49] developed virtual populations for pregnant and non‐pregnant cancer patients, focusing on changes in AAG, CYP3A4, and CYP2C8 expression, and renal function. Pregnancy physiology was used to represent pregnant cancer patients. Of note, some of the physiological changes in cancer and pregnancy are in opposite directions, for example, GFR is decreased in cancer but increased in pregnancy. This model was validated using paclitaxel and docetaxel PK data in pregnant cancer patients. The cancer population is very heterogeneous, and other factors, such as cancer type and tumor size, can affect PK. Salgado et al. reported a semi‐PBPK model for the distribution of therapeutic antibodies from the primary tumor to the tumor‐draining lymph node using human immuno‐positron emission tomography (PET) antibody imaging data [31].

4. QSP/QST Modeling

QSP models represent the biological and pathophysiological underpinnings of specific diseases to predict the therapeutic effects of drugs. QST modeling is an approach to quantitatively understanding the toxic effects of drugs and chemicals on living organisms by integrating computational and experimental methods. Because QSP/QST models account for physiology and pathophysiology, underlying biochemistry, and drug interactions with biological processes, virtual populations can be generated to predict drug efficacy and safety in disease states by understanding the variability due to the represented mechanisms.

QSP models have been published for several diseases, including immuno‐oncology, metabolic, nervous system, and cardiovascular diseases [35]. In QSP models‐derived virtual populations, variability in key biological/pathophysiological processes and PD responses simulates the variability in efficacy in target patients.

Only a handful of QSP models have included the effects of patient demographic characteristics, such as age and sex, on variability in key biological processes and efficacy responses. A review of QSP models in regulatory submissions included a case study where virtual populations representing adult and pediatric patients with inborn errors of metabolism were developed to compare efficacy in these patient groups and inform age‐specific dose regimens [36]. A sex‐specific virtual population was generated with a QSP model of blood pressure regulation that accounted for the sex difference in hypertension pathophysiology [50]. These examples demonstrate that virtual populations representing specific patient groups can be used to predict drug efficacy and optimize treatment regimens. A major challenge is obtaining age, sex, and other relevant biological and physiological data to inform and validate virtual populations.

QST models have been applied to predict drug‐induced toxicity in various organs such as the heart, liver, and bone marrow/hematology [51]. QST model‐derived virtual populations can be used to predict potential drug toxicity in early‐ and late‐stage clinical trials. Extrapolation from healthy volunteers to the target disease is essential to address drug safety, but only a few published reports have addressed this aspect of QST modeling. This section presents a case study of QST modeling of DILI in postmenopausal women.

4.1. Case Study: QST Modeling of Drug‐Induced Liver Injury in Postmenopausal Women

Postmenopausal women are about one‐eighth of the total population worldwide but are underrepresented in clinical trials and virtual populations, which are often skewed toward young and male subjects [52].

To predict the potential hepatotoxic effects of drugs in postmenopausal women, a simulated population (SimPops) was developed in DILIsym. This QST platform models drug exposure, liver biochemistry, and key DILI mechanisms, including oxidative stress, mitochondrial dysfunction, and bile acid transporter inhibition (See Table 2).

TABLE 2.

List of parameters modified between the main NHV SimPops and the post‐menopausal women SimPops and the magnitude of the applicable modifications.

SimPops parameter	BMI < 35	BMI ≥ 35
Liver RNS/ROS clearance scale (Vmax)	Reduce by 10%	Reduce by 10%
Bulk bile acid (i.e., CA) synthesis rate	Reduce by 40%	Reduce by 40%
CDCA baseline synthesis rate	Reduce by 40%	Reduce by 40%
Basal value of mitochondrial ETC flux	Reduce by 12%	Reduce by 20%

4.2. BMI And Body Weight

The differences in BMI, body weight, and NAFLD incidence between postmenopausal and premenopausal women are well characterized [53, 54]. Simulated individuals from the DILIsym NAFLD SimPops and the normal healthy volunteer (NHV) SimPops were combined to create the postmenopausal women SimPops; the body mass and BMI distributions of the SimPops were matched to the literature values (Figure 2).

FIGURE 2.

Distribution of BMIs in the final postmenopausal women SimPops; the final population includes 124 healthy and 120 NAFLD individuals.

