One of the most exciting updates to emerge at DP&AI: Europe 2025 was around complex in vitro models (CIVMs) – advanced, three-dimensional laboratory systems that simulate human tissue structures and physiological functions. In one compelling presentation, Luisa Bell, Neuropathology scientist at Roche, demonstrated how CIVMs offer an unprecedented opportunity to support enhanced preclinical to clinical translation by generating high-quality mechanistic data in human in vitro models.
Here, alongside fellow researcher Nadine Stokar, Bell elaborates on a new digital pathology approach for preclinical toxicity testing using blood-brain barrier (BBB) organoids.
What was the inspiration for your study?
In drug development, we rely on in vitro models, but there has long been a gap between simple cell cultures and the complexity of human biology. Recent advances in CIVMs, such as BBB organoids, offer new potential for preclinical testing. However, to support use in regulatory settings and safety assessment, these models must undergo the same rigorous characterization and validation expected of traditional tissue specimens.
We recognized that applying standard histological techniques to these advanced organoids could help bridge innovative bioengineering with established safety evaluation practices.
Did you face any challenges in your research, and if so, how did you overcome them?
Our challenges were threefold: technical, strategic, and linguistic. Technically, we had to adapt traditional workflows to handle fragile, small CIVMs. Strategically, we needed to identify which model was truly fit for purpose for specific questions, which led us to develop a decision tree to guide model selection. Finally, we had to bridge a communication gap between bioengineers, biologists and model owners, and pathologists. Establishing a shared language was challenging at first, but it ultimately became our greatest strength, turning an interdisciplinary team into a genuinely collaborative one.
What were the main outcomes of your study?
We developed a suite of tissue technology–based methods – ranging from H&E and immunohistochemistry to multiplex immunofluorescence and MALDI mass spectrometry imaging – tailored to the characterization and validation of organoids in preclinical drug development. To support scalability, we also built an AI algorithm to automate the assessment of single-cell toxicities.
One unexpected finding was that the AI performed even better on glioblastoma organoids than on simpler models such as BBB organoids. The more “parenchyma-like” architecture provided richer structural context, enabling highly precise morphological assessments.
In your talk, you noted that the bench-to-bedside pipeline often fails at early decision points. What are the main limitations in current preclinical evaluation methods that contribute to this gap?
Traditional 2D cell cultures are often too simplistic, and animal models do not always reflect human-specific biological responses, which can lead to unexpected toxicities. To help close this gap, we need to shift toward human-relevant new approach methodologies. Complex 3D models are a prime example: they can better reproduce tissue architecture and cellular cross-talk, enabling earlier identification of drug liabilities and safety concerns in the development pipeline.
Which embedding and sectioning methods have proven most effective for generating high-quality slides suitable for routine histology, IHC, and digital pathology assessment?
While there is no one-size-fits-all approach, we found that formalin-fixed, paraffin-embedded processing is generally best for preserving the delicate morphology needed for AI-supported digital pathology. However, the choice of preparation should be driven by the intended readout. For example, MALDI mass spectrometry imaging requires cryo-embedding in gelatin.
How do multiplex IHC, RNA-based markers, or spatial assays aid in validating organoid integrity and detecting subtle diagnostic changes that might otherwise be missed?
Morphology matters. These techniques allow us to confirm organoid architecture, including correct layering and cellular polarization. While 3D imaging is widely used, it is highly labor-intensive. By converting evaluations into high-resolution 2D sections on glass slides, we can increase throughput while still providing the spatial context pathologists need to detect subtle changes in tissue health.
How can algorithm-supported organoid analysis help standardize histopathology evaluation and reduce variability between reviewers?
Pathologists are trained to interpret the complex architecture of human and animal tissues, so moving into organoid assessment requires a shift in perspective. The first step is learning what a “healthy” baseline looks like in a synthetic system. Algorithm-supported analysis can be transformative here, providing objective, quantitative benchmarks that help harmonize evaluations across reviewers and turn qualitative morphological observations into reproducible, data-driven metrics.
Beyond standardization, there is a clear practical benefit: efficiency. These studies often generate large volumes of samples, and manual review can quickly become a bottleneck. AI can rapidly analyze the full cohort, reducing reliance on representative sampling and enabling more comprehensive, consistent assessment – while significantly speeding up the workflow.
What contributions from engineers, computational scientists, or bioengineers have been essential to making organoid-based pathology workflows reliable?
This is clearly a team effort. A pathologist cannot build an organoid alone, and an engineer cannot interpret complex toxicologic injury patterns without pathology expertise. We believe the future of the field lies in multidisciplinary teams: biologists to grow the models, engineers to build the platforms, and pathologists to validate and interpret the results. The pathologist’s eye is essential to confirm that organoids meaningfully reflect human biology and to help define the most appropriate context of use for each CIVM.
Looking ahead, what advancements in 3D modelling, computational staining, or digital workflows do you predict will further strengthen the diagnostic utility of organoid-based pathology?
The field is moving toward more automation and harmonization. The MPS expert field is increasingly exploring multi-organ systems that can show how a drug affects, for example, the brain and liver at the same time. The next step will be incorporating patient-derived samples to better reflect genetic heterogeneity. Ultimately, clinicians may be able to test a specific drug dose on a patient’s own “tumor-on-a-chip” before treatment ever reaches the bedside.
What advice would you give pathology departments interested in piloting organoid-to-algorithm approaches for diagnostic research or laboratory evaluation?
Our advice is to start with a collaborative mindset. Pilot programs work best when pathologists partner closely with the engineers designing the models. Begin by identifying a clear scientific problem – such as quantifying preclinical drug-induced toxicities – that would benefit from high-throughput digital analysis. When developing your algorithm, establish a strong ground truth from day one; whether you rely on IHC-based markers or expert manual scoring, consistency is essential. Ultimately, to gain end-user confidence, you need a robust validation package showing that the digital approach is reliable, reproducible, and offers a clear advantage over traditional manual screening.
And to our fellow pathologists: morphology matters more than ever! As medicine becomes increasingly complex, our ability to interpret structure and form remains essential. Don’t be afraid to step beyond the traditional pathology lab and into the world of complex in vitro models – this is where the next major advances in diagnostics are taking shape.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
