PBPK-Based Dosimetry: Mechanistic Insights

• PBPK-based dosimetry is a modeling approach that transforms administered doses into concentration–time profiles and quantifies metrics such as AUC and Cmax.
• It integrates anatomy, physiology, and compound-specific properties into multi-compartment models for applications in radiopharmaceutical therapy, toxicology, and drug development.
• Recent advances incorporate machine learning and digital twin frameworks to enhance model calibration, prediction accuracy, and personalized treatment planning.

Searching arXiv for recent PBPK-based dosimetry and related papers to ground the article. PBPK-based dosimetry is the use of physiologically based pharmacokinetic (PBPK) models to translate an administered dose of a drug, chemical, radiopharmaceutical, or other xenobiotic into concentration–time or activity–time profiles in blood, organs, tissues, and, in some applications, voxels or microstructural domains. These profiles are then integrated or otherwise transformed into exposure metrics such as area under the concentration–time curve (AUC), $C_{\max}$, time-integrated activity, absorbed dose, biological effective dose (BED), and equivalent dose in 2 Gy fractions (EQD2). Across pharmacology, toxicology, nuclear medicine, dermal exposure science, and brain drug delivery, PBPK-based dosimetry functions as a mechanistic layer between external administration and internal dose, combining anatomical compartments, physiological flows, tissue volumes, transport processes, binding, metabolism, clearance, and, where relevant, radioactive decay or spatially resolved transport. Recent work also places PBPK-based dosimetry inside broader scientific machine learning, digital twin, and verification–validation–uncertainty quantification frameworks (Liu et al., 9 Feb 2026, Zaid et al., 25 Sep 2025, Abdollahi et al., 23 Oct 2025).

1. Definition and conceptual scope

PBPK models are mechanistic, whole-body or organ-focused models that describe how a compound moves through the body using actual anatomy, physiology, and compound-specific properties. In radiopharmaceutical therapy (RPT), they are used to predict the time course of activity in organs and tumors and thereby provide the biokinetic basis for dosimetry (Zaid et al., 25 Sep 2025). In nuclear medicine, PBPK models are described as mechanistic, whole-body models with compartments for organs and tissues such as blood, liver, kidneys, tumors, and bone marrow; physiological flows such as organ-specific blood flows $Q_i$, cardiac output, and lymphatic drainage; volumes $V_i$; tissue:blood partition coefficients $K_{p,i}$; clearance processes such as renal excretion, hepatobiliary clearance, and metabolism; and, for targeted radiopharmaceuticals, binding, internalization, and retention kinetics (Farhadi et al., 23 Nov 2025).

At its broadest, PBPK-based dosimetry is concerned with quantifying how an administered dose translates into concentration–time profiles in blood and tissues, and derived exposure metrics such as AUC, $C_{\max}$, and organ doses that drive efficacy and toxicity (Liu et al., 9 Feb 2026). In RPT, the same logic is expressed in terms of time–activity curves (TACs), time-integrated activity coefficients (TIACs), absorbed doses (ADs), BED, dose–volume histograms (DVHs), tumor control probability (TCP), and normal tissue complication probability (NTCP) (Zaid et al., 25 Sep 2025, Farhadi et al., 23 Nov 2025, Abdollahi et al., 23 Oct 2025). In dermal modeling, PBPK-based dosimetry denotes the conversion of an external skin application into local concentrations in specific skin microstructures and systemic exposure (Sebastia-Saez et al., 2018). In brain PBPK, dosimetry is framed as quantifying concentration–time profiles and exposure metrics such as AUC, $C_{\max}$, $T_{\max}$, and half-life in brain blood, brain mass, cranial cerebrospinal fluid (CSF), and spinal CSF (Wickramasinghe et al., 16 Sep 2025). In liver transporter studies, it denotes the transformation of fluorescence measurements into quantitative estimates of exposure, biliary clearance, and transporter activity, specifically MRP2, in order to assess hepatic ischemia–reperfusion injury (Monti et al., 2023).

