Advancing Evidence Generation in Small, Diverse, Understudied Population

Abstract

Many clinically important patient groups remain systematically underrepresented in drug development. Among these are patients with severe or time‐varying organ dysfunction, pregnant and lactating women, children with ultrarare diseases, and people from socioeconomically or geographically underserved communities. Traditional clinical trials frequently exclude such groups. Little direct evidence is generated as a result, leaving benefit–risk and dosing decisions uncertain and trial findings poorly generalizable to the patients who will ultimately receive the drug. No single method closes this gap. This review argues that progress instead comes from integration, combining model‐informed drug development, registries and real‐world data, external and virtual controls, Bayesian borrowing, and patient‐centric trial designs into a coherent totality‐of‐evidence package. These tools are maturing from a role that supplements clinical data toward one that can, in defined settings, partially replace it. The appropriate approach may differ by population. For example, (1) mechanistic modeling and exposure prediction for organ replacement therapy/changing organ function or from the altered maternal–fetal–infant physiology that makes prospective studies impractical; (2) registry‐based external controls and natural history data where patients are too few to randomize; and (3) decentralized, partnership‐driven designs to reach populations that structural barriers would otherwise exclude. Three case studies (dose justification across hepatic and renal impairment, evidence generation in women of childbearing potential with chronic autoimmune disease, and a registry‐anchored ultrarare pediatric program) illustrated these approaches. Mapped to evolving regulatory frameworks, these strategies can deliver population‐appropriate dosing at approval rather than through postmarketing commitments, accelerating equitable access to safe and effective therapies.

Lack of diverse participant representation is often a critical shortcoming of clinical trials, which reduces the generalizability of research findings, and impedes equitable access to new therapeutics. Specifically, from a clinical pharmacology perspective, this makes dosing recommendations challenging, which may lead to an uncertain benefit–risk profile in these populations. These understudied patient groups include but are not limited to pregnant and lactating women, ¹ , ² pediatrics, ³ organ dysfunction, ⁴ , ⁵ and certain socioeconomic and minority groups. ⁶ The reasons for exclusion of these groups vary—in some, it is physiological and ethical complexity, in others prohibitively small patient numbers, and still others structural barriers to access. But the result is the same and each is left with a clinical pharmacology evidence base that does not allow for confident benefit–risk or dosing decisions. Excluding these patients from trials does not keep them from being treated. It only moves treatment into routine practice without adequate data to guide it, frequently until postmarketing studies close the gap years later. Underlying this shared gap are two distinct problems that require different solutions. The first is participation. Enrolling more of these patients is constrained by ethical concerns, mistrust, logistical burden, willingness to participate, and socioeconomic factors. ⁷ Strategies for improving participation may include expanding trial sites to accessible locations, broadening eligibility criteria in trials, lowering financial and logistical barriers, and culturally tailored outreach programs. ⁸ The second is inference, the challenge of generating reliable evidence when enrolling a sufficient and representative sample is slow, impractical, or impossible.

Most clinical trials automatically exclude pregnant women and women of childbearing potential; as a result, they remain at high risk of potentially inadequate treatment until pharmacokinetic (PK) and/or safety data become available, usually through postmarketing studies. A workshop organized by the International Maternal Pediatric Adolescent AIDS Clinical Trials (IMPAACT) network and the World Health Organization (WHO) encouraged recruitment of pregnant and lactating women in clinical trials, ² once all necessary preclinical reproductive toxicology studies have been completed. They further proposed use of population pharmacokinetic (popPK) and physiologically based PK (PBPK) models to support drug usage in pregnancy and lactation. Similarly, a multidisciplinary working group from ASCO and Friends of Cancer Research proposed expanding inclusion criteria for patients with organ dysfunction. ⁴ The recommendations covered inclusion of (1) patients with creatinine clearance > 30 mL/min, if renal toxicity and clearance were not of direct treatment‐related concern, (2) patients with mild‐to‐moderate hepatic dysfunction, if totality of the data supported that it was safe to do so, and (3) use of ejection fraction values with investigator assessment for heart failure risk to determine eligibility.

The policy landscape too remains unsettled. In 2022, the US Food and Drug Administration (FDA) issued a draft guidance, which was subsequently removed in 2025, providing a framework for sponsors to include broader populations in clinical trails through diversity plans. ⁹ This underscored the importance of diversity in clinical trials, which recommended that sponsors seek diversity in clinical trial enrollment inclusive of race and ethnicity and underrepresented populations defined by demographics such as sex, gender identity, age, socioeconomic status, disability, pregnancy status, lactation status, and co‐morbidity. An International Consortium of Quality and Innovation (IQ) assessment on the implementation of the draft FDA guidance by members of the IQ consortium found that there were general awareness and agreement with the concepts laid out in the draft guidance. ¹⁰ Additionally, the whitepaper envisioned that industry efforts in the inclusion of previously underrepresented populations during global drug development and applications of model informed drug development (MIDD) will culminate in holistic data generation. This will support labels that apply to the intended patient populations at the time of drug or biologic approval rather than through postmarketing requirements.

MIDD is the principal means of addressing the inference problem. By embedding prior knowledge from preclinical data, mechanistic physiology, and related populations within quantitative models, it can predict drug exposure, and in some cases pharmacodynamic response, in settings where direct clinical study is limited. Its methods include population and physiologically based pharmacokinetic modeling, exposure‐response analysis, external and virtual control groups, Bayesian dynamic borrowing, and the structured use of registries and real‐world data. Individually, these methods are rarely sufficient. The applications that have influenced dosing and labeling have generally combined several of them into a single evidence package suited to the population in question. A related shift concerns the role this evidence plays. Historically, it has supplemented clinical data. In defined and well‐qualified settings, MIDD has begun to replace clinical study, as summarized in Table 1 . An IQ consortium white paper anticipated this outcome, in which labeling would reflect the intended population at the time of approval rather than being deferred to postmarketing requirements. ¹⁰ PBPK modeling that supported standard dolutegravir dosing in pregnancy and lactation is one early example, ¹¹ and this review describes many others.

Table 1.

Application of MIDD approaches to replace clinical study

Method	Population	Data replaced	Regulatory precedent	Validation standard	Residual uncertainty
PBPK modeling	Pediatrics (especially infants/children)	Dedicated pediatric PK trials for dose selection	FDA/EMA pediatric extrapolation guidance; used for drugs like voriconazole, midazolam dosing	Comparison to sparse observed pediatric concentrations or biomarkers	Immature/variable ontogeny of enzymes and transporters, especially in neonates
Population PK (sparse sampling) models	Pregnant individuals	Traditional intensive serial‐sampling PK studies	For lenacapavir, pregnancy popPK analysis was used to support dose	External validation against opportunistic/sparse concentration data from pregnancy registries	Gestational‐stage‐dependent physiology and placental transfer not fully captured
Renal function‐based scaling models	Renal impairment (including dialysis)	Dedicated renal impairment trials (in select cases, per waiver criteria)	For infigratinib, popPK modeling using renal function covariate was used to justify dose	Comparison to observed clearance in impaired‐renal‐function cohorts	Sparse data in end‐stage disease and inter‐individual variability in residual renal function
Exposure‐matching/bridging models	Racial/ethnic subgroups	Separate confirmatory efficacy/safety trials in bridging contexts	For eltrombopag, popPK/PD was used to justify dose based on exposure matching	Post‐marketing PK/PD surveillance and external cohort comparison	Genetic polymorphisms and diet/environment effects not always modeled
Model‐informed geriatric extrapolation	Older adults	Additional dose‐finding trials beyond adult data	For berotralstat, covariate analysis showed age had no clinically meaningful effect on exposure	Comparison to real‐world geriatric exposure/outcome data	Polypharmacy, comorbidity burden, and frailty poorly represented in models

