top of page

Welcome to HMBC

Synergy, Not Substitution: How Organ-on-Chip and Digital NAMs Unite for Translational Impact

Christian Maass, PhD, Principal Scientist & Lead, MPSlabs, ESQLabs GmbHHenning Mann, PhD, Founder & CEO, HM.BioConsulting, former Sen. Dir Applied Sciences, Nortis, Inc.

 

Download the full article:


Overview

Over the past decade, Organ-on-Chip (OoC) platforms have matured into powerful microphysiological systems (MPS) capable of recapitulating dynamic human cellular microenvironments. Yet, widespread adoption in pharmaceutical and regulatory workflows has lagged, in many cases due largely to gaps in validation, reproducibility, and translation to whole-body contexts. (GAO, Nature).

Conversely, PBPK (Physiologically Based Pharmacokinetic) and QSP (Quantitative Systems Pharmacology) modeling already enjoy regulatory acceptance and routine use in dose selection, drug–drug interaction prediction, and pediatric extrapolation. PubMed

The critical next step is integration: taking the mechanistic detail from OoC and embedding it in systemic models so that experimental insights become predictive, usable, and regulatory-grade. This conversation explores how that synergy can overcome the adoption barrier, reduce reliance on animal testing, and drive translational human relevance.

 

1 | NAMs in 2025: Parallel Progress and Converging Needs

In 2025, the NAM landscape is defined by two complementary yet historically separate tracks. OoC provides high-fidelity, emergent biology at micro-scales, often focusing on tissue-level readouts like barrier integrity, flow response, or cellular crosstalk. But OoC alone faces hurdles in scaling results to human organ or whole-body context. Meanwhile, PBPK and QSP models excel at scaling, using physiological compartments, flows, and kinetics, but often lack mechanistic grounding in experimental human biology. The gap between these domains is precisely what integration aims to fill.


Henning: Looking back, what first convinced you that PBPK/QSP could meaningfully complement - or even substitute - certain animal studies?


Christian Maass:“PBPK and QSP modelling are established techniques and have been used for FDA IND submission, simulating clinical trials and understanding challenging biological questions for decades now. The new emerging field of OOC/MPS was initially driven by tech-bio aspects, but it became quite evident that biology alone isn’t solving the problems it is trying to solve. Hence my pushing for an adequate math model. If you make your biology more complex you ought to make your analysis tools equally complex to keep up.”

By embedding OoC-derived parameters - permeability, transport rates, metabolic fluxes - into PBPK/QSP frameworks, one can create digital twins of microphysiological systems that extend prediction to human scale.

 

2 | From Competition to Complementarity

Too often, computational and experimental NAMs are pitched as alternatives. But in practice, they address separate layers of the same translational challenge: mechanistic insight vs. predictive scaling. A mature NAM ecosystem treats them as symbiotic rather than rival.


Henning: OoC and NAMs often get framed as competitors. How do you see them fitting together?


Christian Maass:“It’s the perfect match. We see all those AI-only companies going through 150,000 compounds and claiming they can predict DILI when in fact it’s only 60–70%. The same for Biology-only papers – don’t get me wrong, this is good – but is it good enough for a pharma partner to trust 80% only? You still lose a lot of compounds by misclassifying. We can improve that to about 90%+ by simply analyzing the same data with our DigiToCs approach within the virtual testing center of MPSlabs (VTC). Paired with a federated learning approach, where we share a model with clients – not the other way around, results in benefits for all: the model learns from the clients’ data.”

ESQLabs explicitly frames the Virtual Testing Center (VTC) as the nexus where experimental and computational NAMs fuse: benchmarking chip systems, simulating with digital twins, and iterating in closed loop.

 

3 | Mechanistic Complementarity: Digital Twins of Chips

Digital twins are now being deployed not just for organs, but for the chips themselves. These model–device hybrids replicate microarchitecture, flow, transport phenomena, and boundary conditions in silico. That capability turns chip outputs into system-level predictions.


Henning: In what ways can PBPK/QSP models strengthen interpretation of OoC readouts?


