In inhalation toxicology, more complexity is often assumed to mean more relevance. That assumption is convenient, but not always correct. Every extra cell type added to an in vitro model increases biological realism and at the same time creates new sources of variability, higher baseline activity, more difficult interpretation, and lower operational simplicity. The practical question is not whether a model can be made more complex. It is whether the added complexity improves the model for the intended use.
That is the central question addressed by Marescotti et al. (2019), who evaluated a 3D alveolar tetra-culture composed of alveolar epithelial cells, endothelial cells, macrophages, and mast cells organized at the air-liquid interface. The paper goes beyond standard phenotypic characterization. Instead, it uses transcriptomic data and computational biological network models to quantify what each added cell type contributes to the biology of the system. This makes the study especially useful for teams building or selecting advanced in vitro lung models: it does not simply ask whether the model looks more realistic, but whether it behaves in a way that justifies the complexity.
The authors developed and evaluated a 3D alveolar model intended to better mimic the human alveolar barrier for inhalation applications. The system combined A549 alveolar epithelial cells and EA.hy926 endothelial cells as the structural backbone, then added THP-1-derived macrophage-like cells and HMC-1 mast cells to create tri-culture and tetra-culture variants. The cultures were maintained at the air-liquid interface, an important design feature for respiratory exposure work because it better reflects how inhaled materials interact with lung tissue than submerged exposure systems.
The study design is strong because it compares multiple configurations instead of treating the most complex model as automatically superior. The authors assessed mono-culture, bi-culture, tri-culture, and tetra-culture conditions, then used network perturbation analysis to understand which biological pathways changed as cell types were added. They also compared transcriptomic profiles from these in vitro systems with publicly available profiles from healthy human lung tissue and from other 2D and 3D respiratory models. That last comparison is one of the most important parts of the paper, because it moves the discussion from model architecture to human relevance.
The first major finding is that adding endothelial cells to epithelial cells had a measurable but limited effect. The epithelial-endothelial co-culture changed cell-cycle and senescence-related biology, but did not radically transform the system. In practical terms, the structural backbone of the model matters, yet it is the immune component that creates most of the biological shift.
The second and most important finding is that macrophages are the dominant drivers of network perturbation in the complex model. When macrophages were added to the epithelial-endothelial co-culture, the model showed strong increases in inflammatory and oxidative stress signaling. The complete tetra-culture showed the highest overall biological impact, but much of that effect was already attributable to macrophages. Mast cells had a comparatively smaller incremental effect. This is a critical result because it shows that not all added cell types contribute equally to relevance. Some increase signal, some add noise, and some may only be justified for specific use cases.
The third major finding is operationally important: the baseline state of the model can become too activated. The authors observed that the addition of macrophages promoted a pro-inflammatory phenotype with elevated oxidative stress and immune signaling. For a model meant to detect responses to inhaled toxicants, that creates a risk. If the baseline is already highly activated, small or moderate exposure-induced effects may be harder to distinguish from the background. In other words, more complexity can reduce sensitivity if it pushes the system away from a stable homeostatic state.
To address that issue, the paper tested alternative assembly strategies. Resting macrophages for five days after differentiation reduced the pro-inflammatory baseline, while delaying the switch to air-liquid interface for 24 hours had the opposite effect and worsened stress signaling. This is a valuable methodological insight. It shows that model performance is not determined only by which cell types are present, but also by how the model is assembled and conditioned before use.
The final headline result is the most strategic one: transcriptional similarity to healthy human lung tissue increased as the model incorporated more lung-relevant cell types. Tri-cultures and tetra-cultures correlated better with human lung tissue than simpler systems and also outperformed several other 3D organotypic airway models included in the analysis. That is a strong argument for biologically informed complexity. At the same time, the gain from adding mast cells over a tri-culture was modest, which led the authors to question whether that extra layer of complexity is always warranted.
This paper matters because it provides a disciplined way to think about model design in inhalation toxicology. It does not promote complexity for its own sake. It shows that complexity should be added only when it improves the model against a clearly defined objective: physiological relevance, mechanistic coverage, sensitivity, or fitness for a specific testing question.
For research teams working on respiratory toxicology, the message is clear. A more complex model may be closer to the biology of the lung, but that does not automatically make it better for screening or decision-making. If the immune component introduces a high basal inflammatory state, the model may become less useful for detecting subtle responses. Conversely, if the goal is to capture multicellular crosstalk and better approximate human lung biology, then the inclusion of macrophages can be highly valuable.
For companies developing new approach methodologies, this study is also a reminder that validation should include baseline characterization, not just exposure response. The paper demonstrates that transcriptomics and computational network modeling can help determine whether each layer of complexity adds meaningful information or simply increases instability. That is directly relevant for platform developers trying to position advanced human-relevant lung models in industrial or regulatory contexts.
There is also a broader strategic implication. The study supports the argument that human-relevant in vitro models can outperform simpler legacy systems when they are built around the right cellular components and exposure architecture. But it equally supports the principle that the best model is not the most elaborate one. It is the one that is as simple as possible and as complex as necessary for the intended application.
Marescotti et al. provide one of the clearest evaluations of the trade-off between complexity and utility in an alveolar in vitro model. Their data show that adding relevant lung cell types increases similarity to human lung tissue, especially when immune cells are included, but also that this comes at a cost in baseline activation and model stability. Macrophages add major biological value, mast cells add less than expected, and assembly conditions materially affect performance.
The practical conclusion is not that every inhalation model should become a tetra-culture. It is that model designers and end users should define the question first, then build the minimum level of complexity required to answer it robustly. For inhalation toxicology, that principle remains one of the most useful takeaways from the paper.
Citation: Marescotti, D., Serchi, T., Luettich, K., Xiang, Y., Moschini, E., Talikka, M., Martin, F., Baumer, K., Dulize, R., Peric, D., Bornand, D., Guedj, E., Sewer, A., Cambier, S., Contal, S., Chary, A., Gutleb, A. C., Frentzel, S., Ivanov, N. V., … Hoeng, J. (2019). How complex should an in vitro model be? Evaluation of a complex 3D alveolar model with transcriptomic data and computational biological network models. ALTEX - Alternatives to Animal Experimentation, 36(3), 388-402. https://doi.org/10.14573/altex.1811221