This review examines evidence‐generation strategies for four populations, organized around why each population is difficult to study and which tools address those constraints. The populations were selected based on authors' assessment of those most pertinent for worldwide drug development. The case studies were selected based on review of published literature and drug approval summaries from FDA and EMA. Three approaches recur across the sections—(1) application of MIDD where static dosing breaks down, (2) use of registry‐based external controls and natural‐history data for populations too small to support randomization, and (3) decentralized and partnership‐driven trial designs to extend a trial's reach to patients that structural barriers would otherwise exclude. Each population section combines a focused analysis of these strategies with a case study. The aim of this review was to provide a holistic overview to show that a common, integrative approach can characterize these populations at the point of approval (Figure 1 ), rather than through later postmarketing studies.

Figure 1.

Building a coherent evidence base in understudied populations using a combination of approaches, supported by advanced modeling and simulation. Figure was generated using an AI tool.

PATIENTS WITH SEVERE, TIME‐VARYING, OR MECHANISTICALLY POORLY CHARACTERIZED ORGAN DYSFUNCTION

Patients with severe, time‐varying, or mechanistically heterogeneous organ dysfunction are treated frequently in routine care yet are often excluded or represented only by small, carefully selected subgroups in trials. The central issue for such populations is that precisely accounting for the factors impacting drug exposure becomes difficult. For example, impaired renal function can manifest in hours to days during sepsis and acute kidney injury (AKI) and albumin changes following critical illness can alter volume of distribution and free fraction, undermining static dose‐adjustment schemes. Similarly, when renal replacement therapy (RRT) is used, the extracorporeal circuit adds a second, highly variable clearance pathway; modality (intermittent HD, PIRRT/SLED, CRRT), filter properties, delivered dose, and downtime can dominate drug concentrations and are rarely standardized across centers. ¹² , ¹³ Hepatic impairment adds a different kind of uncertainties wherein commonly used severity surrogates (e.g., Child–Pugh) are prognostic rather than mechanistic and do not reliably reflect enzyme/transporter capacity, shunting, or hypoalbuminemia. ¹⁴ In older adults, frailty and sarcopenia can systematically misclassify renal function using creatinine‐based equations and polypharmacy increases interaction risk, compounding PK/PD variability. ¹⁵ , ¹⁶

Regulatory expectations for dosing in organ impairment generally start from a clear accounting of elimination pathways, protein binding, and the anticipated sensitivity of exposure to changes in organ function. The global regulatory landscape features guidance on renal and hepatic impairment as well as geriatrics, dating as far back as the 1990s. ¹⁷ , ¹⁸ , ¹⁹ , ²⁰ , ²¹ , ²² These guidances define when dedicated data are needed and outline a wide range of study designs to address data gaps, with the extent of data generation scaled to the expected difference between these groups vs. the “reference” population. Approaches range from no dedicated studies, for example, for many mAbs in RI, to dedicated studies covering the full impairment spectrum, for example, for small molecules displaying ≥ 30% of unchanged drug in urine. Stepwise approaches are often recommended, first assessing whether a relevant difference exists then characterizing it further if clinically relevant ¹⁷ or starting with a “worst case scenario.” ¹⁸ , ¹⁹ In silico approaches such as popPK and PBPK are mentioned as an integral part of the data generation tools. Earlier inclusion in clinical trials remains a key goal, with avenues to explore including broader or gradually expanded inclusion criteria as benefit–risk becomes clearer (e.g., through adaptive designs or open‐label studies), reduced visit burden, financial and logistical support, appropriate site selection, and targeted engagement with patient and advocacy groups. Regulatory guidance for renal and hepatic impairment, geriatrics, and model‐informed drug development has prompted pragmatic evidence‐generation strategies for understudied populations. Approaches often combine targeted clinical studies, mechanistic modeling, and/or real‐world data to develop actionable dosing recommendations. The common development pattern is to define the relevant phenotype, gather essential data for dose selection, and use model‐informed or external evidence to bridge gaps when prospective trials cannot fully resolve all questions.

Case study: Overcoming the challenge of developing drugs in light of severe/time‐varying/mechanistically poorly characterized organ dysfunction

