Research analysis · Drug discovery

An organoid drug screen that has to survive the mouse

A Huntsman Cancer Institute team converted patient-derived xenografts into stable organoid cultures, screened them against 45 drugs, and then checked the predictions back in the matching mice. Nearly every claim about organoid-based precision oncology is quieter than this one, and the paper itself documents where even this setup goes blind.

Source: A human breast cancer-derived xenograft and organoid platform for drug discovery and precision oncology, Nature Cancer, 2022. Primary source (open access). Read in full: the article, its supplementary figure legends, the companion protocol paper, and the current NIH RePORTER abstract for the continuing U54 award.

What the work claims

Guillen, Fujita, Butterfield and colleagues report a bank of human breast cancer patient-derived xenografts (PDXs) and matched organoid cultures derived from them (PDxOs), built from the tumors with the worst prognosis: endocrine-resistant, treatment-refractory, and metastatic disease1. The central claim is twofold. First, stable PDxO lines can be screened in 384-well plates at moderate throughput while retaining the genomic features of the patient tumor. Second, and more load-bearing, drug responses measured in the dish track responses in the matching PDX in vivo, so a positive organoid result can be forwarded to an animal before anyone talks about a patient.

In one clinical case, a patient with triple-negative breast cancer and early metastatic recurrence had no actionable mutation on commercial genomic profiling. Screening her pretreatment tumor organoid nominated the FDA-approved drug eribulin; she was treated with it and achieved a complete radiographic response of her liver metastases and ascites, with progression-free survival of 138 days versus 41 days on her prior therapy, a 3.5-fold ratio (time to next therapy was 4.8-fold longer, 197 versus 41 days)1. The paper is explicit that this is a single case, not a trial result.

How it works

The pipeline starts in the mouse, not the dish. Tumor tissue is engrafted into immunocompromised mice to establish a PDX, and only then is a PDxO line grown out. The authors derived 40 PDxO lines from 47 attempts, an 85 percent take rate, and have cultured them for more than 200 days with doubling times of roughly three to eight days1. A companion paper turned the method into a protocol, including the mouse-cell removal step and the validation assays2.

For screening, 16 PDxO lines were plated in 384-well format and exposed to 45 compounds in an eight-point dose response, in technical quadruplicate and biological triplicate, over a four-day assay1. Because doubling times vary between lines, raw viability scores would confound comparison, so the authors used growth-rate inhibition metrics (GR50 and the area over the curve) that correct for how fast each model grows3. This matters: without the correction, a slowly growing tumor can look artificially resistant to a cytostatic drug.

The validation loop is the engineering contribution. When the screen found that six of twelve TNBC PDxO lines were markedly sensitive to the SMAC mimetic birinapant, the authors treated seven PDX lines in vivo that the organoid data predicted would span the sensitivity range. The three lines predicted resistant (HCI-001, HCI-002, HCI-019) progressed like controls; the three predicted sensitive (HCI-015, HCI-023, HCI-027) shrank; one intermediate line showed initial shrinkage followed by regrowth1. That is a prospective, blinded-in-spirit concordance test between dish and animal, which is more than most organoid screening papers attempt.

Where a skeptic should push

The single most load-bearing assumption is that a four-day viability readout captures the biology of the drugs clinicians actually use. The paper's own data cut against it. Response to fulvestrant, a standard endocrine therapy, was difficult to discern in the short PDxO assay; 4-hydroxytamoxifen responses did not track estrogen-receptor status; and CDK4/6 inhibitors showed no cytotoxic effect in four days, consistent with PDX tumors needing 10 to 30 days to show a cytostatic response1. The authors conclude plainly that the four-day assay is best at finding cytotoxic drugs. A screen that structurally cannot see cytostasis will systematically over-nominate cytotoxics and under-nominate the drugs that control chronic cancers.

Second, the funnel is narrow and it selects for aggressive disease. Only about 30 percent of breast cancers engraft as PDX, and the engraftable fraction is enriched for tumors that later metastasize1; of those, 85 percent yield a PDxO. Every concordance number in the paper describes this pre-filtered population. That is arguably the right population for metastatic drug development, but it is a specific, selected slice of breast cancer, and the paper is careful not to claim otherwise.

Third, the microenvironment is gone twice over. Human stroma is replaced by mouse stroma during PDX growth, and mouse cells are then removed during organoid propagation, leaving no human immune compartment and no native stroma1. The authors note this limits which drug classes the system can test at all. And the clinical case, however striking, is one patient with no counterfactual; the comparison to the MOSCATO-01 benchmark ratio of 1.3 is offered by the authors themselves as context, not evidence1.

Finally, the scale-up claim deserves separate bookkeeping. The continuing U54 award abstract reports nearly 100 PDxO models across all breast cancer subtypes screened with 40 to 50 drugs each, with high organoid-to-PDX concordance4. That figure comes from a grant abstract, not a peer-reviewed table; treat it as a progress report, not a result.