4.3. Oxidative Stress

Glutathione levels and the expression of liver antioxidant enzymes, such as superoxide dismutase, catalase, and glutathione peroxidase, are lower in postmenopausal women compared to their premenopausal counterparts [55, 56]. Antioxidant enzyme activity is implemented as a single lumped parameter for reactive oxygen species/reactive nitrogen species (ROS/RNS) in DILIsym and is 10% lower for the postmenopausal group.

4.4. Mitochondrial Toxicity

Mitochondrial function also declines with age in many tissues. However, literature data on electron transport chain activity in the human liver is unavailable. Preclinical studies in ovariectomized mice show decreased gene expression for mitochondrial function and biogenesis in the liver, suggesting the importance of estrogen receptor signaling [57]. These empirical findings were incorporated in the SimPops by setting a lower value for mitochondrial electron transport chain (ETC) function in postmenopausal women. An additional reduction of mitochondrial electron transport chain function was implemented for the NAFLD group, which has reduced mitochondrial function compared to the healthy population [58].

4.5. Bile Acid Transporter Inhibition

Bile acid synthesis decreases with age, but there is no evidence for sex differences [59]. Transporter expression differences in premenopausal and postmenopausal women have not been characterized. Decreased bile acid synthesis was implemented in the postmenopausal women SimPops. As there is no evidence that the overall plasma bile acid profile or liver bile acid concentrations are different between pre‐ and postmenopausal women, the transporter expression levels in the population were optimized to result in the same bile acid profile with lower synthesis levels of cholic acid (CA) and chenodeoxycholic acid (CDCA).

4.6. Simulation Results

The postmenopausal women SimPops were used to predict the potential for liver toxicity with acetaminophen (APAP) and tamoxifen compared to an NHV SimPops. As expected, predicted DILI in the postmenopausal women population administered repeat therapeutic dosing of APAP (1 g QID for 2 weeks) tracked the predicted DILI in the NHV population; transient ALT increases greater than three‐fold above the upper limit of normal (ULN) were simulated in only a few individuals (~1%) of both populations. Tamoxifen was simulated in the NHV and postmenopausal women SimPops at a dose of 20 mg BID for 5 weeks. 15.7% of the postmenopausal women SimPops administered tamoxifen were predicted to have an ALT value greater than or equal to three times the ULN. This was approximately twice the number of simulated individuals with increases in ALT compared to the placebo group in the postmenopausal women SimPops (i.e., 7.4%). In addition, the frequency of ALT elevation after tamoxifen administration was greater in the postmenopausal women SimPops compared to the NHV SimPops, consistent with clinical observations. Overall, these results suggest that in postmenopausal women, the SimPops reasonably recapitulated the increased risk of tamoxifen‐mediated ALT elevations compared to healthy subjects. As such, the virtual population approach can predict the DILI risk of new compounds in postmenopausal women.

5. Machine Learning and Artificial Intelligence

Advances in ML and AI provide powerful tools for integrating and analyzing large datasets and identifying key physiological determinants of drug exposure and outcomes.

5.1. Definitions and Terms

AI emulates human intelligence with software algorithms [60]. AI has found many applications in healthcare, ranging from reducing diagnostic errors in radiology and pathology to monitoring medical instrumentation sensors and guiding surgical robots. Machine learning (ML) is a subset of AI that synthesizes knowledge and enables the improvement of algorithms autonomously from data, and can be used to directly solve specific tasks such as feature identification, classification, and clustering [61, 62]. AI can be engineered for relatively complex tasks, for example, face recognition and clinical trial simulations [63]. AI and ML models have powerful capabilities for learning from data that can mitigate the limitations and challenges of building user‐driven models in pharmaceutical modeling and simulation. However, current AI/ML modeling methods typically require larger training and testing data sets compared to compartmental models used in pharmacometric analyses, PBPK, and QSP/QST.

Many important AI algorithms are built with artificial neural networks, which contain interconnected layers of nodes (often described as analogous to neurons) that process input data, which is typically high‐dimensional, via an intervening hidden layer to enable identification of output patterns. Deep learning models are engineered to contain artificial neural networks with multiple hidden layers that can simulate complex decision‐making processes. Deep learning enables semi‐supervised, self‐supervised, reinforcement, and transfer learning. Generative AI uses deep learning to create complex synthetic content, for example, text, data, images, video, or audio [60].