An example of ensuring adequate consideration of impaired organ subpopulations can be found in olaparib. This PARP inhibitor has a clinically relevant degree of renal elimination; however, direct clinical data in severe renal impairment was limited in early development. As such, the developer used PBPK modeling to extend available evidence within a regulator‐reviewable framework. A verified PBPK model was used to simulate exposure in scenarios where clinical data were lacking, including severe renal impairment, and to derive initial dosing recommendations that were subsequently reflected in development and labeling considerations. ²³ The PBPK model for olaparib was also used to recommend reduced dosing in moderate and severe renal impairment on the basis of predefined exposure targets. ²⁴ Taken together, these analyses illustrate how direct clinical data can be supplemented by mechanistic modeling to bridge understudied renal impairment strata in alignment with FDA renal impairment and PBPK guidance. The renal impairment framework arguably becomes even more complex when dialysis is considered since drug exposure reflects both intrinsic kidney dysfunction and extracorporeal clearance. A meropenem PBPK‐PD analysis by Deng et al. addresses the regulatory need to characterize dialysis modality, prescription, timing, and residual renal function, rather than treating dialysis as a homogeneous category. ²⁵ Here, researchers developed and scaled a mechanistic model for chronic kidney disease and hemodialysis populations. They simulated drug exposure during both on‐dialysis and off‐dialysis periods to guide regimen‐specific dosing. By explicitly incorporating dialysis modality, downtime, and residual renal function as covariates in the model, the analysis transformed a heterogeneous clinical population into distinct and measurable dosing scenarios. This enabled the generation of evidence that was directly translatable into clinically relevant and phenotype‐specific dosing recommendations despite limited datasets, supporting more robust and individualized therapeutic decisions. The level of knowledge around hepatically impaired populations remains below that of renally impaired populations, as reflected, for example, in oncology trials where inclusion of severe HI mandates FDA consultation, while RI inclusion is less restrictive. ²² Evaluating this population is also complicated by disconnects between organ impairment classifications and their mechanistic implications for clearance/metabolism. In addition, inconsistencies between classification systems can hinder patient selection and interpretation. Elmeliegy et al. showed that the Child–Pugh and National Cancer Institute (NCI) hepatic impairment criteria frequently assign different severity categories to the same patient, complicating stratification and dosing guidance. ²⁶ A 20‐year review of oncology NDAs and BLAs identified the “success rate” of popPK‐based categorical approaches to be 96% and 100%, and 21% and 50%, and 0% and 25% for mild, moderate, and severe HI patients, respectively, ²⁷ highlighting not only the relevance of model‐based analysis to the question but also the necessity of complementary methods in more severe and/or less prevalent impairment stages. To address these limitations of classification systems, more mechanistic model‐informed approaches are increasingly used alongside traditional hepatic impairment scoring. Willman et al. applied PBPK modeling to rivaroxaban, predicting the effects of various degrees of hepatic impairment on drug exposure by incorporating mechanistic determinants such as hepatic clearance and liver dysfunction. ²⁸ The model was evaluated against clinical pharmacokinetic data, enabling examination of impairment scenarios that would be difficult to study comprehensively in conventional trials. This program and many others ²⁹ , ³⁰ show how PBPK modeling can extend the utility of FDA and EMA hepatic impairment frameworks, using guidance‐defined strata as regulatory anchors while PBPK provides mechanistic interpretation for heterogeneity and dose decisions when direct evidence is limited. Both for RI and HI, more recent research aims to use in silico methods to fully replace clinical trial data rather than supplement existing data. One application based on virtual controls tackles the recurring limitation in conducting OI studies of recruiting a demographically matched control group. Retrospective analyses have demonstrated that using popPK‐generated controls showed good concordance for RI and HI assessments vs. using in‐study controls. ³¹ , ³² The results were primarily based on GMR comparisons, using data from 9 and 11 OI studies across 8 and 7 drugs, respectively. In similar work, Barry et al. compared the use of historical vs. popPK controls for 2 drugs, showing that both approaches are valid alternatives to assess the impact of OI on PK. ³³ While the “proof‐of‐concept” of such approaches is clear, their transition from in progress to established innovations will hinge on establishing best practices and standard methodology, including when to use either historical and/or virtual controls and how to perform matching and popPK model qualification. Further analyses are required on the performance of modeling the relationship between renal function and PK parameters, which can ultimately be preferred over categorical GMR analysis for the derivation of the most adequate dose adjustments. ¹⁸ Further pushing the boundary by using PBPK approaches as primary evidence of OI impact on PK, and waiver of clinical studies, as can be done for certain DDI studies under specific conditions, ³⁴ remains a goalpost. To establish the current performance of PBPK models to predict the impact of OI on PK, a thorough review was performed including data from 106 OI studies across 29 compounds from 19 companies, comparing observed and predicted GMRs across RI and HI categories. ³⁵ PBPK predictions were found to perform particularly well for mild and moderate RI, but less so for severe RI and HI, mostly overpredicting OI impact. Highlighted limitations include known and unknown factors affecting OI processes not accounted for in PBPK OI populations, such as changes in CYP activity due to uremic toxins or reduction in liver transporter capacity. These methods will move closer to confirmatory settings as mechanistic understandings increase over time, supported by advances in fields such as organs‐on‐a‐chip where (patho)physiological models of the renal and hepatic systems are evolving. ³⁶ , ³⁷ This framework extends to geriatrics, where partial organ dysfunction, frailty, altered body composition, and polypharmacy converge. Approaches such as covariate analyses described for renal and hepatic impairment are also relied upon for this subgroup. For example, investigators pooled belzutifan ³⁸ PK data across clinical pharmacology and patient studies, evaluating age, body size, renal function, and mild hepatic impairment within an integrated population PK model, or in the case of bilastine ³⁹ a dual strategy to dose selection in older adults was employed which relied on a PBPK model and a semi‐mechanistic “senescence” population pharmacokinetic model. In addition to model‐based approaches, a key aspect for geriatrics is focusing on trial inclusion due to the large size of this population. Haerrter et al. estimate that the proportion of elderly during trials to be less than half their actual proportion in the different disease populations. ⁴⁰ Key barriers to inclusion of elderly patients in clinical trials and recommendations on how to address them are gaining visibility. ⁴⁰ Decreasing the use of technology and providing additional support for mobility for older patients are examples being explored to decrease the mismatch in elderly participation in clinical trials.

PREGNANT AND LACTATING POPULATIONS, INCLUDING IN RARE AND CHRONIC DISEASES

Pregnant and lactating women remain among the most poorly characterized groups in drug development, even though medicine use during pregnancy and the postpartum period is common and frequently unavoidable. This is particularly important in the context of rare and chronic diseases, where interruption of treatment may compromise maternal health and consequently affect fetal or neonatal outcomes. ⁴¹ , ⁴² Women of childbearing potential with autoimmune disease, hypertension, diabetes, or infectious disease are therefore often managed in the setting of considerable therapeutic uncertainty, not because medicines are avoided, but because they are used in the absence of robust pregnancy‐ or lactation‐specific data. ⁴¹ , ⁴²

This lack of evidence largely reflects the historical exclusion of pregnant individuals from clinical studies, driven by concerns over fetal safety, regulatory complexity, and perceived liability. However, this approach has increasingly been questioned, as it does not prevent exposure but instead shifts treatment decisions into routine practice without robust data on dosing, efficacy or safety. ⁴¹ Consequently, the key ethical issue is not only the risk of conducting research in pregnancy, but the risk of continued exclusion and resulting uncertainty. The National Academies have argued that liability concerns are often overstated, while empirical evidence suggests that participation is influenced by perceived benefit, disease severity and trust in clinical teams, rather than risk alone. ⁴¹ , ⁴³ Taken together, these data indicate that exclusion may perpetuate, rather than mitigate risk. ⁴¹ , ⁴³

Pregnancy produces substantial physiological changes that may alter all major determinants of pharmacokinetics. Increased plasma volume and total body water, enhanced cardiac output, increased renal blood flow and glomerular filtration, reduction in circulating albumin concentrations and gestational changes in metabolic enzyme and transporter activity may all alter absorption, distribution, metabolism, and excretion. ⁴² , ⁴³ These changes are not uniform across compounds and are often trimester‐dependent, which means that extrapolation from non‐pregnant women is frequently unreliable. However, the clinical implications of these changes were often uncertain because pharmacodynamic, exposure–response, and outcome data were sparse. Coppola et al. found that, although pregnancy‐related alterations in drug handling are well‐recognized, the clinical literature remains sparse, and dosing recommendations are unavailable for many medicines used in pregnancy. ⁴⁴

This has important practical consequences. The recommended adult dose may under‐expose pregnant patients if renal clearance increases or metabolic pathways are induced, but the converse may also occur if changes in protein binding or pathway‐specific metabolism increase free concentrations. The challenge is therefore not just that pregnancy alters PK, but that these changes are drug‐specific and mechanistically heterogeneous. ⁴³ This is especially problematic in chronic conditions where sustained therapeutic exposure is required to maintain efficacy. In such settings, treatment failure due to under‐dosing may be as harmful as over‐exposure, particularly when maternal disease activity itself is an important determinant of pregnancy outcome.

