Authors: – Srinivasulu Meruva Pallampati (Sr. Project Manager, Scientific Products)
In brief: the translational gap in drug development — the failure of preclinical models to predict human outcomes, which causes over 25% of drugs entering clinical development to fail — is being narrowed by a combination of PBPK and QSP computational models, organ-on-chip and iPSC biological systems, and AI-augmented pharmacometrics, supported by regulatory frameworks including FDA Modernization Act 2.0 and the newly finalized ICH M15 guideline. The rest of this article explains the mechanisms, the evidence, and what the data layer must supply for these models to deliver on their promise.
Even with decades of progress in drug discovery, over 25% of drugs that reach clinical development still fail. This is mostly because pre-clinical models do not accurately predict outcomes in humans [1]. This ongoing “translational gap” is one of the costliest and most frustrating challenges in pharmaceutical R&D.
The good news? A powerful combination of quantitative modeling and advanced biological systems is finally helping researchers close the gap. At recent events such as the American Conference on Pharmacometrics (ACoP), this progress has been clear. Model-informed methods are not just improving predictions; they are also delivering real results: about 10 months shorter development times, around $5 million saved per program, and almost twice the success rate in establishing proof-of-mechanism [2].
Traditional pre-clinical testing, such as 2D cell cultures, animal studies, and basic PK models, often misses important parts of human biology. These include complex disease processes, differences between individuals, and how organs interact. This leads to high failure rates in Phase 2 and 3 trials, wasted resources, and slower access to new treatments for patients. To solve this, we need a new set of models that work together. These should include both computer-based quantitative models and biological models that closely reflect humans, ideally brought together using digital twins of human physiology.
The intersection of quantitative modeling and clinical decision-making is explored in detail in Excelra’s blog on Exploring New Frontiers: Clinical Pharmacology in Modern Drug Development — providing essential context on how PBPK, PopPK, and exposure-response analysis are being integrated into regulatory submissions and trial design decisions across the drug development lifecycle.
The translational gap: Why Pre-clinical models still fail
Standard pre-clinical methods break down for predictable structural reasons. Animal models are quite different from humans in their metabolic enzymes, receptor pharmacology, and immune systems. Two-dimensional cell cultures remove the mechanical forces, three-dimensional cell-cell contacts, and oxygen gradients that shape the behavior of real tissues. Simple pharmacokinetic models lack sufficient detail to capture differences between patients or the interactions among organs. All of this leads to a Phase 2 attrition problem. This is the stage when the full cost of a clinical program is committed, and most efficacy failures happen.
A review of published quantitative systems pharmacology models showed the extent of the progress still needed. Only 24% of models had documented practices for robustness and reproducibility, and just 41% had been properly tested against pre-clinical or clinical data [3]. These numbers clearly show why standardization is needed. The models are available, but the systems to validate them are not yet in place.
The 24% reproducibility figure is not a modeling failure — it is a data failure. Models break at the input level, not the equation level. Excelra’s blog on Significance of Quantitative Tools and Techniques in Reducing Drug Attrition examines this data quality challenge directly — explaining how the integrity of PK, PD, and biomarker inputs determines whether a quantitative model informs a regulatory submission or fails peer review.
Taxonomy of models for quantitative Decision-Making
Two parallel tracks of models are restructuring how preclinical-to-clinical prediction works.
Track A: Quantitative / Pharmacometric models (In Silico)
These mathematical frameworks characterize drug behavior across populations, time, and physiological states:
- PBPK (Physiologically Based Pharmacokinetic): Mechanistic models that simulate absorption, distribution, metabolism, and excretion across virtual populations.
- Population PK (PopPK): Characterize variability and benchmark against historical data.
- QSP (Quantitative Systems Pharmacology): Multiscale models linking drug exposure to pharmacodynamic effects and disease progression.
- MBMA (Model-Based Meta-Analysis): Optimize trial design, dosing regimens, and patient selection.
- Disease Progression Modeling: Describes the natural history of disease and how therapeutic intervention alters its course over time.
