Rethinking reproducibility in advanced 3D cell models

Drug developers have long sought models that better reflect human biology than conventional 2D cultures. Over the past decade, organoids, spheroids, and patient-derived models have moved steadily from academic proofs of concept toward practical use in toxicity screening, oncology, and disease modeling. Yet broader adoption has exposed friction points that enthusiasm for the technology initially obscured. Distinguishing meaningful biological variation from technical noise remains a persistent challenge, as do the operational demands of keeping complex living systems consistent across sites and studies.

These pressures are pushing the field toward a harder question: what does it actually mean for results from a 3D model to be reliable? The answer is forcing researchers to rethink not only their workflows but also their assumptions about variability and what standardization requires.

Nicola Bevan is a manager of cell imaging applications in the BioAnalytics, Product Development group at Sartorius, where she works at the intersection of 3D model development, imaging, and quantitative analysis. DDN spoke with Nicola Bevan about what realistic progress toward mainstream adoption requires, where expectations still run ahead of practical capability, and why the field's approach to variability may need to change fundamentally.

Let's look ahead to the next several years. If we were having this conversation in 2031, what developments would lead you to say that advanced 3D cell models have become firmly established in mainstream research and development?

Mainstream adoption requires access to fit-for-purpose materials, reproducible workflows, and scalable quantitative methods.

First, researchers would need reliable access to relevant cells, tissues, and matrices, supported by clear sourcing, consent, intellectual-property, and licensing frameworks. Patient-derived material and induced pluripotent stem cell (iPSC)-derived cells should be obtainable through established supply chains rather than through a small number of specialist laboratories or CROs. The matrices used to support these models should be well characterized, ideally with defined composition and consistent mechanical properties.

Second, reproducibility will need to be more sophisticatedly understood. With patient-derived and disease-specific models, the objective should not be to eliminate all biological variation. Instead, we need to characterize that variation and distinguish it from technical variability. Well-documented workflows will be essential, with standardized controls and agreed quality metrics.

Finally, quantifying complex 3D biology must become more streamlined. Researchers should be able to obtain robust morphological, functional, and molecular information from thick tissues, organoids, and co-culture systems. Automated imaging, label-free measurements, multiparametric analysis, and AI-supported interpretation will all contribute to this.

Organoids and other complex 3D models have generated enormous excitement across research and drug discovery. Looking at the field today, where do you think expectations still exceed reality?

Expectations for complex 3D models currently exceed practical capabilities in three main areas: assay throughput, access to patient-derived organoids (PDOs), and our understanding of model phenotype and heterogeneity.

The first area is throughput. There is an assumption that complex organoid workflows can be translated directly into higher-throughput screening workflows. Generating, maintaining, and analyzing 3D cultures at scale remains technically demanding. Matrix handling, cell seeding, culture conditions, timing, and imaging all introduce variables that are less significant in conventional 2D assays.

The second is access to PDOs. Specialist CROs and academic centers have developed valuable PDO capabilities, but access is still uneven. Tissue sourcing, consent, transport, expansion, and characterization all remain important practical constraints.

The third area is our understanding of model phenotype and heterogeneity. We often describe a model as an organoid or spheroid without fully understanding how closely its biology represents the tissue or disease of interest. We still need stronger links between model phenotypes, functional biology, and clinical outcomes.

The field has made significant progress, but these models should not be viewed as universally plug-and-play replacements for existing assays. Their value depends on the biological question, the quality of the model, and the strength of its characterization.

Where have 3D models already delivered clear translational wins, and what can the field learn from those examples?

Commercial tissue models, such as those used in skin and ocular research, have demonstrated the value of human-relevant 3D systems for tissue toxicity, irritation, and barrier-function studies. Their success has been supported by clear contexts of use, standardized protocols, and defined performance criteria.

In oncology, solid-tumor spheroids and tumor-immune co-culture models can provide insights into tumor architecture, immune-cell infiltration, treatment response, and resistance mechanisms that are difficult to reproduce in 2D cultures. Similarly, iPSC-derived and multicellular models, including neural and tri-culture systems, are helping researchers investigate disease biology and toxicity in a more human-relevant context.

The main lesson is that successful translation depends on deep phenotyping. A model does not have to replicate every aspect of human disease, but researchers must be precise about the question it is intended to answer. The strongest examples are well characterized, reproducible, and demonstrably relevant to a specific decision.

Which scientific, technical, or operational challenge do you believe deserves more attention from the community?