For this reason, model‐informed approaches have been increasingly recognized as important tools in pregnancy clinical pharmacology. Ke et al. describe how physiologically based pharmacokinetic (PBPK) modeling can be used to inform dosing in pregnant populations. ⁴⁵ In particular, the authors describe how PBPK modeling allows drug‐specific properties to be integrated with gestation‐dependent physiological changes to predict maternal and fetal exposure under a range of clinical scenarios. Such approaches are especially useful where traditional clinical studies are difficult, slow to conduct, or ethically complex. Importantly, PBPK methods do not circumvent the need for clinical data; rather, they provide a framework within which sparse clinical observations can be integrated, interpreted, and extrapolated rationally. ⁴⁵

A mechanistic maternal–fetal–lactation dolutegravir PBPK model was developed using clinical data from non‐pregnant women, pregnant women, umbilical cord samples, lactating mothers, and breastfed infants. ¹¹ The analysis demonstrated that, although maternal exposure varied according to UGT1A1 phenotype, predicted maternal and fetal concentrations remained above the pharmacological target concentration during late pregnancy, supporting standard dosing in the settings evaluated. The authors also predicted low infant exposure through breast milk, with relative infant daily dose remaining below 2%. This example highlights that pregnancy‐related physiological changes do not necessarily result in dose changes, but beyond PK, it does require quantitative assessment against known pharmacological targets. ¹¹ , ⁴⁵ , ⁴⁶

Given the limited availability of prospective interventional data, post‐marketing pharmacovigilance remains key for evidence generation in pregnancy. Data from registries, spontaneous reports, claims databases, and electronic health records are informative but inherently constrained by reporting bias, lack of denominator data and confounding by indication. The analysis by Sato et al., based on 521 cases of maternal exposure across chronic conditions, illustrates both the value and limitations of such data, identifying signals for adverse perinatal outcomes while highlighting the difficulty of distinguishing drug effects from underlying disease risk. ⁴⁷

Pregnancy pharmacovigilance is increasingly considered within a broader real‐world evidence framework, where linked datasets integrating prescribing, dispensing, and clinical information can improve the accuracy and completeness of medication exposure data and enable more informative comparisons than spontaneous reports alone. However, the value of such data depends on their suitability for the specific clinical or regulatory question, including the quality, completeness, and analytic approach. In practice, the most reliable evidence is likely to come from complementary sources, including registries, observational datasets, and mechanistic models, rather than reliance on any single approach. ⁴⁷ , ⁴⁸

Lactation raises a related but distinct set of challenges. The clinically relevant issue is not simply whether a medicine passes into breast milk, but whether the quantity transferred produces meaningful infant exposure after accounting for maternal concentration, milk composition, feeding patterns, infant oral bioavailability, and developmental pharmacology. Pansari et al. outlined how PBPK approaches may be applied to lactation research to address precisely these questions. ⁴⁹ By combining maternal PK with milk composition and drug physicochemical characteristics, these models can be used to predict milk‐to‐plasma ratios, infant daily dose, and relative infant dose even where clinical lactation studies are unavailable. ⁴⁹ This is an important advance because lactation data are frequently absent at the point of prescribing, leaving clinicians to make decisions on the basis of incomplete or indirect evidence.

The utility of this approach is further demonstrated in the recent PBPK‐led analysis of antimalarial drug concentrations in breast milk. ⁵⁰ Almond et al. showed PBPK modeling could correctly classify relative infant dose as above or below the World Health Organization 10% threshold for antimalarial drugs with available clinical lactation data. In addition, the model could then be used prospectively to identify compounds for which a dedicated lactation study or specific breastfeeding caution might be required. ⁵⁰ They also demonstrated that predictions were sensitive to multiple factors, including plasma protein binding, lipophilicity, milk pH, and creamatocrit, underlining the fact that drug transfer into milk cannot be inferred reliably from simple assumptions based on molecular size or therapeutic class alone. ⁵⁰ Together with the dolutegravir analysis, this work supports a more quantitative and decision‐oriented approach to breastfeeding recommendations. ⁴⁶ , ⁴⁹ , ⁵⁰

More broadly, mechanistic prediction of maternal–fetal drug exposure is increasingly central to clinical pharmacology in pregnancy. Direct measurement of fetal exposure is largely confined to delivery‐associated samples such as cord blood, which provide little information on longitudinal exposure across gestation. In contrast, PBPK models offer a way of integrating maternal physiology, placental transfer, and fetal disposition within a coherent framework that can be interrogated across trimesters, genotypes, and dosing regimens. ¹¹ , ⁴⁵ , ⁴⁶ This does not eliminate uncertainty, but it substantially improves the ability to make structured inferences when clinical studies are impractical or limited. The broader significance is that pregnant and lactating populations are no longer scientifically inaccessible; rather, they are populations in whom evidence generation must rely on a combination of targeted clinical studies, observational data and mechanistic modeling‐based simulations. ¹¹ , ⁴⁵ , ⁴⁶ , ⁴⁹

Case study: Chronic autoimmune disease in women of childbearing potential

Chronic autoimmune diseases provide a clear illustration of the challenges of managing treatment across the reproductive continuum. These conditions commonly affect women of childbearing potential (WOCBP) and often require sustained immunomodulatory therapy before conception, during pregnancy and throughout lactation. This is exemplified by certolizumab pegol (CZP), an Fc‐free anti‐tumor necrosis factor, for the treatment of inflammation and chronic autoimmune conditions. CRADLE ⁵¹ and CRIB ⁵² were prospective postmarketing PK studies designed to quantify certolizumab concentrations in human breast milk and estimate average daily infant dose of maternal certolizumab, and evaluate placental transfer of certolizumab during pregnancy, respectively. These studies demonstrated minimal transfer of certolizumab in breast milk and placenta. Thus, supported use in pregnancy/breastfeeding without a specific maternal dose. ⁵³ However, WOCBP, pregnant, or lactating women are routinely excluded or underrepresented in clinical trials due to concerns about embryo‐fetal risk and complications during pregnancy. Physiological changes during pregnancy can impact the PK and/or PD of drugs; as a result, there might be a need to modify doses. Additionally, disease pathophysiology, autoantibodies, organ involvement, and use of co‐medications can affect maternal and fetal outcomes, so extrapolating from general adult trial populations is often inadequate. ⁵⁴ Evidence suggests that several commonly used therapies, including glucocorticoids, hydroxychloroquine, and ciclosporin, are not associated with increased risk of congenital abnormalities, whereas others, such as methotrexate and cyclophosphamide, remain contraindicated. ⁵⁵ Importantly, adverse outcomes such as prematurity and low birth weight often reflect underlying disease activity rather than treatment alone. ⁵⁵ Management therefore prioritizes disease control with pregnancy‐compatible therapies, supported by preconception planning and careful treatment selection. This example underscores a broader principle in chronic disease: the risks of undertreatment frequently outweigh those of appropriately selected therapy, necessitating an evidence‐informed, rather than avoidance‐based, approach. ⁵⁵ The draft ICH E21 guidance provides recommendations for the appropriate inclusion of pregnant and/or breastfeeding women in clinical trials to facilitate the generation of robust clinical data in that population when WOCBP are the anticipated user population. ⁵⁶ The guidance encourages sponsors to develop strategies and proactively plan to generate data in pregnant and/or breastfeeding women. Sponsors are also recommended to consult with regulatory authorities as early as possible and as needed throughout the development process regarding the plans for the participation of pregnant and/or breastfeeding women in clinical trials. Several general principles could be implemented by the sponsors, such as: (1) clear definition of WOCBP, including criteria for menarche, menopause, sterilization, and alignment with any applicable guidance; (2) clear pregnancy testing schedule throughout the trial; (3) contraception requirements as per fetal risk based on nonclinical reproductive toxicology and genotoxicity data; (4) reduce undue burden such as flexibility in timing of trial assessments, collecting sparse PK samples, and not conduct procedures which carry risk for fetus such as X‐rays; and (5) explicit instructions on reporting of pregnancy during trial.