A recurring distinction is between global mechanistic PBPK and local or imaging-centric tracer kinetic modeling. The former enforces physiological consistency across organs and systemic clearance, whereas the latter often estimates local parameters such as influx $K_i$, volume of distribution $V_d$, or binding potentials from dynamic PET or SPECT (Farhadi et al., 23 Nov 2025). PBPK-based dosimetry frequently combines these levels, using local imaging data to calibrate mechanistic whole-body or organ-level models.

2. Mathematical formulation and dose metrics

The core mathematical structure of PBPK dosimetry is a set of coupled mass-balance ordinary differential equations. A standard compartmental expression for organ $i$ is

$Q_i$0

where $Q_i$1 is the amount of drug in organ $Q_i$2, $Q_i$3 is organ volume, $Q_i$4 is blood flow to organ $Q_i$5, $Q_i$6 is arterial blood concentration, $Q_i$7 is the tissue:blood partition coefficient, and $Q_i$8 is organ-specific clearance (Liu et al., 9 Feb 2026). In concentration form, a typical PBPK tissue equation is

$Q_i$9

with a corresponding blood compartment equation

$V_i$0

(Farhadi et al., 23 Nov 2025). For radiotheranostics, physical decay is added as

$V_i$1

where $V_i$2 (Zaid et al., 25 Sep 2025).

A generic PBPK organ compartment is also often written in amount form as

$V_i$3

or, in concentration form,

$V_i$4

(Hardiansyah et al., 11 Sep 2025). When receptor-mediated uptake is relevant, additional states are introduced, for example

$V_i$5

(Farhadi et al., 23 Nov 2025).

The principal dosimetric outputs are derived by integration or extrema of these trajectories. In pharmacokinetic applications, these include

$V_i$6

(Liu et al., 9 Feb 2026). In radiopharmaceutical dosimetry, PBPK-predicted TACs $V_i$7 are converted to time-integrated activity

$V_i$8

and then to absorbed dose

$V_i$9

or, in voxel notation,

$K_{p,i}$0

(Farhadi et al., 23 Nov 2025). The same relationship is expressed in TIAC notation as

$K_{p,i}$1

with TIACs often defined as $K_{p,i}$2 (Zaid et al., 25 Sep 2025).

In recent PSMA-targeted RPT work, physical dose rate and absorbed dose are written as

$K_{p,i}$3

and

$K_{p,i}$4

while biological dosimetry uses the linear–quadratic framework with the Lea–Catcheside factor $K_{p,i}$5,

$K_{p,i}$6

(Abdollahi et al., 23 Oct 2025). In brain PBPK, the principal exposure metrics are defined as

$K_{p,i}$7

(Wickramasinghe et al., 16 Sep 2025).

This mathematical commonality is one of the defining features of PBPK-based dosimetry: despite domain-specific variations, the task is consistently the transformation of a mechanistic dynamical system into internal dose descriptors.

3. Model structures across application domains

PBPK-based dosimetry is not a single model class but a family of mechanistic representations adapted to route, modality, and scale.

In classical systemic pharmacokinetics, the body is modeled as a network of tissue compartments connected by blood flow. Even simplified structures can support dosimetric analysis. A recent scientific machine learning framework generated synthetic data from a standard 2-compartment PK model,

$K_{p,i}$8

with parameters $K_{p,i}$9, $C_{\max}$0, $C_{\max}$1, and $C_{\max}$2, and used randomized log-normal parameter distributions to simulate inter-individual variability in AUC and $C_{\max}$3 (Liu et al., 9 Feb 2026).

In radiopharmaceutical therapy, model structures are considerably richer. A whole-body [$C_{\max}$4Lu]Lu-DOTATATE PBPK model is described as including tumors, kidneys, GI tract, bone, red marrow, brain, heart, liver, lungs, muscle, skin, spleen, prostate, adrenal glands, and a rest-of-body compartment (Zaid et al., 25 Sep 2025). A PSMA-targeted model implemented in MATLAB SimBiology covers tumor, kidneys, salivary glands, liver, spleen, lungs, red marrow and other bone, GI tract, muscle, prostate, brain, adipose, skin, heart, veins and arteries, plus a peptide–protein serum compartment for albumin-bound ligand. Each organ is subdivided into vascular space, interstitial space, receptor-bound space, and intracellular space, with organ-specific parameters such as blood flow $C_{\max}$5, organ volume $C_{\max}$6, permeability-surface area product $C_{\max}$7, receptor density $C_{\max}$8, glomerular filtration rate, and tubular extraction ratio (Abdollahi et al., 23 Oct 2025).

