3D cell models are cell cultures grown so that they reproduce the organisation and microarchitecture of a living tissue, rather than spreading cells as a single flat layer on plastic. The models range from spheroids and organoids through bioreactors and organ-on-chip devices to bioprinted tissue. They matter because 2D monolayers cannot mimic the tissue microenvironment, and cells grown in 3D respond differently to drugs. A 2025 Scientific Reports paper notes in its background that IC50 values for chemotherapeutics such as doxorubicin can differ more than tenfold between 2D models and patient-derived 3D tumour systems. 3D models bridge the gap between flat 2D culture and animal models.
Why 2D monolayers mislead drug-response studies
In tissue, cell-cell and cell-ECM interactions form a communication network that maintains the tissue’s specificity and homeostasis. A 2D culture grows in a static, rigid but standardised monolayer, and its cells do not take on the shape they have in the body, so proliferation, differentiation and apoptosis can all behave abnormally (Frontiers in Physiology review, 2022). The flat layer also misses spatial architecture, biochemical gradients and cell-ECM interactions, which compromises how well results translate to patients.
The same review gives three reasons 2D cells tend to look more drug-sensitive than 3D cells: they cannot hold normal morphology, their surface receptors are organised differently, and they tend to sit in the same cell-cycle stage, whereas cells in 3D are spread across stages the way cells in tissue are. None of this makes 2D culture useless; it has aided the discovery of many biological and pathological processes. It does mean that a potency figure from a monolayer is a figure about a monolayer.
Spheroids, organoids and organ-on-chip: how the 3D cell models differ
The terms organoids, spheroids and 3D cell cultures have been used interchangeably in the literature (Cancers, 2021). Spheroids and organoids still have distinct and overlapping purposes.
Spheroids
Spheroids are spherical cellular units, generally cultured as free-floating aggregates, and they are of comparatively low complexity in mirroring tissue organisation. Their strength is that they reproduce tissue-like gradients: a mature spheroid develops a necrotic core, a quiescent viable region and an outer proliferative layer, driven by reversed gradients of oxygen, nutrients and waste. They can be formed by static suspension, hanging-drop methods, magnetic levitation, spinner or rotational bioreactors, microfluidic systems or gel inclusion.
Organoids
Organoids are 3D structures that develop from stem cells or organ-specific progenitors through self-organisation and display architecture and function similar to the organ they model. They can be expanded long term, cryopreserved and kept in living biobanks, and they resemble the original tissue both histologically and genetically. One of the clearest clinical uses is cystic fibrosis: intestinal organoids from rectal biopsies swell in response to forskolin when healthy and much less in CF, a test that became a personalised organoid medicine application in the Netherlands.
Organ-on-chip
An organ-on-chip is a microfluidic device, fabricated with techniques such as photo- and soft-lithography, in which living cells are cultured to reproduce organ-level structure and function. The chip recreates tissue-tissue interfaces and gives fine control over fluid flow and mechanical forces, optionally with immune cells or microbiome. That addresses a specific weakness of spheroids and organoids grown in ECM gels: they lack tissue-to-tissue interfaces and organ-level structures.
Bioprinted tissue
3D bioprinting arranges biomaterials, biochemicals and living cells layer by layer with spatial control, using a “bio-ink” that mimics the extracellular matrix. Its open problems are bio-inks that cannot yet define tissue architecture well, mechanical stress during printing that harms cells (iPSCs often do not survive it) and the vascularisation of printed constructs. We would treat bioprinting as promising rather than characterised.
Is an organoid the same as a 3D cell culture?
No: an organoid is one kind of 3D cell culture, defined by self-organisation from stem or progenitor cells and by resemblance to a specific organ, whereas spheroids are simpler free-floating aggregates. In practice, the distinction to check in any paper is the cell source, the protocol and the time needed to establish the model, since that is where spheroids and organoids differ.