ULTRARARE PEDIATRIC DISEASE POPULATION

Ultrarare pediatric diseases represent one of the most methodologically and ethically complex frontiers for pharmacokinetic–pharmacodynamic (PK‐PD) evidence generation within the MIDD paradigm. These conditions, often defined by extremely low prevalence (typically fewer than 1 in 50,000 individuals and sometimes fewer than 1 in 200,000 individuals) ⁵⁷ , present a convergence of constraints: very small sample sizes, substantial biological heterogeneity, developmental variability, and urgent unmet medical need. In this context, conventional paradigms of clinical trial design and dose optimization are not merely inefficient; they are often infeasible. Consequently, advancing PK‐PD understanding in ultrarare pediatric populations requires a synthesis of innovative trial design, real‐world data integration, and mechanistically informed modeling approaches, grounded in the regulatory frameworks set out in ICH E11(R1) on pediatric extrapolation ⁵⁸ and the FDA's 2023 draft guidance Considerations for the Design and Conduct of Externally Controlled Trials. ⁵⁹

Challenges and opportunities in enrollment and trial design

The primary barrier to evidence generation in ultrarare pediatric diseases is the sheer scarcity of patients. Recruitment is often geographically fragmented, requiring multinational coordination even to identify a handful of eligible participants. This logistical complexity is compounded by diagnostic delays, limited disease awareness, and heterogeneity in clinical presentation. For many ultrarare diseases, natural history is poorly characterized, further complicating eligibility criteria and endpoint selection. The FDA's 2019 guidance Rare Diseases: Natural History Studies for Drug Development formalized this position. ⁶⁰ It noted that natural history information is usually unavailable or incomplete for most rare diseases and that well‐designed prospective natural history studies can serve as external control groups for single‐arm trials when randomization is impractical or unethical. As a result, prospective natural history studies are increasingly initiated in parallel or in advance of the interventional program itself.

From a PK‐PD perspective, pediatric populations introduce additional layers of complexity due to ontogeny. Age‐dependent changes in organ function, drug‐metabolizing enzyme expression, transporter activity, plasma protein binding, glomerular filtration, and body composition modify drug absorption, distribution, metabolism, and excretion in ways that are not linear with body weight and are not adequately captured by simple weight‐based scaling. ⁶¹ These developmental processes are quantitatively characterized through several complementary approaches—PBPK models incorporating ontogeny functions for individual clearance pathways, ⁶² allometric scaling with theoretical exponents of 0.75 for clearance and 1.0 for volume of distribution, ⁶³ and sigmoid maturation (Hill) functions describing the gradual approach of clearance to adult values during early life. ⁶⁴ In ultrarare diseases, these developmental factors intersect with disease‐specific pathophysiology, making extrapolation from adult or even more common pediatric conditions highly uncertain. As a result, traditional dose‐finding studies are often impractical, and reliance on sparse sampling strategies, population PK modeling with maturation covariates, and PBPK‐based exposure prediction becomes the standard methodological response. ⁶⁵

Ethical considerations also shape trial design. ICH E11(R1) explicitly states that pediatric clinical investigations should expose participants to no more risk than is justified by the anticipated benefit, and that the balance of risk and anticipated clinical benefit must be at least comparable to available alternatives so that the child is not disadvantaged by enrollment. ⁵⁸ Given the severity and progressive nature of many ultrarare pediatric conditions, placebo‐controlled trials may be considered unacceptable by patients, families, and clinicians. This ethical constraint limits the use of randomized controlled designs and increases reliance on single‐arm studies or externally controlled comparisons. However, such approaches introduce challenges in causal inference and bias, particularly when historical or external controls are not well matched, and require prespecified analytic plans, sensitivity analyses, and tipping‐point assessments to be regulatorily defensible. ⁵⁹

Adaptive trial designs and Bayesian methods have become central in this space. Specific tools include Bayesian dynamic borrowing with meta‐analytic‐predictive (MAP) priors, ⁶⁶ commensurate and power priors that down‐weight external information when it conflicts with concurrent data, master and basket protocols that share controls across mechanistically related indications, and continual reassessment dose‐finding designs adapted to pediatric exposure targets. These approaches enable formal incorporation of adult, preclinical, and related‐disease information into the pediatric inference, partially mitigating the limitations imposed by small sample sizes.

Use of rare disease registries

Rare disease registries have become indispensable resources for overcoming many of the challenges associated with ultrarare pediatric populations. Examples relevant to pediatric PK‐PD include disease‐specific natural history registries such as DEM‐CHILD for CLN2, ⁶⁷ TREAT‐NMD for muscular dystrophies, ⁶⁸ the Pompe registry, ⁶⁹ the CINRG Duchenne Natural History Study, ⁷⁰ and the Hunter Outcome Survey for MPS II, ⁷¹ alongside cross‐disease platforms (NORD's IAMRARE, Sanford CoRDS, Orphanet, the EURORDIS‐supported European Reference Networks) and increasingly sponsor‐led prospective natural history studies designed in parallel with the interventional program. For PK‐PD research, registries provide a critical foundation for understanding disease natural history, identifying potential endpoints, and informing model development. Several of these have served directly as the external comparator cohorts in regulatory submissions for products approved of single‐arm pediatric trials. ⁶⁷ , ⁶⁸

One of the key contributions of registries is their ability to contextualize PK‐PD relationships within the broader trajectory of disease progression. In the absence of large, randomized trials, understanding baseline variability and progression rates is essential for interpreting treatment effects. Registries can also support the identification of covariates that influence drug response, such as genotype, disease severity, or concomitant therapies.

However, the use of registry data is not without challenges. Data quality and completeness can vary significantly across registries, and standardization of data elements is often lacking. Inconsistencies in sampling timing, assay methods, and outcome definitions can introduce substantial noise. Furthermore, registries are typically observational in nature, raising concerns about confounding and selection bias when used to inform causal inference. Regulatory agencies have responded with formal frameworks for assessing registry fitness‐for‐purpose. The EMA Patient Registries Initiative ⁷² and the 2021 guideline on registry‐based studies ⁷³ and the FDA's 2023 guidance on assessing registries to support regulatory decision‐making proposed the data‐quality, governance, and methodological criteria a registry must meet before its data can support a regulatory submission. ⁷⁴ Despite these limitations, integration of registry data with clinical trial data is increasingly feasible through advanced statistical and computational methods. Techniques such as hierarchical modeling and data harmonization enable the combination of heterogeneous data sources, enhancing the robustness of MIDD. As discussed below for CLN2 disease, registries have been used to construct external control arms, ⁷⁵ providing a comparator for single‐arm trials.

Case study: Ceroid lipofuscinosis neuronal 2 (CLN 2) disease

The challenges of ultrarare pediatric populations push evidence generation into the MIDD and model‐integrated evidence (MIE) paradigm. No single component (adaptive design, registry‐based external controls, MBMA, Bayesian dynamic borrowing) is sufficient on its own but together they constitute the totality of evidence package that support regulatory approval. Augmenting evidence generation in ultrarare pediatric populations is no longer a methodological aspiration. Recent approval of cerliponase alfa in CLN2, ⁷⁶ relied on registry‐derived external controls, model‐based exposure extrapolation, biomarker‐anchored bridging, and small open‐label confirmatory cohorts.