Together, these tools excel at dose optimization, exposure prediction, and de-risking go/no-go decisions.
The role of QSP specifically in target selection and validation — one of the earliest stages where quantitative modeling can reduce attrition — is examined in Excelra’s blog on Improving Target Selection & Validation: Role of QSP. For teams building MBMA capabilities, Excelra’s dedicated blog on Data Curation for MBMA — an Integral Part of MID3 explains how structured literature extraction and data standardization directly determine MBMA model reliability.
Track B: Biological Pre-Clinical models (In Vitro / Ex Vivo)
Biological model complexity has advanced markedly over the past decade, progressing from simple 2D cell cultures to 3D spheroids, organoids, organ-on-chip microfluidic systems, and iPSC (induced pluripotent stem cell)-derived tissue models. These systems better recapitulate human tissue architecture, cell-cell interactions, and disease phenotypes than traditional 2D cultures, providing critical data on target engagement, mechanism of action, efficacy, and toxicity.
Track A answers dose and exposure questions. Track B answers biology and mechanism questions. The frontier is where they converge.
Where the Real Power Emerges: Integration and Digital Twins
The most exciting new development is digital twins. These are multiscale biophysical models that use real data from organoids and organ-on-chip systems to simulate virtual patient groups before any human receives a dose. Today, two clear examples show how valuable this integration can be.
PBPK for special populations
For pediatric patients, those with hepatic impairment, and pregnant individuals, mechanistic ADME models can predict safe and effective dosing without a dedicated clinical trial for each group. The FDA’s 2018 PBPK guidance formally recognized this ability [4], and regulators now accept PBPK analyses to waive some clinical pharmacology studies when appropriate. From 2020 to 2024, more than 26% of FDA-approved new drugs included PBPK models as key regulatory evidence in NDAs and BLAs. This number has increased each year.
Excelra’s case study on PK/PD Modeling for Dose Regulation demonstrates how exposure-response modeling and PBPK-informed dose selection translated into a concrete regulatory outcome for a clinical program — illustrating the practical pathway from quantitative model to NDA submission evidence.
QSP in rare disease
When patient populations are too small for a traditional pharmacokinetic study, QSP models rely on biological rather than statistical methods. This is especially important in rare disease programs, where even a Phase 2 trial might have fewer than 30 patients. Target-mediated drug disposition (TMDD) modeling for biologics is the standard here. Mechanistic models of drug-target binding kinetics explain the nonlinear PK behavior that happens when drug concentrations get close to target concentrations. They do this without needing the large datasets that empirical methods require. The ICH M15 guideline, finalized in January 2026, sets the regulatory framework for documenting and submitting this kind of model-based evidence in all major regions [5].
Excelra’s Intelligent Quantitative Systems Pharmacology (iQSP) service is specifically designed to support the kind of mechanistic modeling described here. For an overview of how iQSP integrates target biology, dose selection, and rare disease modeling into regulatory-ready evidence packages, see our iQSP service page.
Engineering challenges and honest bottlenecks
Organ-on-chip technology has advanced rapidly, with endothelial-lined channels, perfusable vasculature, and the ability to apply mechanical strain that more closely mimics the physiological environment of living tissue. Yet significant gaps remain:
- Building a fully functional, long-lasting perfusable blood vessel network
- Accurately recreating a strong and reliable blood-brain barrier
The EMA’s 3Rs Working Party is working on context-of-use criteria for organ-on-chip data in regulatory submissions. This includes planned annexes on regulatory acceptance criteria for microphysiological systems in specific situations [6]. This process recognizes both the potential of these systems and the need for more validation. Sponsors who want to submit organ-on-chip data should talk to EMA’s Innovation Task Force early to get scientific advice.
For iPSC technology, cell maturity is still a major technical challenge. iPSC-derived heart cells now show strong functional maturity, but iPSC-derived liver cells still struggle with full metabolic zonation and stability over time. This limits their reliability for predicting liver toxicity. These challenges remind us what is truly ready for regulatory decisions today.