The most critical operational challenges are transport logistics, automation-compatible matrices, and quantifying thick tissues. Moving patient-derived tissues or living 3D cultures between hospitals, biobanks, CROs, and research sites can affect viability, phenotype, and reproducibility. Standardized shipping conditions, validated containers and media, defined time limits, and environmental monitoring will be needed to make these workflows reliable.

A second major challenge is the development of animal-free, defined, and automation compatible matrices. Many commonly used matrices are biologically variable and can be difficult to dispense consistently. Future matrices should ideally have well defined composition and mechanical properties, be compatible with automated liquid handling, and support consistent culture at scale.

A third challenge is quantifying thick and complex tissues. More attention should be given to volumetric imaging, label-free methods, non destructive measurements, and multiparametric analysis. If we want to scale 3D models, we need materials that can be handled consistently, logistics that preserve the biology, and analytical methods that can extract reliable information without destroying scarce samples.

How should researchers think differently about variability when working with patient-derived or disease-specific models?

Variability should be treated as biological information rather than simply as experimental noise.

Reliable controls are the first requirement; healthy donor-derived models, reference cell lines, characterized patient-derived models, or controls with known treatment responses, included consistently across batches and, where possible, across sites. This helps separate technical variation from genuine biological differences.

Rather than reducing each model to a single average value, researchers should consider subpopulations, phenotypic clusters, response distributions, and relationships to donor or disease characteristics. Because patient material is often limited, label-free and multiparametric methods can support non-destructive, longitudinal measurements, allowing researchers to monitor changes over time and preserve samples for additional analyses.

The aim should not be to make every model look the same. It should be to understand which sources of variation matter, which are acceptable, and which are predictive of patient response or toxicity.

Everyone agrees that standardization is critical for broader adoption. But when it comes to organoids and advanced 3D models, what does "standardized" actually mean in practice?

Standardization needs to cover the complete workflow, not just the final assay. At the biological level, this includes consistent requirements for tissue sourcing, transport, processing time, temperature, media, matrix composition, tissue or organoid size, culture duration, and passaging, along with defined acceptance criteria for viability, morphology, growth, and differentiation.

At the analytical level, standardization means using consistent imaging settings, assay conditions, controls, data formats, and analysis pipelines. Researchers should be able to understand what was measured, how it was measured, and whether the results meet predefined quality thresholds.

Standardization does not necessarily mean that every laboratory must use the same model or protocol. The important point is that protocols, critical variables, and performance characteristics are clearly defined, so results can be compared and interpreted. Automation can help embed standardization into routine workflows, but it is only effective when the underlying biology and process have already been sufficiently understood.

How important will computational tools and AI be in unlocking the full value of complex 3D systems?

Computational tools and AI will be essential because the complexity of 3D models generates more information than can be reliably assessed manually.

One application is the use of historical control data as a predictive reference, allowing new data to be compared with well-characterized historical datasets. This may reduce the amount of sample required while still allowing researchers to identify outliers. AI-driven image analysis can also improve consistency and reduce user bias by quantifying subtle morphological and spatial features that are difficult to score reproducibly by eye.

Linking experimental results to donor information, biobank records, treatment history, and clinical outcomes could help researchers identify which model characteristics are genuinely predictive. In the longer term, AI could support automated culture and passaging. For example, using imaging data to determine whether a model is ready for passage or whether culture conditions need adjustment.

AI should not be viewed as a substitute for biological understanding. Models need diverse training data, appropriate validation, version control, and transparent performance assessment. Otherwise, there is a risk of automating bias rather than removing it.

Which application area do you believe will see the greatest impact from advanced 3D models over the next five years, and why?

Over the next five years, 3D models will have the greatest impact on predictive tissue toxicity assays and oncology research.

Tissue toxicity models address a clear need for more predictive, human relevant assays. Models of skin, eye, gut, liver, and other tissues can provide information about tissue structure, barrier function, irritation, and toxicity that is difficult to obtain from 2D cultures. Commercial tissue models have already shown how a clearly defined context of use, standardized handling, and robust performance data can support adoption.

In oncology, tumor spheroids, organoids, and tumor-immune co-cultures can capture aspects of three-dimensional architecture, cell-cell interaction, immune infiltration, and treatment resistance. These features are particularly relevant for understanding why apparently similar tumors respond differently to the same therapy.

The greatest value will come when these models are combined with deep phenotyping and patient linked information. It is not enough to generate a complex model; we need to determine which features predict drug response, toxicity, or clinical outcome. I expect the strongest progress to come from applications where the biological question is well defined, the model can be characterized rigorously, and the results can be connected to a meaningful development decision.