For dynamic PET with FDG, an irreversible 2-tissue compartment model is used voxel-wise. The states are free FDG $C_{\max}$9, bound FDG $C_{\max}$0, and the total tissue concentration $C_{\max}$1, governed by

$C_{\max}$2

and a measured voxel activity concentration

$C_{\max}$3

to account for unresolved blood volume fraction $C_{\max}$4 (Benetti et al., 2023). Here the dosimetric object is a voxel TAC, not merely an organ average.

Brain PBPK adds barrier-specific physiology. A permeability-limited 4-compartment brain model includes brain blood $C_{\max}$5, brain mass $C_{\max}$6, cranial CSF $C_{\max}$7, and spinal CSF $C_{\max}$8, with arterial concentration $C_{\max}$9 as an exogenous input. Volumes $T_{\max}$0, $T_{\max}$1, $T_{\max}$2, and $T_{\max}$3; flows $T_{\max}$4, $T_{\max}$5, $T_{\max}$6, $T_{\max}$7, $T_{\max}$8, and $T_{\max}$9; permeability-surface area products $K_i$0, $K_i$1, and $K_i$2; active transport clearances $K_i$3, $K_i$4, $K_i$5, $K_i$6, and $K_i$7; and unbound and unionized fractions all modulate brain and CSF exposure (Wickramasinghe et al., 16 Sep 2025).

Dermal PBPK replaces organ networks with layered skin geometry. Three major categories are distinguished: QSPR/QSAR models, compartmental PBPK models, and discretised-microstructure PBPK (DM-PBPK) models (Sebastia-Saez et al., 2018). Compartmental approaches treat each skin layer as a well-mixed compartment with diffusion-type ODEs such as

$K_i$8

while DM-PBPK solves diffusion or diffusion–convection PDEs in explicit 2D or 3D representations of vehicle, stratum corneum, viable epidermis, dermis, hair follicles, and pilosebaceous units (Sebastia-Saez et al., 2018).

Liver-centric PBPK can narrow the scope further. In sodium fluorescein modeling for ischemia–reperfusion injury, the compartments are liver blood, hepatocytes, and bile, with systemic circulation represented through empirical inflow functions. SF and its metabolite SF-glucuronide are transported between blood and hepatocytes, converted by glucuronidation, and effluxed into bile via MRP2-mediated transport (Monti et al., 2023).

These examples show that PBPK-based dosimetry is defined less by one canonical architecture than by a modeling principle: explicit physiological structure linked to measurable exposure endpoints.

4. From concentration–time curves to organ, voxel, and biological dose

The most direct output of a PBPK model is a time course. Dosimetry begins when this time course is converted into cumulative or biologically weighted exposure.

In organ-level pharmacokinetics, this is often sufficient in itself. The concentration–time profiles in plasma and tissues determine AUC, $K_i$9, organ AUCs, half-life, and related metrics (Liu et al., 9 Feb 2026, Wickramasinghe et al., 16 Sep 2025). For first-in-human translation, environmental risk, or toxicology, such metrics may already constitute the relevant dose surrogates.

In nuclear medicine, TACs are the central intermediate object. PBPK models produce organ- and lesion-specific TACs $V_d$0 for target tissues and organs at risk such as kidneys, salivary glands, and bone marrow (Farhadi et al., 23 Nov 2025). These TACs support metronomic injection schedule optimization, therapy planning, and adaptive dosing by allowing one to vary injection profiles, uptake rates, retention, and clearance (Farhadi et al., 23 Nov 2025). The same framework underlies single-time-point and sparse-time-point dosimetry, where PBPK compensates for limited imaging and can estimate TIACs and absorbed doses from reduced data (Zaid et al., 25 Sep 2025, Hardiansyah et al., 11 Sep 2025).