Christian Maass:“Our digital twins of MPS/OoC follow the main and basic physiological micro-architecture observed in humans. So we bring PBPK/QSP to the OoC/MPS systems. In doing so, we are able to describe the distribution of compounds more accurately than is currently possible.”

This integration allows compound concentration/ time profiles in chips to be contextualized, correcting for scaling artifacts, boundary limitations, and microfluidic idiosyncrasies.

 

4 | Validation Frameworks: Aligning Experimental and Computational Standards

Regulators and industry alike demand robust, transparent validation strategies. Yet OoC and computational NAMs often operate under divergent standards. A true translational framework demands a unified validation logic: retrospective training and prospective prediction.


Henning: GAO highlights gaps in validation. What does a validation framework for in silico NAMs look like, and how should it align with OoC validation?


Christian Maass:“Good point! Ideally you train and build and fit a computational model (mechanistic model) on retrospective data. Then you can use this to actually fully predict new situations. Comparing those predictions against new data then validates the approach. ESQLabs also has a qualification framework within our own ecosystem.”

The alignment is crucial: you must treat computational and experimental models under fit-for-purpose validation, verifying predictive performance in the contexts they will be used - not over-engineering or universal certification.

 

5 | Bridging Variability: When Hardware and Cells Differ

One of the thorniest issues in OoC is inter-lab variability: chip design differences, cell donor heterogeneity, and batch effects. Computational models can absorb and interpret that variability rather than treat it as a barrier.


Henning: OoC variability often comes from chip architecture and donor cells. How can PBPK/QSP models help contextualize or correct for that variability?


Christian Maass:“We can either include this and build bespoke models per chip experiments, or we include variability on parameters, biological functions from chips, e.g. liver clearance or gut permeability, in the human PBPK models and simulate the effects more on a population level; this depends on the context-of-use.”

By modeling distributions rather than fixed parameters, you can propagate uncertainty through to outputs, enabling sensitivity analysis, comparison across devices, and population predictions.

 

6 | Standardization and Harmonized Data

Lack of shared metrics, nomenclature, file formats, and metadata is a persistent choke point in NAM adoption. Standardization efforts (CEN/CENELEC, NIH ring trials, OoC roadmaps) are underway - but integrating them across computational and experimental domains is imperative (CEN/CENELEC).


Henning: How do you envision harmonizing data formats, endpoints, and reporting standards across OoC labs and PBPK/QSP modeling groups?


Christian Maass:“I was part of the CEN/CENELEC standardization roadmap helping to shape what standards are needed and how to perform experimental setups, data to report and the format. The recent NIH Center for organoids is really pushing hard towards reproducibility, and we see that in the Liver Ring Trial we are part in. But, even when you fully follow the same protocol, variation remains.”

Defining units (e.g. permeability coefficients, flux), metadata (cell passage, media composition), and embedding computational metadata (model version, parameter provenance) in shared standards is a desired and highly anticipated goal industry-wide.

 

Multi-site qualification is essential for NAM credibility, yet sharing raw OoC data is constrained by IP, GDPR/PII, and clinical data silos. Federated learning (FL) replaces data pooling with model pooling: each site trains locally on its raw chip datasets (and linked outcomes) and shares only encrypted model updates for secure aggregation. For OoC + PBPK/QSP, this enables (i) federated calibration of digital twins (e.g., DigiLoCS) to harmonize device parameters across chip architectures; (ii) federated meta-models that map chip readouts to PBPK inputs (CLint, permeability) without exposing measurements; and (iii) federated population PBPK, where sites contribute parameter distributions (enzyme/transporter variability) to a global predictive population. Secure aggregation, differential privacy, and auditable provenance align this chip-to-cloud pathway with fit-for-purpose validation: pre-registered structures, retrospective local fitting, and prospective prediction and an aggregated global model. Practically, FL allows vendors, CROs, and sponsors to share models, not data and thereby accelerating generalizable performance while protecting confidentiality and regulatory traceability.

 

7 | Regulatory Acceptance: Building Trust Through Integration

PBPK and QSP modeling already serve as accepted regulatory tools; OoC data remains on the frontier. The path to broader acceptance is through integrated evidence - morphing chip results into regulatory-ready predictions.