CLN2 disease (classical onset 2–4 years of age) is caused by biallelic loss‐of‐function variants in TPP1, the gene encoding the lysosomal protease tripeptidyl peptidase 1. ⁷⁷ Incidence in Western populations is approximately 1 per 100,000 live births (most recently estimated at 0.99 per 100,000 in a French nationwide population‐based study, 2017–2023), with global estimates ranging more broadly depending on founder effects in specific populations. ⁷⁸ Affected children present with seizures and language regression, then progress through motor decline, blindness, and death typically by mid‐teens. ⁷⁷ , ⁷⁹ Because the disease is uniformly fatal in childhood, no adult population exists from which to extrapolate and all dose selection, endpoint validation, and efficacy inference had to be derived from pediatric data.

The clinical endpoint that anchored the cerliponase alfa development program was the CLN2 Clinical Rating Scale, ⁷⁶ originally proposed by Steinfeld and colleagues in 2002 as a quantitative descriptor of disease progression. ⁸⁰ This scale was subsequently refined and psychometrically validated for clinical‐trial use. ⁸¹ Its decline characteristics were then established prospectively in the DEM‐CHILD natural history database, a multi‐site international registry of genotype‐confirmed CLN2 patients with longitudinal motor‐language assessments. ⁷⁹ Nickel and colleagues documented a mean rate of decline of approximately 2 points per 48 weeks in untreated patients across two independent international cohorts, providing the quantitative comparator subsequently accepted by both EMA and FDA. ⁷⁹

The pivotal evidence package for cerliponase alfa represents a well‐characterized example of a registry‐anchored externally controlled trial in an ultrarare pediatric indication. A single‐arm phase 1/2 open‐label dose‐escalation study (NCT01907087) enrolled 24 patients across five centers in Germany, Italy, the United Kingdom, and the United States (23 in the efficacy population), with the comparator constructed from 42 historical untreated patients in the DEM‐CHILD database. ⁷⁵ The natural history cohort defined the untreated trajectory at 2.12 ± 0.98 points lost per 48 weeks, while treated patients lost 0.27 ± 0.35 points per 48 weeks (mean difference 1.85; 95% CI 1.51–2.18; P < 0.001). ⁷⁵ The drug is recombinant human TPP1 delivered intraventricularly every two weeks at 300 mg via a surgically implanted intraventricular access device, a delivery approach that was developed alongside the molecule itself and which now has its own best‐practice operational guidelines.

FDA approved Brineura on April 27, 2017. ⁷⁶ The European Commission granted marketing authorization on May 30, 2017, following a CHMP positive opinion in March of that year. ⁸² Subsequent data have extended the package in two directions that further illustrate the MIDD/MIE strategy. A 240‐week open‐label extension study (NCT02485899) provided more than five years of follow‐up in the original cohort, showing durable slowing of motor‐language decline through the expected phase of most rapid disease progression, ⁸³ and real‐world data from the DEM‐CHILD registry, with 1:1 matching of treated patients to natural‐history controls on baseline age and motor‐language score, has independently replicated the trial‐derived effect outside the trial population. ⁸⁴ The July 2024 supplemental BLA (Study 190–203, NCT02678689) then expanded the indication to children of all ages and to presymptomatic patients identified by genotype prior to clinical onset. ⁷⁶ The presymptomatic approval is the conceptually interesting piece with no concurrent control and no possibility of one, the disease‐progression model derived from the natural‐history registry is the entire counterfactual.

The cerliponase alfa program demonstrates a pragmatic totality of evidence approach that ultrarare pediatric programs can realistically deliver—(1) 24‐patient single‐arm trial, (2) a curated natural‐history registry providing 42‐patient external comparator, (3) an endpoint with documented decline characteristics in both cohorts, (4) a novel CNS‐targeting delivery mechanism developed in parallel, four years from first patient to FDA approval, (5) long‐term follow‐up confirming durability, (6) real‐world data replicating the trial effect, and (7) subsequent regulatory expansion built on natural history based disease progression modeling. None of these components alone meets the standard of a conventional phase III program. However, combined, they have supported initial approval, expansion to younger ages, and treatment of children who have not yet developed clinical disease.

SOCIOECONOMIC AND GEOGRAPHIC GROUPS

The generation of robust PK‐PD evidence in small, diverse, and understudied cohorts is fundamentally constrained by structural inequities tied to geography, socioeconomic status, and functional ability. These inequities shape not only who is recruited into clinical trials, but also how data are generated, interpreted, and ultimately translated into dosing, safety, and efficacy conclusions. Addressing these disparities requires rethinking trial architecture, embedding flexibility into data collection models, and leveraging cross‐sector partnerships that can bridge infrastructure and access gaps.

Rural populations: Geographic exclusion and decentralization as a corrective

Rural populations represent one of the clearest examples of structural exclusion in clinical research. Traditional site‐centric trial designs inherently privilege proximity to academic medical centers, which are disproportionately located in urban areas. As a result, rural participants face increased travel burden, time costs, and logistical complexity, all of which contribute to under‐enrollment and higher attrition rates. Geographic isolation has been repeatedly identified as a primary barrier to participation, particularly for populations already experiencing healthcare access limitations.

The emergence of decentralized clinical trial models offers a partial corrective to this imbalance. ⁸⁵ By shifting trial activities away from centralized sites and toward patient‐centric environments—such as the home or local community, these approaches reduce travel requirements and improve retention. Decentralized trials leverage telemedicine, remote monitoring, electronic consent, and home‐based biospecimen collection to enable participation independent of geography. For PK‐PD studies, where repeated sampling and longitudinal observation are critical, such approaches can substantially reduce participant burden while preserving data density. ⁸⁵

However, decentralization is not a panacea. The “digital divide” introduces new inequities, as rural populations may have limited broadband access or lower digital literacy, potentially shifting exclusion from geographic to technological domains. Hybrid models, combining local healthcare providers, mobile units, and selective central visits, have therefore emerged as a pragmatic compromise. These models are particularly relevant for PK‐PD studies requiring controlled sampling conditions or specialized assays, where complete decentralization may compromise data integrity. ⁸⁶

Low‐income and uninsured populations: Economic barriers and structural underrepresentation

Economic disadvantage intersects with geography but introduces additional layers of exclusion. Low‐income and uninsured individuals face both direct and indirect costs associated with trial participation, including transportation, childcare, lost wages, and healthcare‐related expenses. Even when trials nominally cover medical costs, ancillary burdens remain significant deterrents. ⁸⁷

This economic gradient has measurable consequences for PK‐PD evidence generation. Underrepresentation of low‐income populations can bias pharmacokinetic estimates, particularly when socioeconomic factors correlate with nutritional status, comorbidities, environmental exposures, and concomitant medication, all of which influence drug absorption, distribution, metabolism, and excretion. As PK‐PD modeling increasingly informs dose selection and labeling, such biases risk embedding inequity directly into therapeutic recommendations.