A major recent transition is from organ-level to voxel-level dosimetry. The field has moved from organ-level, population-based dosimetry using simplified kinetics toward patient-specific, voxel-based dosimetry driven by quantitative imaging, Monte Carlo, and PBPK (Farhadi et al., 23 Nov 2025). In dynamic PET, voxel-wise TACs can be estimated by a self-supervised spatio-temporal UNet that outputs parametric images of $V_d$1, $V_d$2, $V_d$3, and $V_d$4 and reconstructs each voxel TAC through the kinetic forward model (Benetti et al., 2023). Once $V_d$5 is known, voxel-level cumulated activity can be computed as

$V_d$6

or, when physical decay is included,

$V_d$7

with organ-level cumulated activity obtained by summing over voxels (Benetti et al., 2023). Voxel absorbed dose then follows from voxel S-values or dose kernels,

$V_d$8

(Benetti et al., 2023).

The same organ-to-voxel progression appears in RPT planning. A PBPK-informed conditional GAN has been reported to predict voxel-wise post-therapy dosimetry from pre-therapy PSMA PET using simulated $V_d$9Ga-PSMA-11 PET, segmentation masks, and organ-specific PBPK-derived dose ranges. The model is constrained by PBPK priors to generate biologically plausible 3D dose distributions for $i$0Lu-PSMA-I&T therapy, with improved DVH accuracy and reduced organ-level errors relative to conventional methods (Farhadi et al., 23 Nov 2025). This suggests a hybrid regime in which PBPK does not directly calculate every voxel dose, but constrains a learned surrogate operating at image resolution.

Biological dosimetry adds a further transformation. In the PSMA virtual theranostic trial framework, PBPK-generated TACs yield physical dose, which is then converted to BED and EQD2 using organ-specific $i$1 ratios and repair rates $i$2, with the Lea–Catcheside factor accounting for protracted dose delivery (Abdollahi et al., 23 Oct 2025). This formalism makes the dose-rate shape of the TAC, not merely its integral, dosimetrically relevant.

5. Calibration, inference, and computational acceleration

PBPK-based dosimetry has long been limited by calibration difficulty, identifiability, and computational cost. Several recent developments address these constraints.

Classical and likelihood-based calibration

Conventional PBPK calibration uses dynamic PET/SPECT TACs, blood sampling, and prior distributions for flows, volumes, partition coefficients, and binding parameters (Farhadi et al., 23 Nov 2025). Goodness-of-fit is assessed graphically and quantitatively. A recent VVUQ review emphasizes fitted TACs versus observed data, observed-versus-predicted plots, residual diagnostics, and metrics such as SSE,

$i$3

MSE,

$i$4

$i$5,

$i$6

coefficients of variation for estimated parameters, and correlation matrices to identify parameter dependency and poor identifiability (Zaid et al., 25 Sep 2025). Model selection is supported by AICc,

$i$7

with derived $i$8 and Akaike weights $i$9 (Zaid et al., 25 Sep 2025).

In radiopharmaceutical therapy, hybrid PBPK–PopPK approaches use nonlinear mixed-effects models to stabilize individual estimates when imaging is sparse. Individual observations $Q_i$00 are modeled as

$Q_i$01

with population structure

$Q_i$02

or a general covariate-dependent mapping $Q_i$03 (Hardiansyah et al., 11 Sep 2025). This combination is particularly relevant for single-time-point dosimetry and sparse-data RPT settings (Hardiansyah et al., 11 Sep 2025, Zaid et al., 25 Sep 2025).

Self-supervised and inverse neural inference

Dynamic PET has motivated self-supervised parameter estimation. A spatio-temporal UNet takes dynamic PET slices as input and outputs voxel-wise $Q_i$04, $Q_i$05, $Q_i$06, and $Q_i$07 maps. No ground-truth parameter maps are required; instead, the loss enforces agreement between measured TACs and TACs reconstructed through the embedded kinetic forward model,

$Q_i$08

(Benetti et al., 2023). Parameter ranges are explicitly constrained to physiologically plausible intervals, a feature of direct dosimetric relevance because implausible parameters distort residence times and dose (Benetti et al., 2023).