Henning: Where have regulators (FDA, EMA, BfArM) already shown comfort with digital models, and how can OoC data strengthen that trust?


Christian Maass:“See above, PBPK/QSP models as well as popPK approaches have been used in many IND submissions and led to fewer clinical trials or animal experiments. In general, data analysis using computational modelling is key. The power of OoC data is limited. It’s cool if you can kill your tumor cells in vitro, but you don’t know if you treat your patient or kill the kidney with the same dose…you need a translator, that’s what computational modelling is and what we develop at MPSlabs.”

Chip-based dosage–response curves become interpretable only when embedded in systemic simulations that forecast human exposure, target engagement, and off-target risk.

 

8 | Scaling Translation: From Chip to Clinical Prediction

Integration isn’t conceptual - it’s operational. MPSlabs’ Virtual Testing Center is one platform where data flows from chip to model to prediction, benchmarking across chip types and guiding decision-making.


Henning: How does ESQLabs’ MPSlabs connect OoC experimental data with digital models to create predictive translational platforms?


Christian Maass:“Within our Virtual Testing Center, the world’s first integrated tool to analyse, visualise and translate OoC data for specific context-of-use situations, we have developed many services and showcases. For example, if you want to know how fast a new compound will be metabolized, you can study this in a liver-chip quite well. The problem is from an end-user perspective: if you search for liver-chip online, you now get 20 different companies offering you their services  -  how would you choose the best? We help by benchmarking the performance of different chips and enabling a comparison against each other. Once you have run your PK study in the chip, you often end up analyzing it with a simple Excel equation  -  how should the math model know your chip was a hypercomplex or a very simplistic system? That’s where we come in with our DigiLoCs approach:  a digital liver-chip simulator taking into account physics and chemistry, hardware and physiological information about the chip and cells. With this we can improve the prediction of human clearance by more than 80%.”

This end-to-end workflow is a working proof-of-concept: standardized benchmarking + in silico modeling + translational output.

  

9 | Therapeutic Applications: Where Synergy Shines

Every therapeutic area can benefit from integrated NAMs, but certain domains - oncology, brain, metabolic disease - are especially poised for translation, because they require both mechanistic microenvironments and whole-body context.


Henning: Which therapeutic areas (oncology, diabetes, rare diseases) stand to gain most from combining OoC + PBPK/QSP?


Christian Maass:“I don’t see a clear winner, it’s everyone and every field. If I had to pick, it’s definitely oncology and brain-targeting drugs. The amount of personalization possible with these models is enormous, but you again need to include that in your decision-making with computational models. We are developing hybrid models: integrating mechanistic pieces but also AI/ML into a single framework that helps us better understand individual differences, and allows us to pre-select potentially effective drugs per patient.”

In oncology, tumor–immune chips provide mechanistic insights while QSP models simulate systemic drug exposure and immune response; in neurology, BBB chips feed brain PK models to guide central exposures.

 

Case Study: Drug-Induced Liver Injury (DILI) — Chip-informed PBPK with DigiLoCS

Drug-induced liver injury remains a leading cause of attrition and post-marketing restrictions (Aravindakshan et al., see below). While liver-chip platforms yield human-relevant readouts, quantitative translation requires accounting for device physics, adsorption, and microenvironmental constraints that depress intracellular exposure and bias IVIVE. DigiLoCS is a digital liver-chip twin that integrates hardware/flow characteristics, physicochemical partitioning/binding, and hepatocellular metabolism in a three-compartment ODE framework (media–interstitium–intracellular) to infer effective intrinsic clearance (CLint) and separate passive transport from active biotransformation using time-series depletion data. As reported by the authors, “DigiLoCS, our developed digital liver-on-chip simulator, facilitates the accurate description of on-chip complex biology to disentangle biological processes.” Embedding these chip-derived parameters in human PBPK yields calibrated hepatic exposures (AUC, C ), time-above injury thresholds, and population variability driven by donor differences, enabling risk tiers rather than binary calls and reducing false positives/negatives relative to conventional scaling. In this integrated OoC → digital-twin → PBPK workflow, DigiLoCS functions as the quantitative translator from device-level kinetics to human-scale exposure and DILI risk, supporting prospective, fit-for-purpose evaluation across compound series.

