Tumor Models Are Getting Real: What's New in 3D Cancer Drug Screening

Team Dynamic Matrices | 7 June 2026

For a decade the pitch for 3D tumor models was simple: grow a piece of a patient's cancer, test drugs on it, predict what will work before the patient receives anything. That promise is now being tested at scale, and the results are forcing a more honest conversation about what these models can and cannot do. Our summary: the biology has caught up to the hype, and the matrix underneath it has quietly become the rate-limiting step.

The evidence for clinical predictivity is real

The landmark dataset is still Vlachogiannis et al. (Science, 2018). Patient-derived organoids from metastatic gastrointestinal cancers predicted patient response with 100 percent sensitivity, 93 percent specificity and an 88 percent positive predictive value. Ooft et al. (Science Translational Medicine, 2019) showed colorectal organoids predicted irinotecan response in more than 80 percent of patients. Tiriac et al. (Cancer Discovery, 2018) built pancreatic organoid signatures that tracked with clinical chemotherapy outcomes. Driehuis et al. (PNAS, 2019) extended this to head and neck cancer, including radiotherapy response.

The Human Cancer Models Initiative and national biobanks now hold hundreds of lines across a dozen tumour types, with histopathology and driver mutation concordance to the parent tumour generally reported above 80 to 90 percent. What was anecdote is now dataset, and the dataset says functional testing catches responders that genomic panels miss. That raises an uncomfortable question: if organoid screening outperforms the biomarkers we gate treatment on, why is it still a research tool?

The immune system was the missing character

Most tumour organoids have been epithelial monocultures: cancer cells alone, stripped of stroma, vasculature and immune infiltrate. Tolerable for chemosensitivity, useless for a checkpoint inhibitor.

Two approaches are closing the gap. Neal et al. (Cell, 2018) used air-liquid interface culture to keep the native tumour microenvironment, including infiltrating T cells, intact and showed PD-1 blockade responses in the dish. Dijkstra et al. (Cell, 2018) co-cultured organoids with autologous peripheral T cells to expand tumour-reactive clones. These immune-inclusive models are a different object from a clonal organoid: closer to a functional slice of the tumour ecosystem. The trade-off is harder standardisation, because immune composition varies patient to patient in ways epithelial organoids do not.

Reproducibility is still the uncomfortable truth

None of this changes a structural problem: protocols are not harmonised, and the matrix is a major reason. Basement membrane extracts vary lot to lot in stiffness, ligand density and growth factor content (Hughes et al., Proteomics, 2010; Kozlowski et al., Communications Biology, 2021). In a chemosensitivity assay that is not cosmetic. A stiffer or growth-factor-richer lot can shift apparent drug response independently of anything meaningful about the patient's cancer.

This matters more in oncology than almost anywhere else, because the whole value of the screen is a quantitative readout, an IC50 or a sensitivity call, that a clinician might act on. A model that behaves differently depending on which lot it grew in cannot support that decision. The Vlachogiannis result is compelling precisely because it came from a controlled dataset. Scale that across hundreds of labs on undefined matrices and the signal-to-noise gets much worse.

Matrix mechanics are also biology here, not just noise. Tumours are stiffer than the tissues they arise in, and stiffness drives drug resistance through integrin and YAP signalling (reviewed in Northey, Przybyla and Weaver, Cancer Discovery, 2017). If you cannot set stiffness, you cannot study that.

What a defined, tunable matrix buys you

The fix is not more biology; it is more control over the scaffold the biology sits in. A chemically defined synthetic matrix removes lot variation by design. A tunable one lets you set stiffness, stress relaxation and ligand presentation to match a specific tumour type rather than accepting whatever an extract provides. A pancreatic organoid and a lung organoid can be grown in matrices tuned to their native mechanics, and a resistant subclone's response to a stiffer environment becomes something you probe deliberately instead of an artefact you explain away.

Our own recent work uses DyNAtrix as a force probe to study metastatic behaviour of breast cancer organoids, exploiting the fact that the DNA crosslinks report on the forces cells apply. That kind of experiment is only possible when the matrix is defined enough to be an instrument.

As functional screening moves toward informing treatment, the assay has to survive a tumour board, not just a grant reviewer. That means knowing exactly what the tumour grew in and why it responded as it did. The field has proven the concept works. The next job is making the platform underneath precise enough to be believed.