Inverse physics-informed neural networks extend this logic to PBPK ODE systems. In PBPK-iPINN, a neural network $Q_i$09 approximates concentration profiles in four brain compartments while simultaneously estimating selected PBPK parameters. Training minimizes a composite loss,

$Q_i$10

where data loss matches observed concentrations, ODE loss enforces the brain PBPK system, and IC loss enforces initial conditions (Wickramasinghe et al., 16 Sep 2025). The study reports that correct convergence requires careful weighting of the loss components and tuning of network depth, width, activation, learning rate, optimizer, and collocation points (Wickramasinghe et al., 16 Sep 2025). The same work compares PBPK-iPINN to SAEM and differential evolution and reports very small absolute errors for recovered volume parameters, with somewhat larger but still small deviations for $Q_i$11 and $Q_i$12 (Wickramasinghe et al., 16 Sep 2025). This suggests that inverse neural solvers can recover dosimetrically relevant parameters when traditional estimation is difficult.

Surrogates and foundation models

A broader scientific machine learning direction treats PK or PBPK trajectories as sequence data. A “Foundation PBPK Transformer” uses the first $Q_i$13 observed points

$Q_i$14

to forecast the remainder of the profile

$Q_i$15

with standard self-attention

$Q_i$16

and sinusoidal positional encodings (Liu et al., 9 Feb 2026). On synthetic datasets, the model reconstructs an entire PK curve from the first 5 time points, corresponding to 10% of the trajectory, with MSE $Q_i$17 (Liu et al., 9 Feb 2026). The same framework introduces Physiologically Constrained Diffusion Models (PCDM) to synthesize virtual physiologies under explicit constraint penalties, and “Neural Allometry” combining graph neural networks with Neural ODEs for cross-species PK scaling (Liu et al., 9 Feb 2026). These methods are positioned as ways to accelerate simulation, reduce physiological violations, and improve translational prediction.

A key quantitative result is that PCDM reduces a physiological violation rate from 2.00% to 0.50% under constraints (Liu et al., 9 Feb 2026). This matters for dosimetry because unphysiological organ volumes or flow combinations can induce implausible exposure profiles. The same paper reports human leave-one-species-out PK prediction with test MSE $Q_i$18 using Rat + Dog training and Human testing in the Neural Allometry setup (Liu et al., 9 Feb 2026).

6. Virtual populations, digital twins, and uncertainty

PBPK-based dosimetry increasingly operates on populations rather than isolated patients. Virtual cohorts, sensitivity analysis, and uncertainty propagation are central because dosimetric predictions are only as credible as their underlying physiological and pharmacological assumptions.

Constrained generative models provide one route to virtual population synthesis. In PCDM, the forward diffusion process is

$Q_i$19

with a learned reverse model trained under a DDPM loss plus a physiological penalty

$Q_i$20

where a sample constraint can be

$Q_i$21

(Liu et al., 9 Feb 2026). This produces virtual patients whose organ volumes are consistent with specified mass-fraction constraints.

Digital twin formulations go further by embedding PBPK dosimetry in a bidirectional clinical-model feedback loop. Theranostic digital twins (TDTs) are defined as a physical patient, a virtual representation, and a bi-directional feedback loop linking clinical data to model updates and model outputs to treatment decisions (Zaid et al., 25 Sep 2025). PBPK is the PK engine of the TDT, producing TACs, TIACs, absorbed doses, and inputs to TCP/NTCP or BED models (Zaid et al., 25 Sep 2025). This framework supports adaptive planning across cycles, in which first-cycle data are used for internal validation and later-cycle predictions are externally evaluated against observed doses and TACs (Zaid et al., 25 Sep 2025).

A concrete PSMA-targeted implementation generated 640 virtual patients and simulated 15,360 TACs across tumors and five organs under multiple imaging and therapy tracers (Abdollahi et al., 23 Oct 2025). Tumor volume, tumor heterogeneity, PSMA receptor density, TER, GFR, body height, and body surface area were varied, and a 20% Gaussian perturbation was applied to kinetic parameters:

$Q_i$22

(Abdollahi et al., 23 Oct 2025). Machine learning models trained on PET TAC-derived features then predicted physical and biological dosimetry endpoints with MAPE-based evaluation,

$Q_i$23

(Abdollahi et al., 23 Oct 2025). Cu-64-based imaging yielded the most robust predictions, with dose prediction MAPE as low as 8% for tumors and 10–20% for different organs, whereas F-18 was more volume-dependent and Ga-68 exhibited higher variability (Abdollahi et al., 23 Oct 2025).