Development of AI‐guided products has surged, particularly in the radiology and cardiovascular specialties. As of May 2024, the FDA had approved over 850 devices with AI [64].

5.2. Example of AI/ML Use in Drug Development

Because the utilization of AI methods in drug development is nascent, there are few real‐world drug development examples. The interleukin‐1 inhibitor anakinra was approved by the FDA in 2022 for treating the coronavirus disease of 2019 (COVID‐19) based on AI/ML methods to identify the appropriate patient population for a clinical trial (https://www.fda.gov/media/163081/download). This represented the first time AI/ML methods were used for regulatory decision‐making.

COVID‐19 patients with soluble urokinase plasminogen activator receptor (suPAR) levels ≥ 6 ng/mL were reported to be at high risk for requiring mechanical ventilation and mortality [65, 66]. However, an approved suPAR assay was unavailable in the United States. Training and test data from two trials were used to develop an AI/ML model to identify individuals with suPAR levels ≥ 6 ng/mL from baseline characteristics [65, 66, 67]. Elastic net regression was used to identify key features, and an artificial neural network was used to determine a cutoff value to maximize sensitivity and a positive predictive value ≥ 0.95.

5.3. Generative AI

Generating synthetic data emerges as a practical solution in clinical trial simulations where data availability is limited, for example, groups with a low frequency of participation in clinical trials. Generated data can be used in simulations to assess the full range of factors that could contribute to drug efficacy, effects, and disposition.

Generative adversarial networks (GANs) [68] and variational autoencoders (VAEs) [69] are commonly used generative AI approaches. Both use systems of deep neural networks and can learn the high‐dimensional joint distribution from training data. However, GANs and VAEs use distinctly different computational strategies. GANs employ adversarial training via classification, whereas VAEs perform nonlinear dimensionality reduction. The AAE contains the encoder and decoder features of the VAE, but its neural networks are updated with an adversarial loss function from a discriminator. Figure 3 shows a schematic representation of GAN, VAE, and AAE architectures.

FIGURE 3.

(A–C) are schematic representations of the generative adversarial network (GAN), variational autoencoder (VAE), and adversarial autoencoder (AAE) methods, respectively. A GAN is comprised of generator and discriminator neural networks. The generator takes random variables from a latent space and conditioning variables as input and computes generated data via its neural network. The discriminator is a binary classifier that takes training data containing biomarkers and conditioning variables and the generated data from the generator as inputs to compute the generator and discriminator loss functions. The loss functions are used to update the generator and the discriminator neural networks via backpropagation. A VAE consists of an encoder neural network and a decoder neural network. The encoder conducts nonlinear dimensionality reduction to obtain a latent space vector representation of the distributional parameters ($\mu$, the mean vector, $\mathrm{\Sigma }$, the covariance matrix, and $\epsilon$, the noise vector). The encoder and decoder are updated via a variational loss function that takes training data and decoder output. The AAE contains the encoder and decoder features of the VAE, but their neural networks are updated with the aid of an adversarial loss function from a discriminator.

5.4. Case Study

5.4.1. GAN for Physiological Determinants of Drug Dosing

Statistical approaches for generating virtual populations, such as multivariate distributions, copulas, and Bayesian models require strong distributional assumptions and extensive user input that limit their utility to small numbers of variables and known distributions [60, 64, 68]. Approaches requiring less user input, for example, information‐theoretic and ML algorithms such as random forests, have also been investigated to identify and model the interdependencies among key pharmacometric covariates [69, 70]. Empirical approaches, such as resampling, do not require models and are easy to implement because they reuse data vectors in an existing dataset by random or stratified random sampling. However, they do not generate new cases that are not present in the original data. AI approaches capable of reliably generating random variate vectors from the joint distribution can provide virtual populations for the drug and disease‐relevant biomarkers for clinical trial simulations.

Generative AI methods fit this bill to a tee as they can estimate the high‐dimensional joint distribution of a PDODD biomarker panel containing disease status variables, providing a comprehensive statistical description of the trends, pairwise correlations, multivariate associations, and inter‐individual variability among the constituent variables, including the disease status dependence of PDODD [70]. The main advantage of generative AI is its ability to learn complex data distributions from training data, which frees the user from formulating complex models. A representative and adequate training dataset is essential.

