Part XVII · Research methods, evidence, and emerging therapy · Chapter 93

Experimental models and discovery platforms

Why so much preclinical breast cancer work does not translate, argued from what each model can and cannot represent.

1 · Cell lines and their persistent limitations

The breast cancer cell line panel is small, old and unrepresentative, and the field knows this while continuing to depend on it. A handful of lines account for the large majority of published experiments. MCF7 and T47D carry most of what is claimed about oestrogen receptor biology. MDA-MB-231 carries a disproportionate share of what is claimed about triple-negative disease, despite being a poor representative of the common triple-negative transcriptional programmes described in Why TNBC classifiers disagree with one another.

Three limitations are structural rather than fixable by better technique. Lines were selected for the ability to grow indefinitely on plastic, which is a phenotype almost no primary tumour has. That selection is not neutral with respect to the biology being studied. Lines drift, so the same named line differs between laboratories after decades of separate passage. And lines have no microenvironment, so every finding about tumour-stroma or tumour-immune interaction described in Functional states of the microenvironment and Spatial organization and what it adds is unavailable in them by construction.

The luminal problem is the sharpest. Hormone-receptor-positive disease is the majority of breast cancer and is the hardest to model, because the lines that grow reliably have been selected away from dependence on the signalling being studied. That mismatch is one reason the preclinical literature over-represents triple-negative biology relative to its share of the disease.

2 · Patient-derived xenografts and organoids

Patient-derived xenografts preserve tumour architecture, stromal content at least initially, and intratumoural clonal diversity, which no cell line does. They retain sensitivity patterns that track the donor patient's clinical course better than lines do. They are the closest available system for asking whether a drug works against a real tumour rather than an adapted one.

Their limitations are specific and predictable. Engraftment rate is strongly biased by subtype. Triple-negative and high-grade tumours engraft readily, while hormone-receptor-positive tumours engraft poorly. The resulting model collection is therefore skewed away from the commonest disease, repeating the bias described in Cell lines and their persistent limitations by a different route. The host is immunodeficient, so the model cannot address immunology. Mouse stroma replaces human stroma over successive passages, so the microenvironment being studied is not the patient's.

Organoids grow from patient tissue in three dimensions and preserve epithelial architecture and differentiation. They establish faster than xenografts, can be biobanked, and are tractable for screening. What they lack is everything outside the epithelium. An organoid has no vasculature, no immune compartment and, in most protocols, no fibroblasts. A drug whose mechanism runs through the microenvironment will look inactive in an organoid for reasons that have nothing to do with the tumour cell.

3 · Explants and ex vivo drug sensitivity

Explant culture keeps a fragment of tumour intact with its native stroma and resident immune cells for a short window. That window is the point. For a few days the system contains the actual microenvironment of the actual patient, which neither organoids nor xenografts provide.

The trade is duration against fidelity. Explants degrade over days, so they can address immediate pharmacodynamic questions such as target engagement, apoptotic response and immune cell activation. They cannot address anything requiring weeks, which includes acquired resistance, dormancy in Dormancy, residual disease, and late recurrence, and most of what determines clinical outcome.

Ex vivo drug sensitivity testing as a clinical tool has a long history of disappointing results and the reasons are instructive. The assay measures short-term cytotoxicity in a sampled fragment. Clinical response depends on drug delivery, on the fraction of the tumour the fragment represents, and on resistant subclones that may be rare at the time of sampling. The sampling problem here is the same one that Part VI raises for biomarkers, and it is not solved by a better assay.

4 · Immune-competent and humanized models

Anything said about immunotherapy requires an immune system, which excludes standard xenografts.

Syngeneic mouse models place a mouse tumour line in an immunocompetent mouse of the same strain. The immune system is complete and functional. The tumour is murine, so its antigens, mutational burden and signalling are murine. The commonly used mammary lines carry mutational burdens and immune phenotypes that do not match human breast cancer. The field's tendency to work in the most immunogenic of them is one reason preclinical immunotherapy results look better than clinical ones.

Genetically engineered mouse models induce tumours in situ from defined driver alterations, in an intact host with an intact immune system. They model tumour initiation and progression, which transplant systems do not. They are slow, expensive, and typically driven by one or two alterations, producing tumours with far less genomic complexity than human disease.

Humanised mice reconstitute a human immune system in an immunodeficient host carrying a human tumour. The reconstitution is partial. T-cell development occurs in a mouse thymus on mouse major histocompatibility complex, myeloid reconstitution is poor, and graft-versus-host reactivity limits the experimental window. The immune biology in Breast cancer immunology is only partly accessible in these systems, and myeloid biology is largely inaccessible.

5 · Single-cell and spatial profiling platforms

Single-cell transcriptomics resolved the population-average problem that had made every earlier expression measurement ambiguous. A bulk profile of a tumour is a weighted average over epithelial cells, fibroblasts, endothelium and leukocytes. Single-cell profiling separates those compartments and showed that much of what was attributed to tumour cells was stromal¹.