Uncertainty quantification itself is becoming formalized. A recent review organizes PBPK VVUQ for theranostic digital twins around verification, validation, and uncertainty quantification (Zaid et al., 25 Sep 2025). Uncertainty propagation assigns distributions to parameters, samples the joint space with Latin hypercube or Monte Carlo methods, and computes TIAC, AD, and BED distributions. Accuracy relative to a reference can be summarized using relative deviation,

$Q_i$24

RMSE,

$Q_i$25

and MAPE (Zaid et al., 25 Sep 2025). This makes PBPK-based dosimetry not merely a point-estimate procedure but a probabilistic prediction task.

7. Validation, interpretability, and limitations

PBPK-based dosimetry derives much of its appeal from mechanistic structure, but this does not eliminate the need for rigorous credibility assessment. The literature repeatedly emphasizes that verification, validation, and uncertainty quantification are indispensable for clinical or regulatory use (Zaid et al., 25 Sep 2025).

Verification asks whether the model has been implemented correctly: compartments, connections, units, mass balance, parameter definitions, limiting cases, and solver stability must all be checked (Zaid et al., 25 Sep 2025). Validation asks whether the model is adequate for its intended purpose. Internal validation examines goodness-of-fit to calibration data; external validation tests generalizability to unseen patients, cycles, or imaging conditions (Zaid et al., 25 Sep 2025). In nuclear medicine, voxel-level validation may involve comparing predicted dose maps against Monte Carlo reference, assessing DVH agreement, and measuring clinical impact when dose constraints are mispredicted (Farhadi et al., 23 Nov 2025).

Clinical oversight remains a central theme. Integrated clinical-computational nuclear medicine emphasizes “physician-in-the-loop” use of PBPK-based dosimetry, because segmentation, priors, generative models, and learned surrogates can fail in subtle ways (Farhadi et al., 23 Nov 2025). The RELAINCE framework is recommended for evaluation of AI and modeling tools, focusing on Reliability, Lack of bias, Interpretability, Non-inferiority vs standard methods, and Prospective evaluation (Farhadi et al., 23 Nov 2025).

Interpretability is an active concern in PBPK–ML hybrids. In the PSMA digital twin study, SHAP analysis on tree-based regressors identified AMax, Clearance, $Q_i$26, $Q_i$27, $Q_i$28, Skewness, and P25 as influential TAC-derived features, with importance varying by organ, endpoint, and tumor volume (Abdollahi et al., 23 Oct 2025). This does not replace mechanistic interpretability, but it localizes which kinetic signatures drive surrogate predictions.

Several limitations recur across domains.

Data scarcity remains acute. Dynamic PET/SPECT datasets for PBPK calibration are scarce, and multi-timepoint post-therapy imaging is resource-intensive (Farhadi et al., 23 Nov 2025). This motivates sparse-sampling PBPK, PopPK regularization, and machine learning surrogates, but also constrains validation.

Model complexity is both a strength and a weakness. Robust PBPK models require numerous physiological, biochemical, and drug-specific parameters, many difficult to measure directly (Hardiansyah et al., 11 Sep 2025). Complex models are computationally demanding, and the dimensionality of parameter estimation can make identifiability poor (Zaid et al., 25 Sep 2025).

Scope limitations often remain. The 2026 SciML framework is explicitly proof-of-concept and validated on synthetic datasets rather than clinical or toxicological data; full multi-organ PBPK structures with explicit tissue compartments, partition coefficients, and enzyme kinetics are not yet embedded in its neural components (Liu et al., 9 Feb 2026). The PBPK-iPINN study models only brain compartments and uses systemic plasma input as an exogenous function rather than a full-body PBPK (Wickramasinghe et al., 16 Sep 2025). Liver fluorescein modeling is liver-centric rather than whole-body and embeds renal clearance empirically (Monti et al., 2023). The dermal DM-PBPK literature is described as early-stage with scarce validation data (Sebastia-Saez et al., 2018).