Regulatory tailwinds and the rise of MID3
The regulatory environment is increasingly aligned with the adoption of non-animal, model-informed approaches.
On December 29, 2022, President Biden signed the FDA Modernization Act 2.0 into law [7]. The Act replaced the longstanding requirement that animal testing precede human clinical trials. Now, non-animal methods like organ-on-chip systems, iPSC-based cell assays, and in silico models are officially allowed for IND applications. In April 2025, the FDA shared a plan to phase out routine animal testing and released a roadmap for moving to human-relevant alternatives, with monoclonal antibodies as a key focus.
In Europe, the EMA’s 3Rs Working Party released a concept paper in December 2023. It describes a step-by-step update to its regulatory guidelines, adding annexes with specific criteria for organ-on-chip and other microphysiological systems [6]. The framework is still being developed, but the direction is clear.
ICH M15 and evolving MIDD guidance continue to mature [5]. The M15 guideline establishes a harmonized assessment framework for MIDD evidence across ICH regions, covering planning, evaluation, and documentation of model-informed submissions. Its finalization in early 2026 marks a significant step toward consistent global regulatory treatment of quantitative modeling evidence.
MID3 (Model-Informed Drug Discovery and Development) is the next step forward. It brings quantitative modeling into the early discovery phase, not just late-stage development. While MIDD focused on decisions during development, MID3 treats modeling as an ongoing process from target identification to post-market surveillance.
The data infrastructure required to support MID3 is explored in detail in Excelra’s blog on Data Curation for MBMA — an Integral Part of MID3 — examining how structured literature extraction, data standardization, and analysis-ready dataset construction are foundational requirements for MID3 implementation, and why the 24% reproducibility figure cited earlier reflects a data problem more than a modeling problem.
The AI opportunity with guardrails
Machine learning is being layered onto traditional pharmacometric models to augment their capabilities in three specific areas:
- Covariate selection: finding which patient characteristics best predict how they will respond to a drug.
- Virtual patient generation: turning small clinical datasets into larger, statistically representative virtual populations for simulation.
- Surrogate modeling: using faster machine learning models instead of complex QSP simulations for sensitivity analyses and scenario testing.
The difference between interpretable ML and black-box deep learning is not just academic; it is a regulatory issue. Regulators require models to be explainable. Tools like SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) are quickly becoming standard for showing how a model makes its predictions [8]. Black-box AI predictions cannot be used for important regulatory decisions unless there are clear, context-specific frameworks. There is real interest in using AI for MIDD, but there are also real limits on how it can be used.
ACoP in action: Two case studies
The American Conference on Pharmacometrics offers the clearest window into how integrated modeling is being applied at the bench-to-bedside interface. Two presentations from recent ACoP meetings illustrate the translational impact directly.
Case Study 1: TMDD modeling for biologic dose decision support (ACoP14, 2023)
At ACoP14, Corey J. Bishop introduced an interactive application based on Target-Mediated Drug Disposition (TMDD) to make pharmacokinetic simulation easier and help guide drug development for high-affinity biologics [9]. The tool showed how mechanistic binding models let teams quickly explore dose-response situations where standard two-compartment PK models do not work well. This is especially true for the nonlinear effects seen when drug concentrations get close to receptor saturation, which often happens in rare disease programs with limited clinical data. This decision-support platform shows QSP being used as the main evidence for dose selection, rather than just as an extra analysis after the fact.
Case Study 2: Mechanistic translational modeling for First-in-Human dose in T-Cell engagers (ACoP15, 2024)
At ACoP15, a mechanistic translational modeling framework was introduced to predict first-in-human doses for T-cell engager therapies. These are a type of bispecific antibody where the risk of cytokine release syndrome makes moving from pre-clinical to clinical studies especially important [10]. The study compared this mechanistic approach with the standard Minimal Anticipated Biological Effect Level (MABEL) method, using the bispecific HPN536 as an example. By combining pre-clinical pharmacodynamic data with quantitative binding kinetics, the mechanistic model improved the accuracy of first-in-human dose predictions and gave a clear biological explanation that MABEL alone cannot provide. This approach is helping to close the translational gap in real time, not by collecting more animal data, but by using better mechanistic understanding in a validated quantitative framework.