Comparing the 3D cell culture models at a glance
| Model | What it is | What it captures well | Main gap |
|---|---|---|---|
| 2D monolayer | Cells on a flat, rigid surface | Standardisation | Unable to mimic the tissue microenvironment; distorted drug sensitivity |
| Spheroid | Free-floating cell aggregate | Oxygen and nutrient gradients, necrotic core | Low organisational complexity |
| Organoid | Self-organised stem-cell-derived structure | Organ-specific histology and genetics, biobanking | Usually no neighbouring-tissue microenvironment |
| Organ-on-chip | Microfluidic device with organ-level culture | Tissue interfaces, flow, mechanical forces | Systematic predictive validation not reported before the Liver-Chip study |
| Bioprinted tissue | Layer-by-layer printed cells in bio-ink | Defined spatial arrangement | Cell stress, vascularisation, bio-ink limits |
How 3D cell culture models are built
3D cultures are either scaffold-based, where cells grow on synthetic materials (ceramics, metals, polymers), natural ones (polysaccharides, proteins, ECM-derived or acellular matrix) or hydrogels, or scaffold-free, as with spheroids, organoids and self-assembly.
Organoid cultures are embedded in Matrigel or other ECM gels; Matrigel is a mouse-sarcoma basement-membrane extract used for more than four decades. One published organoid protocol mixes cells with Matrigel and medium 50/50 and lets the dome solidify for 15 minutes at 37 °C before adding medium. Its weakness is composition: Nature Reviews Materials (2020) describes Matrigel as complex, ill-defined and variable, with mechanical and biochemical variation within and between batches that undermines reproducibility, and it resists tuning toward specific cell behaviours. Synthetic scaffolds now offer xeno-free, chemically defined and tunable alternatives, including a fully synthetic PEG hydrogel protocol for human intestinal and lung organoids.
Moving from static to dynamic culture is the other building choice. Bioreactors apply flow and mechanical, physical or chemical stimuli, and the 2022 review frames them as the intermediate step between a static 3D model and an organ-on-chip. Perfusion bioreactors and microfluidic devices keep fresh or recirculating medium flowing past cells on scaffolds, which improves nutrient and oxygen exchange.
Even simple spheroids are sensitive to media, oxygen and serum levels. A 2025 Scientific Reports study found that 3% oxygen produced smaller spheroids with more necrosis, serum above 10% produced dense spheroids with distinct zones, and serum-free culture shrank MCF-7 spheroids more than threefold to about 200 µm. Seeding density changes spheroid size, and seeding around 6,000 to 7,000 cells caused some spheroids to rupture. The authors’ conclusion is that drug screening needs standardised media and serum, and that comprehensive reporting of culture parameters supports both reproducibility and regulatory approval.
How well do 3D models predict drug toxicity?
Ewart et al. (Communications Medicine, 2022) tested 870 Liver-Chips against a blinded set of 27 drugs with known hepatotoxic or non-toxic behaviour, and compared the results with animal models and 3D spheroids of primary human hepatocytes. The Liver-Chip met qualification guidelines with a specificity of 100% and, with two donors and protein-binding correction, a sensitivity of 87%.
The comparison with spheroids depends on the analysis, so it is worth reading carefully. On 15 toxic drugs tested with two donors, the chip’s sensitivity was 80% against 42% for 3D hepatic spheroids on the same drug set from previously published data. The chip flagged no safe drug as toxic, whereas the spheroids had 67% specificity. The 87% figure comes from including the 18 drugs tested in both donors with protein-binding correction. The chip did miss zileuton, whose toxicity likely involves reactive metabolites from cytochrome P450 metabolism.
This is one organ, one device and 27 drugs, so the result supports the Liver-Chip for drug-induced liver injury rather than organ-on-chip as a category. The authors place the chip between in vitro tests and dose-range-finding animal studies, as a way to deprioritise toxic compounds early. Their economic model estimates about $3 billion a year in industry R&D productivity from broad adoption of Liver-Chips alongside existing preclinical models to predict small-molecule drug-induced liver injury, which is a modelled projection, not an observed result.