Most datasets in pharmaceutical sciences settings contain tabular data. Tabular data generation presents unique challenges compared to image generation because tabular data can contain a mix of discrete and continuous variables. Additionally, the order of the rows and columns in tabular data does not have local structure. The distributions of continuous variables on conditioning variables, for example, sex, genetic variants, etc., that are of clinical interest can be multimodal. The underlying density function can have multiple peaks. The typical GAN architectures designed for images are not particularly good at generating multimodal data because of a phenomenon termed “mode collapse”. Mode collapse reduces the diversity of output samples, and it occurs when the generator can only produce a single output type or an incomplete set of output types that fool the discriminator [71]. Data vectors for underrepresented groups of interest are, by definition, low‐frequency variables that occur less frequently in training batches. This can affect the learning of the joint distribution of the low‐frequency group. Learning the joint distribution of the smaller class can be improved by oversampling.

Nair et al. [70] evaluated the utility of GAN models to generate data sets emulating the age and disease differences in physiological determinants of drug dosing (PDODD). The National Health and Nutritional Examination Survey (NHANES), which contains data from a nationally representative sample for anthropometric measures, laboratory measurements, physical screening, and survey responses, was used. The GAN was engineered to model the multivariate joint distribution of a PDODD panel containing 14 continuous variables and four categorical variables. The variables selected included sex, age, anthropometric measures, hepatic and renal function biomarkers, blood cell counts, and hepatitis disease state variables.

The conditional dependence of PDODD on hepatic and renal function was assessed. Hepatic function was evaluated with the R‐VALUE, a computed measure of liver function [72] obtained from serum alanine aminotransferase and serum alkaline phosphatase activity measurements in a standard complete metabolic panel. R‐VALUE can be used to discriminate between hepatocellular and cholestatic liver injury. In NHANES, active hepatitis B virus (HBV) and hepatitis C virus (HCV) infection status were defined from the outcomes of a two‐stage screening and diagnostic test process. GFR, a clinically essential measure of renal function, was obtained from serum creatinine using the CKI‐EPI study 2021 formula [73].

The univariate probability density histogram of GFR in the GAN‐simulated and the test datasets (Figure 4A) overlap across the entire range of GFR values. The age‐dependence of GFR from the GAN‐generated data was visualized as bivariate distributions using scatter plots (Figure 4B) with loess lines to highlight trends. The GAN simulated data points were well dispersed with the test data points, and the loess lines overlapped over the assessed age range. Thus, GAN simulations of GFR provide a good model for learning the age‐dependent decline and variability patterns of renal function. GFR was 12% lower in the group with active hepatitis C (Figure 4C), and GAN‐generated and test data histograms for the groups with and without hepatitis C overlapped extensively. Notably, GAN can simulate virtual populations with multiple comorbidities.

FIGURE 4.

(A–C) compare generative adversarial network‐generated data vs. test data for estimated glomerular filtration rate (GFR). The overall probability density histogram is shown in (A), and the age dependence is summarized in (B). The GFR distributions are compared in the groups with and without hepatitis C in (C). The variables GFR and RIAGEYR represent the min‐max transformed values of the logarithms of estimated glomerular filtration rate and age in years.

VAEs have also been evaluated for generating PDODD [74]. However, VAEs performed poorly at emulating categorical variables, especially those with a class frequency of less than 10% [74]. GANs and VAEs differ in how they handle conditioning categorical variables.

5.5. Large Language Models

Large language models (LLMs) present transformative opportunities and challenges in clinical pharmacology [75, 76]. LLMs trained on enormous datasets have catalyzed groundbreaking advancements in the processing and generation of natural language, in the form of text and voice data. Some examples of LLMs are ChatGPT, Gemini, Llama, Claude, and DeepSeek [77, 78, 79, 80, 81]. The appeal of LLMs arises from their ability to emulate human‐like text generation and comprehension [82]. LLMs are versatile and user‐friendly and can assist with diverse tasks ranging from document editing to solving complex, problem‐oriented queries [83].