Innovations aimed at mitigating these barriers include participant reimbursement frameworks, flexible scheduling, and integration with routine care pathways. More structurally, embedding trials within community health systems, such as federally qualified health centers or safety‐net hospitals, can reduce barriers to access. ⁸⁶ Decentralized elements again play a role, particularly in enabling asynchronous data capture and reducing opportunity costs associated with participation.

Nevertheless, there remains a tension between inclusivity and data standardization. PK‐PD studies often require controlled conditions to reduce variability; expanding participation to economically disadvantaged populations may increase heterogeneity. Rather than viewing this as a limitation, contemporary modeling approaches should treat such variability as informative, enabling more realistic characterization of exposure–response relationships across real‐world populations.

LMIC populations: Necessity, variability, and global inequities in evidence generation

Inclusion of populations from low‐ and middle‐income countries (LMICs) is both a scientific necessity and an ethical imperative. ⁸⁵ Disease burden for conditions such as tuberculosis, malaria, and many neglected tropical diseases is disproportionately concentrated in LMICs, making these populations essential for externally valid PK‐PD characterization. Moreover, environmental, genetic, dietary, and co‐infection factors can significantly alter pharmacokinetics, necessitating region‐specific data generation.

Despite this, clinical trials in LMICs remain underrepresented relative to global disease burden, and when conducted, are often led by investigators from high‐income countries. This imbalance raises concerns regarding relevance, ownership, and equity in evidence generation. Transnational research partnerships have emerged as a dominant model to address these gaps, combining financial and technical resources from high‐income settings with local expertise and contextual knowledge from LMIC partners.

Adaptive and platform trial designs have shown particular promise in LMIC settings, enabling efficient evaluation of multiple interventions within a shared infrastructure. ⁸⁶ These designs can accelerate PK‐PD learning by allowing iterative refinement of dosing strategies based on accumulating data. However, their implementation highlights persistent inequities in leadership, regulatory capacity, and data ownership, which can undermine local capacity building if not carefully managed.

Public–private partnerships (PPPs) play a central role in LMIC clinical research ecosystems (Table 2 ). These partnerships often involve collaborations among governments, non‐governmental organizations, academic institutions, and industry stakeholders, with the goal of improving health outcomes and strengthening research capacity. Successful PPPs are characterized by local ownership, alignment with national priorities, and robust data systems for monitoring progress. For PK‐PD studies, PPPs can facilitate access to diverse patient populations, support infrastructure development (e.g., laboratories, data systems), and enable long‐term follow‐up necessary for understanding exposure–response relationships.

Table 2.

Role of public–private partnerships (PPP) in enhancing inclusive PK‐PD evidence generation

PPP function	Stakeholders involved	Contribution to PK‐PD studies	Key success factors	Risks
Infrastructure development	Governments, NGOs, industry	Establish labs, trial sites, data systems	Alignment with national priorities; sustained funding	Dependency on external funding
Funding and resource mobilization	Philanthropic organizations, industry	Supports large‐scale, multi‐site PK‐PD trials	Transparent governance; equitable allocation	Power asymmetries
Capacity building	Academic institutions, local partners	Develops personnel in PK analysis and modeling	Local ownership; knowledge transfer	Brain drain
Data sharing and governance	Multinational consortia	Enables pooled PK‐PD datasets across regions	Robust data systems; equitable access	Data ownership disputes
Community engagement	Local NGOs, community leaders	Enhances recruitment and retention	Trust‐building; cultural competence	Tokenistic engagement

Individuals with disabilities: Overlooked heterogeneity and methodological implications

Individuals with disabilities represent a highly heterogeneous group that is frequently excluded from clinical trials due to perceived complexity, communication barriers, or concerns about consent and compliance. This exclusion has significant implications for PK‐PD evidence generation, as disability may be associated with altered physiology, comorbid conditions, or concomitant therapies that influence drug kinetics and dynamics.

Barriers to participation include physical inaccessibility of trial sites, lack of accommodation for sensory or cognitive impairments, and rigid protocol requirements that fail to account for functional limitations. Decentralized and hybrid trial models offer opportunities to address some of these barriers by enabling home‐based participation and flexible data collection methods. For example, remote monitoring devices and caregiver‐assisted reporting can facilitate inclusion of individuals who might otherwise be excluded.

However, inclusion requires more than logistical adaptation; it demands methodological innovation. PK‐PD models must account for variability introduced by disability‐related factors, and trial designs must incorporate stratification or subgroup analyses to ensure that data are interpretable. Ethical considerations are also paramount, particularly in ensuring informed consent and equitable risk–benefit assessment.

Integrating socioeconomic and geographic diversity into PK‐PD frameworks

Across these groups, a common theme emerges—variability is both a challenge and an opportunity. Traditional PK‐PD paradigms have often sought to minimize variability through restrictive inclusion criteria and controlled conditions. In contrast, contemporary approaches increasingly recognize that capturing real‐world heterogeneity is essential for generalizable and equitable evidence generation.

Decentralized trial designs, adaptive methodologies, and advanced modeling techniques (e.g., population PK modeling, Bayesian frameworks) enable integration of diverse data sources while maintaining analytical rigor. These approaches can incorporate covariates related to geography, socioeconomic status, and functional ability, allowing for more nuanced understanding of exposure‐response relationships. ⁸⁶

At the same time, these innovations must be implemented with careful attention to equity. Augmenting PK‐PD evidence generation in small, diverse, and understudied cohorts requires a deliberate shift from exclusionary, site‐centric paradigms toward inclusive, flexible, and partnership‐driven models. Digital tools, for example, can both expand access and exacerbate disparities, depending on how they are deployed. Similarly, global research partnerships can either empower local stakeholders or reinforce existing hierarchies. Table 3 summarizes the role of trial innovation in evidence generation.

Table 3.

Clinical trial innovations in LMICs relevant to PK‐PD evidence generation

Innovation type	Description	Relevance to PK‐PD	Advantages	Limitations
Decentralized/Hybrid trials	Remote data collection, telemedicine, home sampling	Enables longitudinal PK sampling in dispersed populations	Reduces travel burden; improves retention	Data variability
Adaptive platform trials	Multi‐arm, iterative designs with shared control groups	Efficient dose optimization and exposure–response learning	Resource‐efficient; flexible	Complex implementation; regulatory challenges
Community‐embedded trials	Integration with local healthcare systems	Captures real‐world PK variability	Improves representativeness	Potential variability in data quality
Mobile health technologies	Wearables, remote monitoring	Continuous PK/PD‐relevant data (e.g., vitals)	High‐frequency data capture	Device access and literacy barriers
Capacity‐building initiatives	Training, infrastructure development	Improves local PK assay and data capabilities	Sustainable research ecosystems	Requires long‐term investment

FUTURE DIRECTIONS & PERSPECTIVES

Patient involvement in clinical trials is critical for success; however, participating in a clinical trial is not always easy for patients. By creating a patient‐centric clinical trial experience that reduces the barriers facing participants, participation can be increased and ultimately bring medical advances to the public sooner. ⁸⁸ Patient‐centered trial design for understudied populations should focus on reducing participation barriers through decentralized, flexible, and culturally relevant approaches. Key strategies include using telemedicine, home visits, and engaging community partners to build trust, thereby enhancing equity and improving retention for minority, elderly, or remote populations. Some strategies for improving inclusion of understudied populations are—(1) decentralized trials (DCTs), which incorporate telemedicine, local nursing, and remote monitoring; reduce the logistical and financial burden of travel, which is a major barrier for underrepresented groups, ⁸⁹ (2) community partnership for building trust by engaging community leaders and local doctors early in the trial planning process, (3) designing flexible protocols that accommodate specific needs, such as providing transportation, offering child care, simplifying consent processes, improving accessibility, and (4) routine inclusion of PROs utilizing meaningful trial outcomes that matter to patients, rather than only laboratory values, ensures the trial addresses the community's actual needs.