Regulators have started to make room for these methods. The FDA describes New Approach Methodologies as including in vitro human-based systems such as organs-on-chips, and in April 2025 it released a roadmap for a stepwise move from animal testing to validated NAMs, beginning with monoclonal antibodies. The FDA also lists a reconstructed human cornea-like epithelium model that replaced rabbit eye-irritation tests, and a 3D reconstructed human epidermis model accepted, when warranted, for dermal irritation. Acceptance is framed around validation, with a framework setting regulatory expectations for non-animal methods.
Where machine learning image analysis helps with organoids, and where it fails
The bottleneck in organoid screening is measuring responses across large numbers of microscopy images, because organoids vary widely in size and shape and drift in and out of the focal plane. OrganoID (PLOS Computational Biology, 2022), a deep learning tool among AI applications in biotechnology, was built to close that gap. It combines a convolutional neural network, a contour-resolving identification module and a tracking module, and works on brightfield and phase-contrast images without cell labelling. Trained on pancreatic cancer organoids, it reached a mean intersection-over-union of 0.74 on unseen pancreatic images and 0.79 on lung, colon and adenoid cystic carcinoma organoids, against a pass mark of 0.5. Its tracking held above 89% accuracy across a 92-hour time-lapse, and morphology readouts such as circularity and solidity gave sigmoidal dose responses to gemcitabine. Following individual organoids frame to frame is the same association problem as multiple object tracking.
Generalisation has limits, though. On mouse small intestinal organoids OrganoID averaged 0.70, but three images failed the benchmark; retraining on a subset of those images lifted the result to 0.82 while holding 0.76 on the original test set. The tool assigns one organoid per pixel, so it cannot separate organoids that overlap across the focal plane.
Co-cultures are harder still. A 2023 Frontiers in Medicine study found that Incucyte and OrganoSeg software reached precision and recall between 0.00 and 0.43 on organoids in immune-cell co-cultures, while the authors’ own StrataQuest-based Organoid App, a trained classifier using intensity, morphology and Haralick texture features, reported 0.91 to 0.95 (the authors’ own evaluation, data not shown). The same study names three failure modes that any detector on these images has to survive:
- Contour mimicry: dense immune-cell clusters near organoids mimic organoid morphology and become false positives.
- Contour fusion: neighbouring organoids are recognised as one structure, undercounting.
- Contour disruption: poor imaging or overlapping immune cells break organoid outlines.
The MultiOrg dataset (2024) adds the underlying causes: 3D cultures imaged in 2D produce overlapping structures, organoids can be hard to tell from dust and debris, and expert labels are themselves subjective, which is why it provides multi-rater annotations.
Our reading, as computer vision engineers, is that these are the familiar problems of any detection system pushed into a new domain: ground truth is uncertain, the deployment distribution differs from the training one, and an aggregate metric hides which conditions fail. We would report precision and recall per tissue type and per culture complexity rather than one pooled figure, score against more than one annotator where labels disagree, and keep multiple focal planes where overlap matters, because a single projection throws away the information that would separate touching organoids. We would also version the detection model and its training images with the assay protocol, since a retrain changes the measurement. That engineering is the core of computer vision R&D applied to lab imaging, and the Frontiers authors’ own caveat sets the bar: no algorithm should be expected to detect 100% of target structures.
Limits that still apply to every 3D cell model
Size is a physical ceiling for spheroids: above about 500 µm in diameter, the multilayered structure restricts diffusion of gases and nutrients and impedes waste removal. That gradient-driven zonation is part of what spheroids are used to model.
Organoids still struggle with controlled self-organisation, achieving physiologically relevant shape and size, and lifespans long enough to mature. In tumour work, missing stroma, fibroblasts, endothelial and immune cells can compromise prediction of clinical outcome. Co-culture is one way to add neighbouring tissue, though it complicates imaging, as the detection results above show.
Variability runs through the stack: Matrigel batches, undefined serum composition and cell-line differences all undermine reproducibility. And the 2022 review is plain that no current method yet reproduces in vivo conditions completely. The practical question for any programme is which human outcome a given model has been validated against, because that, rather than its dimensionality, decides how far its results can be trusted.