The road ahead
The translational gap is narrowing. The evidence appears in the regulatory calendar: FDA Modernization Act 2.0 signed, ICH M15 finalized, EMA’s organ-on-chip criteria actively developed. It also shows in conference halls, where ACoP presentations on QSP-informed first-in-human decisions and PBPK-supported pediatric dosing now occupy prime slots once held by classical PK/PD. The direction is set.
The next phase demands not more models but better inputs for existing ones. A PBPK simulation is mechanistically valid only if its ADME parameters are. A QSP model is predictive only if its biological calibration data is curated with the precision the model expects. An MBMA analysis synthesizes literature, but someone must extract, standardize, and quality-control that literature first. The 24% reproducibility figure is not a modeling failure but a data failure. Models do not break at the equation level; they break at the input level.
Organizations that close the translational gap fastest will not necessarily have the most sophisticated models. They will be those who feed models clean, structured, analysis-ready data — PK, PD, biomarker, and clinical endpoint data built for the specific question. This work makes the difference between a model that informs a regulatory submission and one that fails peer review.
This is the problem Excelra was built to solve. If your quantitative modeling programs are ready to move, the conversation to have is about data.
The translational gap, as defined in this article, is the systematic failure of pre-clinical models — 2D cell cultures, animal studies, and simple PK models — to predict human clinical outcomes, resulting in over 25% of drugs entering clinical development failing despite successful preclinical results. The approach to closing it requires integrating Track A quantitative models (PBPK, PopPK, QSP, MBMA) with Track B biological models (organoids, organ-on-chip, iPSC systems) and ensuring that all models are fed analysis-ready, curated data. As this article concludes: models do not break at the equation level — they break at the input level.
Excelra’s Clinical Data Services provide the structured, analysis-ready PK, PD, biomarker, and clinical endpoint data that quantitative models require. For an overview of how Excelra builds the data layer that model-informed drug development depends on, visit our Clinical Data Services page.
References
[1] https://www.nature.com/articles/nrd4609
[2] https://ascpt.onlinelibrary.wiley.com/doi/10.1002/cpt.3636
[3] https://pmc.ncbi.nlm.nih.gov/articles/PMC6482280
[5] https://database.ich.org/sites/default/files/ICH_M15_Step4_Final_Guideline_2026_0129.pdf
[7] https://onlinelibrary.wiley.com/doi/10.1111/aor.14503
[8] https://ascpt.onlinelibrary.wiley.com/doi/10.1002/psp4.70155
What is the translational gap in drug development?
The translational gap in drug development is the failure of preclinical models to accurately predict drug efficacy and safety in humans — resulting in drugs that succeed in animal studies and cell culture experiments failing when they enter human clinical trials. Over 25% of drugs entering clinical development fail, with the majority of late-stage failures driven by efficacy gaps identified only in Phase 2 or Phase 3 — the point at which the full cost of a clinical program has already been committed. The structural causes are predictable: animal models differ from humans in metabolic enzymes, receptor pharmacology, and immune biology; 2D cell cultures lack the mechanical forces and 3D architecture of living tissue; and simple PK models cannot capture patient variability or organ-organ interactions. Closing the translational gap requires integrating quantitative computational models with advanced biological systems that more accurately recapitulate human physiology.
What is MID3 and how does it differ from MIDD?
MIDD (Model-Informed Drug Development) is the FDA and ICH framework for using quantitative modeling and simulation — including PBPK, PopPK, QSP, and MBMA — to inform decisions during clinical drug development, particularly for dosing, trial design, and regulatory submissions. MID3 (Model-Informed Drug Discovery and Development) extends this concept upstream into the early discovery phase. While MIDD focused on decisions after a clinical candidate is identified, MID3 treats quantitative modeling as an ongoing process from target identification through post-market surveillance — a continuous thread connecting discovery biology, preclinical pharmacology, clinical development, and real-world evidence. The ICH M15 guideline, finalized in January 2026, provides the harmonized regulatory framework for documenting and submitting model-based evidence across ICH regions, establishing a consistent global standard for MID3 evidence packages.