LLMs are deep learning AI algorithms built on transformers [84], a specialized architecture comprising an encoder and decoder with features for positional encoding, multi‐head self‐attention, and neural networks. Transformers enable LLMs to learn long‐range dependencies in the input data.

LLMs are a generative method and intrinsically lack the determinism and reproducibility of many computer algorithms. The same prompt can produce results that differ somewhat. “Temperature” is a built‐in LLM parameter that controls the balance between the precision of information retrieval and generative flair. A higher temperature setting is more likely to yield creative results, and a lower temperature is more likely to conform to the facts. However, lowering the temperature in the NONMEM code generation setting did not reduce coding errors [76]. The generative capabilities of LLMs might be helpful in drug discovery to identify lead compounds with novel structures, whereas reproducibility might be favored in a regulatory setting.

As LLM models require extensive training on enormous datasets, they can be updated only periodically. However, many research and clinical applications require a focus on a specialized knowledge base and up‐to‐date information. Retrieval‐augmented generation (RAG) is an AI framework that harnesses data from specialized user‐provided information sources, for example, clinically relevant or regulatory documents, to enhance LLM outputs. RAG‐based systems produce results that make them traceable, more trustworthy, and authoritative [85]. Technologies such as RAG may enable the development of secure healthcare applications. LangChain is an example of a framework that can supplement LLM with other resources to incorporate contemporaneous information without retraining [86].

LLMs can integrate diverse sources of information into easily understandable text. They can aid in collecting and assessing relevant patient data, planning and decision‐making in the face of uncertainty, and integrating information to supplement the pharmaceutical scientist's judgment. LLMs have been investigated for solving PK problems, writing R and NONMEM code, and curating RWD [75, 76, 87].

LLMs are also well‐suited for enhancing pharmacists' tasks in clinical and community pharmacies [88, 89]. ChatGPT performs well at interpreting prescriptions and providing decision support for patient consultations in community pharmacy [90].

5.6. Algorithmic Bias

Algorithmic bias refers to prediction errors that create unfair outcomes for certain groups of individuals. It is emerging as an essential consideration in AI because the datasets used for training are often insufficiently representative of real‐world scenarios. Algorithmic bias is a concern for virtual populations because they are used to evaluate inter‐individual differences. Algorithmic bias is generally not assessed in population pharmacometrics, PBPK, and QSP/QST.

Algorithmic bias evaluations of early commercial face recognition algorithms showed significant disparities: the maximum error rate for darker‐skinned females was 34.7% versus 0.8% for lighter‐skinned males [91]. Publicly available image data sets of skin disease lesions have provided a framework for testing the disparities in the accuracy of diagnostic algorithms for dermatology in people with different skin tones [92, 93]. Dangerous diagnostic bias can also occur with medical devices. For example, pulse oximeters used in critical care have been known to overestimate oxygen saturation in individuals with darker skin tones [94].

Algorithmic bias can be mitigated by removing bias in the data and/or by ensuring fair training. Bias in the data can be removed by building balanced data sets, which can be challenging for low‐frequency events and patient subsets, and the resulting balanced data set might have a smaller overall sample size. An alternative way to build balanced data sets is data augmentation [95, 96, 97], wherein a generative method, such as GAN or VAE, provides random variates. Existing GAN and VAE for PDODD and PK data could be adapted for data augmentation [70, 74, 98, 99]. An algorithmic approach to address bias in generative AI incorporates a reward for fairness into the discriminator loss function during training [96, 100]. A combination strategy that addresses fairness via the neural architecture design, training algorithm choices, and data augmentation has also been proposed for the skin lesion diagnosis problem [101]. The approach involves the L1‐norm, a metric that measures the deviations between the overall accuracy and the accuracy for the individual groups of interest. Ultimately, human supervision and review of AI methods to prevent algorithmic bias from affecting product performance and impacting patient care decisions are indispensable.

6. Challenges, Opportunities, & Strategies for Integrating In Silico Methods and Ai/ml in Clinical Pharmacology

The utility of incorporating inter‐individual variability assessments in the characterization of the clinical pharmacology properties for every drug candidate and ensuring its consideration in the observe, orient, decide, and act (OODA) loop of drug development is well established. A scoping assessment of the prevalence patterns of the indication and the variabilities of disease progression, drug target receptor distribution, and factors related to absorption, distribution, metabolism, and elimination should be conducted early on.