Spatial methods restore the coordinates that dissociation destroys². Position is not decoration. Whether a cytotoxic T cell is inside the tumour nest or excluded at its margin is the distinction that separates an inflamed from an excluded phenotype, and that distinction is invisible in dissociated data. The argument that this reshapes is in Spatial organization and what it adds.

Each platform makes a specific trade. Sequencing-based spatial methods cover the transcriptome at limited spatial resolution. Imaging-based methods reach single-cell or subcellular resolution across a targeted panel of a few hundred genes. Protein-level spatial methods measure what is actually present rather than its transcript, at panel sizes of tens of markers.

That last distinction is the one that matters for this book. Spatial proteomics across 280 tumour regions found receptor subtypes co-existing within single tumours, and regional proteomic states that the receptor classification the tumour was given does not resolve³. Proteomic heterogeneity rose with grade, was independent of genomic heterogeneity, and tracked microenvironmental differences. Note what the comparison was against. The regions were not subtyped transcriptionally, so the finding is protein against the assigned immunohistochemical call. A transcript-based platform would not have generated it, which is why the choice of platform determines whether a particular kind of heterogeneity is findable at all.

6 · Functional genomics and perturbation screens

Association tells you a gene is present. Perturbation tells you whether it is required. CRISPR knockout, interference and activation screens supply the second, at genome scale, and they are what separate a driver from a passenger that travels with one.

Screen design determines what can be found. A dropout screen in standard culture finds genes required for proliferation on plastic, which is a narrow question. A screen run under drug pressure finds resistance and sensitisation genes, which is a more useful one. A screen run in vivo finds genes required for growth in a host, including genes involved in immune evasion that are invisible in vitro.

Two failure modes recur in the breast cancer literature. Screens are run predominantly in the same few unrepresentative lines from Cell lines and their persistent limitations, so the hit list inherits that bias. And screens identify requirement, not druggability. A gene can be absolutely required and have no tractable binding site, which is the distinction What separates a real target from a publishable one formalises.

Combinatorial and single-cell readout screens partly address the first problem by testing interactions and by reading transcriptional consequence rather than survival alone. They do not address the model-choice problem, which is upstream of all of them.

7 · Model choice as a determinant of the conclusion

The claim of this section is stronger than the usual caveat about model limitations. Model choice does not merely constrain how far a conclusion generalises. It frequently determines what the conclusion is.

Consider a drug whose activity depends on antibody-dependent cellular cytotoxicity. In a cell line it is inactive, because there are no effector cells. In an immunodeficient xenograft it is partly active through mouse natural killer cells. In a syngeneic model it is active. Three systems, three answers, one drug. The experiment that gets published is the one that worked, and the model that produced it is reported as a method rather than as a premise.

The same structure applies across the field. A microenvironment-directed agent fails in organoids and works in explants. An endocrine agent fails in the lines that grow well and works in the poorly engrafting models nobody uses. An immunotherapy combination works in the most immunogenic syngeneic line and fails in the least.

8 · Why so much preclinical work fails to translate

The failure is usually attributed to species differences, which is too vague to act on. Four specific mechanisms account for most of it, and each maps to something in this chapter.

The first is population mismatch. Models over-represent triple-negative and high-grade biology because those engraft and grow, while hormone-receptor-positive disease, the majority of the disease, is under-modelled. Preclinical enthusiasm is therefore systematically miscalibrated against the clinical population.

The second is missing compartments. A model without stroma cannot show stromal resistance. A model without an immune system cannot show immune escape. Drugs are selected in systems that lack the mechanisms most likely to defeat them clinically, so the failure appears late.

The third is homogeneity. A cell line is effectively clonal, a xenograft is clonally narrowed by engraftment, and an organoid is narrowed by culture. Each is more uniform than the tumour it came from. A therapy that eliminates a homogeneous population may leave a heterogeneous one substantially intact, which is the argument of Part VI restated as a modelling problem.

The fourth is endpoint mismatch. Preclinical endpoints are tumour volume over weeks. Clinical endpoints are survival over years and depend on residual disease and dormancy, which short-term volume measurements do not capture at all. A drug that shrinks tumours and does not eradicate the dormant population in Dormancy, residual disease, and late recurrence will succeed preclinically and fail in the adjuvant setting. It will do so for reasons the preclinical design could not have detected.

References

All 3 references

1. Wu SZ, Al-Eryani G, Roden DL, et al. A single-cell and spatially resolved atlas of human breast cancers. Nat Genet 2021 53:1334-1347. PMID 34493872
2. Stahl PL, Salmen F, Vickovic S, et al. Visualization and analysis of gene expression in tissue sections by spatial transcriptomics. Science 2016 353:78-82. PMID 27365449
3. Mardamshina M, Karagach S, Mohan V, et al. Integrated spatial proteomic analysis of breast cancer heterogeneity unravels cancer cell phenotypic plasticity. Nat Commun 2025 16:10482. PMID 41290667