Personalized medicine tailored to prevention, diagnosis, and treatment based on an individual's genetics, environment, and lifestyle, can improve health equity for understudied populations. An example of this is tailored dosing of allopurinol for the Hmong ethnic group, who are at increased risk of developing gout. ⁹⁰ But a major limitation to effective implementation of personalized medicine is that many understudied populations are underrepresented in the genomic and/or biomarker data that underpin personalized medicine, which creates significant knowledge gaps. A study of data from biobanks found that of the nearly 5 million samples of DNA contributed to biobanks worldwide, 68% come from individuals of European ancestry, 22% from East Asian and 10% from other non‐European ancestry. ⁹¹ The underrepresentation of people with African, Asian, and American heritage is a major pitfall for personalized medicine because diseases more commonly found in minority populations are understudied. Similarly, diseases that are more common across populations but have different progression and prognosis depending on ethnicity, such as chronic kidney disease, also remain understudied. Thus, the diversity gap in biobanks may hinder our understanding of disease in some population groups, contributing to global health disparities, such as the significantly poorer health outcomes in certain populations. ⁹² For personalized medicine to become truly “personalized,” the data must become more globally inclusive, which means investment in collecting data from diverse populations. Investment not only in terms of capital and resources but also building trust, improving data privacy and community engagement. Initiatives such as H3Africa. All of US and UK Biobank diversity efforts are needed to address the gaps in genomic data globally. ⁹³ Biomarker data, another lynchpin of personalized medicine, is also heavily skewed toward European ancestry and molecular drivers of diseases remain under‐characterized in many populations. Even though there is overwhelming evidence of variation across populations, for example, higher prevalence of triple‐negative breast cancer in women of African origin, higher prevalence of loss‐of‐function alleles of CYP2C19 in Asian populations, racial and ethnic variations of HbA1c, which has significant implications of diabetes diagnosis, to list a few. Like genomic data, approaches to improve biomarker data include—recruitment of diverse cohorts in clinical trials, establishment of population‐specific reference ranges, integration of multi‐omics data with clinical data, community engagement, and ensuring ethical use of the data. Initiatives such as AACR Project GENIE, which aims to collect and make public clinical outcomes data, including biomarkers, across 10 solid tumor cohorts are aimed to improve precision cancer medicine across populations. ⁹⁴ Such data sharing consortia are needed to globalize personalized medicine. Similarly, future development in severe, time‐varying, or mechanistically heterogeneous organ dysfunction will likely depend on moving beyond one‐size‐fits‐all evidence generation toward phenotype‐driven strategies that are feasible in small, clinically unstable populations. A central priority is to define the intended population more precisely, distinguishing, for example, stable chronic kidney disease from dynamic acute kidney injury, or decompensated cirrhosis from milder hepatic dysfunction, so that dosing recommendations are linked to clinically meaningful phenotypes rather than broad labels. ⁹⁵ Standardized capture of key covariates such as renal function determination method, renal replacement therapy settings, albumin, body composition, and comedications will be essential to improve the comparability of study findings. At the same time, future progress will likely require trial designs that promote earlier and broader inclusion of older adults and medically complex patients, including pragmatic eligibility criteria, embedded geriatric cohorts, and fit‐for‐purpose assessments of frailty and function, as emphasized in recent workshops and trial‐design recommendations for older adult accrual. Emerging strategies also point to greater use of microphysiological systems and other human‐relevant experimental platforms to strengthen mechanistic interpretation and reduce dependence on animal or difficult‐to‐conduct clinical studies, while advances in popPK, QSP, and artificial intelligence are likely to support increasingly adaptive, individualized dosing frameworks. Together, these approaches can move the field toward regulator‐ready, patient‐centered evidence that is explicit about uncertainty but actionable at the bedside.

As described throughout this review, MIDD affords an integrative approach to evidence generation in understudied populations as no single methodological advancement is sufficient on its own. Instead, a combination of these approaches, supported by advanced modeling and simulation, is required to build a coherent evidence base (Figure 1 , Table 4 ). By incorporating data from clinical trials, registries, and preclinical studies, these models can characterize variability, support individualized dosing strategies, and extrapolate to other age groups. ⁹⁶ Bayesian approaches further enhance this framework by enabling the incorporation of prior knowledge and continuous updating as new data become available. The recently finalized ICH M15 guidance intends to facilitate such a multidisciplinary understanding of MIDD and associated evidence generation by providing general recommendations for planning, model evaluation, and documentation of MIDD evidence. ⁹⁷ Several novel applications of MIDD to guide dosing decisions in understudied populations have emerged in recent years—application of virtual population and PBPK modeling in obese children, ⁹⁸ geriatric patients, ⁹⁹ prediction of milk exposure of antimalarial drugs in lactating women in absence of clinical lactation data, ⁵⁰ and opioid disposition during pregnancy to predict fetal & neonatal exposure. ¹⁰⁰

Table 4.

Established practices and emerging innovations to augment evidence generation in understudied populations

Initiative	Renal/hepatic impairment	Pregnancy and lactation	Ultrarare pediatric	Socioeconomic and geographic
Study design	Recruitment focusing on organ impaired population only; matched healthy controls leveraged from sources external to the study or virtual	Recruitment of pregnant and lactating women in clinical trials, once all necessary preclinical reproductive toxicology studies have been completed	Single arm trial, patient potentially serving as their own control	Access is key (large population), decentralized/hybrid trials, community‐embedded, mobile health technology
In‐study Analysis methodologies	Propensity score matching	Subgroup analysis from embedded trial design	Bayesian	Bayesian
Model‐based evidence	Generation of virtual controls, simulations as supportive evidence, confirmatory contributions in the making (popPK‐PD, PBPK)	Model based analysis of sparse clinical observations that can be integrated, interpreted and extrapolated rationally	Natural history based synthetic control	PK‐PD evidence generation using inclusive, flexible, and partnership‐driven clinical trials
Real‐world evidence	Limited	Pharmacovigilance	Registries	Real world use and compassionate use programs

In conclusion, augmenting evidence generation in understudied populations requires a paradigm shift toward flexibility, integration, and innovation. By leveraging novel trial design, use of RWD, embracing advanced modeling techniques, and exploring strategies such as pathway‐based data pooling, it is possible to overcome many of the challenges posed by small sample sizes and heterogeneity. While significant methodological and ethical challenges remain, continued progress in this area holds the potential to transform the development of therapies for some of the most vulnerable and underserved patient populations.

FUNDING

No funding was received for this work.

CONFLICT OF INTEREST

The authors declared no competing interests for this work.