How is PBPK being used to support FDA regulatory submissions?
Physiologically-Based Pharmacokinetic (PBPK) modeling is now formally accepted by the FDA as primary regulatory evidence in several context-of-use scenarios. The FDA’s 2018 PBPK guidance established the formal criteria for format and content of PBPK analyses in regulatory submissions. From 2020 to 2024, more than 26% of FDA-approved new drugs included PBPK models as key evidence in NDAs and BLAs — and this proportion has increased every year. The most common applications are dosing recommendations for special populations (pediatric patients, hepatic impairment, renal impairment, and pregnant individuals) and drug-drug interaction predictions — both scenarios where conducting dedicated clinical pharmacology studies for every population would be impractical. PBPK analyses accepted by the FDA can waive the need for some clinical studies, reducing development timelines and patient burden while providing regulators with mechanistically grounded dosing recommendations.
What are digital twins in drug development?
Digital twins in drug development are multiscale biophysical computational models that integrate real biological data — from organoids, organ-on-chip systems, and patient-derived iPSC models — to simulate how a specific drug will behave in a virtual patient or virtual patient population. Unlike purely statistical pharmacokinetic models, digital twins incorporate physiological structure: organ compartments, blood flow, metabolic enzyme activity, and disease-specific tissue changes. They allow researchers to explore dose-response scenarios, patient subgroup variability, and disease progression trajectories in silico before any human receives a dose. The key enabling development is the convergence of Track A quantitative models (PBPK, QSP) with Track B biological data from organoids and organ-on-chip systems — using the rich biological context those systems provide to calibrate and validate the computational models. This convergence is what transforms a standard pharmacokinetic model into a true digital twin of human physiology.
What is the FDA Modernization Act 2.0 and what does it mean for drug development?
The FDA Modernization Act 2.0, signed into law on December 29, 2022, replaced the longstanding legal requirement that animal testing must precede human clinical trials for IND applications. The Act explicitly permits the use of non-animal testing methods — including organ-on-chip microfluidic systems, iPSC-based cell assays, 3D organoids, and in silico computational models — as primary evidence for IND filing decisions. In April 2025, the FDA announced a roadmap for phasing out routine animal testing, with monoclonal antibody development identified as a key initial focus. For drug developers, this creates a legal and regulatory pathway to substitute organ-on-chip and computational data for some animal studies — potentially reducing development costs, shortening timelines, and improving translational relevance. However, the practical implementation requires sponsors to engage with the FDA early, demonstrate the validity of non-animal methods in their specific context of use, and produce data in formats that regulators have established acceptance criteria for.
Why does data quality determine quantitative model reliability in drug development?
Quantitative pharmacokinetic and pharmacodynamic models — including PBPK, QSP, and MBMA — are mechanistically structured: their outputs are mathematically derived from their inputs. This means that even a well-designed model with a rigorous validation framework will produce unreliable outputs if its input data is incomplete, inconsistent, or not curated to the precision the model expects. A PBPK simulation is only as valid as the ADME parameters it is built on. A QSP model is only as predictive as the biological calibration data used to tune it. An MBMA analysis synthesizes published literature, but the quality of that synthesis depends entirely on how rigorously the source data was extracted, standardized, and quality-controlled. A review of published QSP models found that only 24% had documented reproducibility practices and only 41% had been properly validated against external data — figures that reflect inadequate data infrastructure, not inadequate modeling methodology. The organizations that close the translational gap fastest are those that build analysis-ready, curated data foundations for their models.
Ready to Close Your Translational Gap with Analysis-Ready Data?
Excelra builds the curated, structured, analysis-ready PK, PD, biomarker, and clinical endpoint data that PBPK, QSP, and MBMA models depend on — the data layer that determines whether a quantitative model informs a regulatory submission or fails peer review. If your quantitative modeling programs are ready to move, the conversation to have is about data.