The critical gap in PBPK/QSP/QST modeling is the paucity of data for anatomical and pathophysiological differences in diverse populations to inform system parameters for virtual populations. Also, virtual populations in PBPK models have focused on representing individual conditions, but not multiple comorbidities. The range of disease areas and patient sub‐groups represented is limited. For example, the disease and organ impairment populations mostly represent adults. Disease sub‐groups (e.g., different cancers) are poorly characterized or inconsistent across various models and platforms (e.g., NAFLD/NASH). In general, while PBPK modeling and virtual populations have been widely applied to small molecules, such validation is often lacking for other therapeutic modalities such as biologics, gene, and cell therapies. Likewise, the characteristics of virtual populations in QSP/QST models are rarely defined or reported, which can limit the application of QSP/QST models to inform efficacy/safety prediction and optimize treatment protocols for specific patient groups.

AI/ML brings unique and new capabilities to clinical pharmacology, including the ability to learn from data and efficiently dissect large, high‐dimensional real‐world datasets. AI/ML algorithms can learn the univariate and multivariate dependencies and patterns of inter‐individual variability and covariance from RWD [62, 70, 99]. Harnessing the potential of AI/ML in clinical pharmacology requires proper integration with in silico approaches.

Integrating AI/ML methods successfully requires a well‐defined strategy and rigorous attention to technical considerations, given the diverse methods with distinct capabilities, limitations, and implementation requirements. Unfortunately, many in drug development often view AI/ML as a powerful, all‐purpose, ready‐to‐use computational tool—a “Swiss army knife” that will work “right out of the box.” Selecting AI/ML algorithms, engineering them, and tuning hyperparameters can be as time‐ and expertise‐intensive as in silico modeling methods.

ML algorithms are more useful when the datasets are relatively small [62]. AI algorithms are more adept at complex outcomes for which large datasets are available. Predictive, generative, and LLM deep learning have domains of utility in clinical pharmacology. As a rule of thumb, predictive AI models are suitable for disease modeling and treatment response classification, generative AI models are useful for clinical trial simulations, and LLMs are helpful as productivity tools (e.g., coding, report writing, search assimilation) and decision‐making. All AI/ML methods also require rigorous validation to mitigate bias.

Integration of AI/ML with in silico approaches can be accomplished via bridging, hybrid, embedded, or direct strategies. The bridging approach is exemplified by the generalized pharmacometric modeling framework, which deploys AI/ML models that leverage RWD to enhance the covariate model component of population PK/PD models. The hybrid approach uses AI/ML for population PK model selection or to build data‐enabled population PK and QSP models for dynamic modeling. Examples of the hybrid approach include the use of genetic algorithms for model selection in population modeling [102] and the use of Bayesian networks to identify QSP‐like model structures to drive stochastic modeling of the effects of aging on systems of metabolic and cardiovascular biomarkers [103]. In the embedded approach, the AI/ML is contained within the PK‐PD/PBPK/QSP/QST model. The neural ordinary differential equation (neural ODE) embedding of neural networks in the differential equations of PK structural models is an example of the embedded strategy [104]. AI/ML methods can also directly model survival curves [105], dissect PK/PD data, and build disease progression models [106].

In conclusion, in silico and AI/ML methods are highly complementary and can be used in various strategies to leverage RWD, enhance the modeling of inter‐individual differences, and generate representative virtual populations. Integrating these approaches could improve drug exposure, efficacy, and safety predictions across the lifespan, enhancing drug development and regulatory decision‐making.

Disclosure

Disclaimer: Parts of this review article were presented as a symposium at the American Society for Clinical Pharmacology and Therapeutics 2024 Annual Meeting.

Conflicts of Interest

Dr. Gonzalez receives support from industry for services related to drug development in adults and children. All other authors declared no conflicts of interest.

Yang K., Gonzalez D., Woodhead J. L., Bhargava P., and Ramanathan M., “Leveraging In Silico and Artificial Intelligence Models to Advance Drug Disposition and Response Predictions Across the Lifespan,” Clinical and Translational Science 18, no. 6 (2025): e70272, 10.1111/cts.70272.