Improving DILI classification ultimately depends on getting the on-chip clearance right. If intracellular exposure is under- or over-estimated, every downstream decision is affected. By recovering actual intrinsic clearance (CL ) and intracellular unbound exposure from liver-chip time courses (accounting for flow, adsorption, and partitioning), a digital twin analysis converts raw depletion kinetics into physiologically interpretable parameters. Using these corrected parameters in place of simpler depletion rates enables both discrimination (separating low- vs high-risk compounds) and calibration (aligning predicted risk with observed injury). Simulating the correct exposure in vitro is essential, and embedded within a hybrid mechanistic-AI/ML approach can be used to improve the classification of bio-only or AI-only predictions of DILI. Additionally, this approach is [CM1] 

 

10 | Education and Cross-Disciplinary Fluency

One of the unspoken barriers to NAM adoption is the cultural divide between biologists, engineers, and computational modelers. Bridging that gap is essential - without interdisciplinary fluency, even perfect tools remain underutilized.


Henning: How do we train scientists to use both experimental NAMs (OoC) and computational NAMs (PBPK/QSP) effectively?


Christian Maass:“I see three groups here. First, super experts in bio. Second, super experts in data science, stats and modelling. Third, the group that knows a bit of both. That group is relevant to speak the same language. It’s the awareness we have to raise and break down barriers that modelling is uncool, difficult or just math equations. It is actually a tool that adds value and enables the translation of insights from bench to bedside  -  without that, bio results are just a data point on a poster.”

Training programs, cross-disciplinary curricula, and collaborative hackathons can help create this “language bridge” layer within the scientific workforce.

 

11 | Open Science, Transparency, and Community Trust

Transparency is not optional - it’s foundational. Open source modeling, open standards, and community-driven validation accelerate trust and adoption. For both OoC and computational NAMs, reproducible, inspectable workflows are critical.


Henning: What role do open science communities (e.g., Open Systems Pharmacology) play in scaling adoption of NAMs?


Christian Maass:“We push for open science and community-driven approaches. We see increasing numbers of FDA IND submissions with OSPS (Open Systems Pharmacology Suites). If the code and science is open, you can monitor and cross reference, allowing for accelerated adoption. Lastly, of course, removing licensing fees also reduces barriers and serves as additional accelerant.”

Open modeling and open data pipelines make validation transparent, reproducible, and auditable - especially important for regulators and end-users seeking confidence in NAM-derived predictions.

 

12 | Vision 2035: Integrated NAMs Ecosystem

The ingredients for a fully integrated NAM ecosystem already exist: OoC, 3D tissues, PBPK/QSP, AI/ML. The remaining task is packaging, coordination, and shared platforms that connect them seamlessly into translational pipelines.


Henning: What does a fully integrated NAMs ecosystem (OoC + in silico + 3D tissues) look like in 10–15 years?


Christian Maass:“I don’t think we need that [long]. Technically we have all the tools ready today to start integrating that and building that ecosystem. In fact, that’s exactly what we have been doing for the last two years and the Virtual Testing Center is the first step towards that mission. Ideally you take a biopsy from a patient, cultivate it in the lab, maybe add a liver and kidney for tox to it. Screen pre-selected 5–10 compounds, come back a week later and get an optimal, digital-twin-based dose and compound. This should increase your treatment efficacy. If not, re-do the cycle. It’s data- and patient-driven.”

This feedback loop - where patient-derived tissues fuel chip models, which feed computational twins that suggest next-step experiments - is the foundational vision of translational precision medicine.

 

Conclusion

Integration across Organ-on-Chip and computational NAMs is not optional; it's becoming essential. OoC gives mechanistic resolution; PBPK/QSP lends interpretability and scale. Together, they form a unified engine for human-relevant, regulatory-ready evidence.

With robust validation strategies, open standards, and interdisciplinary fluency, we can advance the next generation of translational science - reducing animal reliance and bringing safer therapies to patients faster.

 

📚 References & Resources

 
 
 

Comments

bottom of page