Regulatory qualification is still developing. PBPK is widely accepted in small-molecule drug development, but radiopharmaceutical-specific regulatory standards remain limited; transparent documentation, fit-for-purpose VVUQ, and professional society guidance are repeatedly identified as prerequisites for broader adoption (Zaid et al., 25 Sep 2025, Hardiansyah et al., 11 Sep 2025).

A common misconception is that mechanistic structure alone guarantees reliable dosimetry. The literature does not support this. Mechanistic models can still be misspecified, underidentified, poorly calibrated, or numerically unstable. A related misconception is that AI surrogates remove the need for physiology; the emerging consensus is the opposite. Recent work explicitly preserves mechanistic PBPK principles while adding data-driven flexibility, uses physiological constraints in generative models, and stresses clinician-guided evaluation for safe deployment (Liu et al., 9 Feb 2026, Farhadi et al., 23 Nov 2025).

8. Emerging directions

Several directions now define the research frontier of PBPK-based dosimetry.

One is multi-scale integration. The 2026 SciML framework links mechanistic ODE-based PBPK or 2-compartment models, Foundation PBPK Transformers, physiologically constrained generative models, and Neural Allometry for cross-species scaling (Liu et al., 9 Feb 2026). Nuclear medicine work similarly points toward linking PBPK with voxel-level microdosimetry, BED, microenvironmental effects, radiomics, genomics, and clinical informatics (Farhadi et al., 23 Nov 2025).

Another is hybrid PBPK–PopPK. Reviews of RPT personalization describe PBPK structure as the mechanistic scaffold and PopPK or NLMEM as the statistical layer that estimates population distributions, covariate effects, and individual posterior parameter values, especially under sparse imaging (Hardiansyah et al., 11 Sep 2025). This hybridization is particularly important for clinically feasible single-time-point dosimetry.

A third is predictive and adaptive dosimetry. PBPK-based digital twins aim to move RPT from empirical fixed dosing to simulation-driven activity selection, cycle timing, and adaptive replanning (Zaid et al., 25 Sep 2025, Abdollahi et al., 23 Oct 2025). In this setting, dosimetry is no longer retrospective but prospective: the model predicts future dose and toxicity under candidate regimens.

A fourth is microstructure-resolved dosimetry outside classical whole-body RPT. Dermal DM-PBPK replaces homogeneous barriers with explicit bricks-and-mortar stratum corneum, hair follicles, and pilosebaceous units, solved with finite element or lattice Boltzmann methods (Sebastia-Saez et al., 2018). This yields local concentration–time fields, penetration-pathway decomposition, and depth-dependent dose metrics unavailable to classic compartmental PBPK.

A fifth is inverse and self-supervised inference. Self-supervised TAC reconstruction in dynamic PET and inverse PINNs for brain PBPK suggest a future in which PBPK parameters and exposure metrics are inferred directly from image sequences or sparse concentration data while preserving the governing equations (Benetti et al., 2023, Wickramasinghe et al., 16 Sep 2025).

A plausible implication is that PBPK-based dosimetry is evolving from a simulation methodology into a computational ecosystem. In that ecosystem, mechanistic models generate or constrain trajectories, statistical layers capture inter-individual variability, neural surrogates accelerate forecasting, digital twins support treatment planning, and VVUQ frameworks determine whether the resulting dose estimates are credible enough for practical use.

PBPK-based dosimetry therefore occupies a distinctive position in quantitative biomedical modeling. It is neither purely empirical curve fitting nor purely first-principles simulation. Rather, it is a mechanistically structured inference problem whose outputs are internal dose metrics and whose relevance spans drug development, toxicology, imaging, radiopharmaceutical therapy, dermal exposure, organ-specific injury assessment, and personalized treatment planning. The contemporary literature indicates that its future lies in tighter integration of physiology, imaging, uncertainty quantification, and scientific machine learning, with credibility assessment and clinical oversight remaining central conditions for translation (Liu et al., 9 Feb 2026, Zaid et al., 25 Sep 2025, Farhadi et al., 23 Nov 2025).