Pharmacokinetic-pharmacodynamic modelling is a cornerstone of modern drug development, enabling the optimisation of safety profiles. In his latest commentary, Professor Alain Astier explores the role of exposure-response modelling in anticipating the toxicities of small-molecule drugs, outlines its current limitations, and considers future directions for model-informed drug development.
Adverse events such as cardiac toxicity remain leading causes of late-stage attrition and post-marketing withdrawal for drugs. Predicting toxic effects at early preclinical development stages is therefore both ethically sound and cost-effective.
Pharmacokinetic-pharmacodynamic (PK-PD) modelling plays a central role in anticipating and managing toxicity risks for small-molecule drugs across discovery and development. By quantitatively linking drug exposure to adverse outcomes, these models provide a structured framework for translating preclinical findings into clinical risk management.
Marketing authorisation agencies now actively encourage model-informed drug development, in which decisions are supported by mathematical models and simulations that predict the drug’s likely success, and embed exposure-response analyses into regulatory reviews.1
Where PK-PD modelling matters most
Mechanistically, PK-PD frameworks range from empirical concentration-effect models to physiologically based pharmacokinetic (PBPK) and multiscale systems models that estimate tissue exposure and downstream injury dynamics.
These approaches have proven particularly valuable in predicting proarrhythmic and haematologic toxicities, although their performance is limited when underlying mechanisms are multifactorial or idiosyncratic.
Cardiac repolarisation and corrected QT interval analysis
The best-established safety application of PK-PD modelling is in assessing proarrhythmic risk. Updates to the ICH E14 guideline permit concentration-corrected QT interval (C-QTc) modelling as the primary method to rule out clinically relevant QTc effects early, often replacing stand-alone thorough QT studies.2,3
Beyond simple linear C-QTc analysis, the Comprehensive in Vitro Proarrhythmia Assay combines multi-ion channel data, in silico ventricular models and early ECG biomarkers to differentiate actual torsade risk from benign QT prolongation.4–6
Traditional safety margins remain useful but are now supported by these integrated models.7
Haematologic toxicity
For cytotoxic and some targeted agents, semi-mechanistic models combining a proliferative compartment, a maturation ‘transit’ chain and feedback regulation, predict the time course of neutropenia and leukopenia from exposure data.
The canonical model developed by Friberg and colleagues has been evaluated across agents and drug schedules to predict and manage chemotherapy-induced myelosuppression and neutropenia in PK-PD studies.8,9
Organ-specific risks
PBPK models translate preclinical and in vitro data into human organ exposure, accounting for physiological differences, disease, age and drug-drug interactions.10,11
For acetaminophen, PBPK-PD and multiscale models link metabolic activation and spatial liver injury to exposure, helping define thresholds for intrinsic hepatotoxicity and evaluating sensitive populations such as pregnant, geriatric or paediatric patients.12–14
In the context of aminoglycoside nephrotoxicity, PK-PD and PK-toxicity models link metrics such as trough concentrations to renal injury risk, supporting extended-interval dosing strategies that mitigate toxicity while preserving efficacy.15
PK-PD in decision support and extrapolation
The most robust use of PK-PD modelling is therefore as decision support, coupled with transparent uncertainty analyses and focused clinical verification at relevant exposures and patient subpopulations.
Quantitative models enable ‘what if’ simulations exploring untested doses, regimens or special populations, and provide interpretable safety margins, such as exposure relative to in vitro potency.
Current C-QTc guidance explicitly supports extrapolation within observed concentration ranges to predict QT effects under alternative dosing conditions or potential drug-drug interactions, while emphasising the need for appropriate caution when extending predictions beyond the available data.2,3
Key limitations of PK-PD modelling
Despite its value, PK-PD modelling has limitations. Certain toxicities remain difficult to model mechanistically, extrapolating in vitro data to clinical outcomes introduces uncertainty, and both structural assumptions and lack of sufficient data can limit reliability. Moreover, variability across populations and the rarity of severe adverse events challenge the precision and external validity of model predictions.
Idiosyncratic and immune-mediated toxicities
Idiosyncratic drug-induced liver injury (iDILI) is a rare adverse reaction that develops independently of drug dose, route or duration of administration. It remains challenging to predict, as risk depends on host traits, including genetic, immunological and environmental factors that interact with exposure. Even sophisticated PK-PD or PBPK models cannot fully capture these multifactorial mechanisms.
Risk stratification is improving but not yet definitive, so these models are best used to contextualise exposure and test hypotheses rather than to predict IDILI directly.16,17
In vitro-in vivo gaps and assay variability
Translating in vitro potency or cell injury data to in vivo organ risk involves assumptions about unbound fractions, metabolism and transport. Variability in assay conditions can skew potency estimates and predicted safety margins. Mechanistic in vitro–in vivo frameworks are improving but continue to introduce uncertainty.10,11,13
Uncertainty in PK-PD model structure and parameters
Common models, including C-QTc and myelosuppression frameworks, rely on assumptions about structural form and covariate effects. Over-fitting within narrow concentration ranges can compromise extrapolation to higher exposures or drug-drug interactions.18–20 Best practice emphasises pre-specification, rigorous diagnostics and simulation-based sensitivity analyses.2,3
Generalisation across populations
Differences in physiology, comorbidity and co-medication alter exposure and susceptibility to toxicity. PBPK models can capture these effects, but their accuracy depends on the quality of the data. Predictions should therefore include uncertainty bounds and be prospectively verified in target populations.10,11
Rare event detection and statistical power
Early-phase PK-PD datasets are typically too small to detect rare but severe adverse events. Exposure-response methods can identify systematic shifts, such as QTc effects, but are not designed to predict one-in-a-thousand events. Models should therefore guide decision-making, not provide false certainty.2,3
Conclusion
PK-PD modelling has evolved from a specialist research tool into a practical safety engine supporting drug development from discovery to labelling. In small-molecule programmes, its capabilities translate diffuse safety signals into dose- and context-dependent predictions that can be tested and prospectively validated.