What this changes for organoid drug screening credibility

The non-obvious lesson is that the organoid is not the product; the validation loop is. This program buys credibility by submitting every interesting dish result to a matched in vivo check before making a claim, and by quantifying the assay window (what a four-day readout can and cannot see). For anyone building drug discovery on patient-derived organoids of any organ, that is the transferable blueprint: decide in advance which classes of drug your readout can detect, validate a panel of predictions in a second, orthogonal system, and report the funnel honestly, from patient tumor to engraftment to organoid take rate.

The opportunity is a maturing quality bar. Growth-rate corrected metrics, replicate structure, and an explicit concordance experiment are cheap to adopt and would sharply raise the evidentiary value of the average PDO screen.

The threat is subtler. Short-term viability assays bias the entire field toward cytotoxic hits, which is exactly the drug class where organoids, lacking stroma and immune context, are most likely to agree with simple models and least likely to add information over existing methods. A field that standardizes on four-day cytotoxicity screens will keep publishing concordance numbers while systematically missing the cytostatic, microenvironment-dependent biology that determines most real treatment outcomes. The deeper caution from this paper is that its most convincing result, a single complete response, came from a platform that its own authors show is blind to several major drug classes; extrapolating from one well-validated corner of assay space to "organoids predict therapy response" is the generalization failure to guard against.

The bottom line

Established: stable PDxO lines retain driver mutations over long culture, a 45-drug growth-rate-corrected screen is reproducible, and organoid birinapant predictions matched PDX outcomes in seven of seven tested lines. Suggestive: the eribulin case, with its 3.5-fold PFS ratio, is hypothesis-generating and the authors say so. Not established: that direct-from-patient organoids (without the PDX intermediate) behave the same way, that short-term assays can rank cytostatic drugs, or that concordance in engraftable, aggressive tumors generalizes to the broader patient population. What would confirm the platform: a prospective trial where PDxO or PDO results alter treatment and beat physician choice, and dose-response data with exposure times matched to drug mechanism. What would break it: repeated failures of cytostatic or microenvironment-dependent drugs that looked inert in the dish but worked in patients.

Frequently asked questions

What is a PDxO, and how is it different from a PDO?

A PDxO is an organoid grown from a patient-derived xenograft, a tumor that has already passed through a mouse. A PDO is grown directly from the patient tumor. The PDX intermediate expands aggressive tumor cells and replaces human stroma with mouse stroma, so PDxOs are more homogeneous but one step further from the patient.

How many models and drugs were actually tested?

In the peer-reviewed paper, 16 PDxO lines were screened against 45 compounds in an eight-point dose response, in quadruplicate technical and triplicate biological replicates, over four days. The continuing grant abstract reports a larger in-house dataset of nearly 100 PDxO models with 40 to 50 drugs each, which has not been published in full.

Did the organoid screen help a real patient?

In one documented case, yes. A patient with metastatic triple-negative breast cancer and no actionable genomic finding was treated with eribulin after her organoid screen nominated it, and her liver metastases and ascites resolved completely. Her progression-free survival on that therapy was 138 days versus 41 days on the prior therapy. This is a single case without a control arm, so it shows feasibility, not efficacy.

What can a four-day organoid drug assay not detect?

According to the paper itself, mainly cytostatic effects. Fulvestrant responses were hard to discern, 4-hydroxytamoxifen responses did not track estrogen receptor status, and CDK4/6 inhibitors showed no effect in four days even though the same models respond over 10 to 30 days in mice. The assay is best suited to cytotoxic drugs.

Why does growth-rate correction matter in organoid screens?

Models with different doubling times convert the same drug effect into different viability numbers. Growth-rate inhibition metrics, developed by Hafner and colleagues, normalize for proliferation so sensitivity can be compared across lines. Without them, slow-growing tumors can be misclassified as resistant.

How should a drug-discovery team reuse this work?

Borrow the discipline, not just the protocol: predefine which drug classes your readout can see, validate a subset of predictions in an orthogonal in vivo or tissue-level system, correct for growth rate, and publish the attrition funnel from patient sample to screenable model. Those steps, more than any single culture condition, are what make a screen interpretable.

References

  1. Guillen KP, Fujita M, Butterfield AJ, et al. A human breast cancer-derived xenograft and organoid platform for drug discovery and precision oncology. Nature Cancer. 2022;3(2):232-250. PMC8882468. Accessed 2026-09-13.
  2. Scherer SD, Zhao L, Butterfield AJ, et al. Breast cancer PDxO cultures for drug discovery and functional precision oncology. STAR Protocols. 2023;4(3):102402. doi:10.1016/j.xpro.2023.102402. Accessed 2026-09-13.
  3. Hafner M, Niepel M, Chung M, Sorger PK. Growth rate inhibition metrics correct for confounders in measuring sensitivity to cancer drugs. Nature Methods. 2016;13:521-527. doi:10.1038/nmeth.3853. Accessed 2026-09-13.
  4. Welm AL. Research Project 2: Identify and validate efficacious therapies for metastatic breast cancer (U54 CA224076, PDX Trial Center). NIH RePORTER. Project record. Accessed 2026-09-13.