Looking ahead, the most significant impact is likely to come from integration rather than complexity. Future work should link PBPK-derived tissue exposures to mechanism-anchored biomarkers, incorporate assay quality-by-design principles and pre-specification analyses to address key development decisions, such as dose selection, scheduling or the need for PK boosting. Model assumptions, data provenance and code should be versioned and reviewable for full regulatory traceability.
Recent efforts integrating machine learning with PK inference, such as Deep-PK, aim to prioritise compounds and identify potential liabilities earlier while maintaining interpretability for regulatory acceptance.21
When applied with mechanistic rigour and an awareness of its limits, PK-PD modelling can continue to accelerate development, reduce unnecessary clinical exposure, and improve patient safety.
The value lies not in promising certainty but in making toxicity risks explicit, quantifiable and testable, so that benefit-risk decisions become faster, more transparent and more reproducible.
Author
Alain Astier PharmD PhD
Honorary head of the Department of Pharmacy, Henri Mondor University Hospital, and French Academy of Pharmacy, Paris, France
References
1 US Food and Drug Administration. Model-Informed Drug Development (MIDD) Paired Meeting Program. 2025;Feb 25.
2 US Food and Drug Administration. E14 Clinical Evaluation of QT/QTc Interval Prolongation and Proarrhythmic Potential. Guidance for Industry. 2024; May 31.
3 Garnett C et al. Scientific white paper on concentration-QTc modeling. J Pharmacokinet Pharmacodyn 2018;45(3):383–97.
4 Colatsky T et al. The Comprehensive in Vitro Proarrhythmia Assay (CiPA) initiative – Update on progress. J Pharmacol Toxicol Methods 2016;81:15–20.
5 Li Z, Ridder BJ, Han X, et al. Assessment of an in silico mechanistic model for proarrhythmia risk prediction under the CiPA initiative. Clin Pharmacol Ther 2019;105(2):466–75.
6 Vicente J et al. Assessment of multi-ion channel block in a Phase I randomised study design: results of the CiPA Phase I ECG biomarker validation study. Clin Pharmacol Ther 2019;105(4):943–53.
7 Redfern WS et al. Relationships between preclinical cardiac electrophysiology, clinical QT interval prolongation and torsade de pointes: evidence for a provisional safety margin in drug development. Cardiovasc Res 2003;58(1):32–45.
8 Friberg LE et al. Model of chemotherapy-induced myelosuppression with parameter consistency across drugs. J Clin Oncol 2002;20(24):4713–21.
9 Soto E et al. Predictive ability of a semi-mechanistic model for neutropenia in the development of novel anti-cancer agents: two case studies. Invest New Drugs 2011;29(5):984–95.
10 Isoherranen N. Physiologically based pharmacokinetic modeling of small molecules: How much progress have we made? Drug Metab Dispos 2025;53(1):100013.
11 Fairman K et al. Physiologically based pharmacokinetic modeling in translational research and regulatory toxicology. Curr Opin Toxicol 2020;23-24:17–22.
12 Jiang XL et al. Application of physiologically based pharmacokinetic modeling to predict acetaminophen metabolism and pharmacokinetics in children. CPT Pharmacometrics Syst Pharmacol 2013;2(10):e80.
13 Dichamp J et al. In vitro to in vivo acetaminophen hepatotoxicity extrapolation using classical schemes, pharmacodynamic models and a multiscale spatial-temporal liver twin. Front Bioeng Biotechnol 2023;11:1049564.
14 Mian P et al. Population pharmacokinetic modeling of intravenous paracetamol in fit older people displays extensive unexplained variability. Br J Clin Pharmacol 2019;85(1):126–35.
15 Rougier F et al Aminoglycoside nephrotoxicity: modeling, simulation, and control. Antimicrob Agents Chemother 2003;47(3):1010–16.
16 Fontana RJ et al. The Evolving Profile of Idiosyncratic Drug-Induced Liver Injury. Clin Gastroenterol Hepatol 2023;21(8):2088–99.
17 Björnsson HK, Björnsson ES. Review of human risk factors for idiosyncratic drug-induced liver injury: latest advances and future goals. Expert Opin Drug Metab Toxicol 2023;19(12):969–77.
18 Ji Y, Johannesen L, Garnett C. FDA’s insights: implementing new strategies for evaluating drug-induced QTc prolongation. J Pharmacokinet Pharmacodyn 2025;52(4):37.
19 Sager PT et al. Rechanneling the cardiac proarrhythmia safety paradigm. Am Heart J 2014;167(3):292–300.
20 Park JS, Polak S, Wisniowska B. Introduction to in silico models for proarrhythmic risk assessment under the CiPA initiative. Transl Clin Pharmacol 2019;27(1):12–18.
21 Myung Y, de Sá AGC, Ascher DB. Deep-PK: deep learning for small molecule pharmacokinetic and toxicity prediction. Nucleic Acids Res 2024;52(W1):W46–W475.
