A proprietary ranker takes living tumor tissue to the clinic

What the work claims

The claim is that drug sensitivity measured ex vivo on a patient's own tumor cells, run through a quadratic phenotypic optimization model, can rank treatment options, including combinations, well enough to direct therapy after standard options fail. Two kinds of work are bundled here, and they should be weighted differently.¹

The first is the Q-SAM trial itself: an observational proof-of-concept study, registered as NCT04986748, sponsored by the National Cancer Centre Singapore, recruiting, with an estimated 100 patients across sarcoma and melanoma cohorts. Patients undergo resections or biopsies as part of standard care; from that material the team derives two- and three-dimensional patient models; the models are screened against a panel of up to 14 drugs; patients meanwhile receive standard-of-care treatment, and QPOP reports are generated to assess feasibility, radiological response rates and survival outcomes. The design is explicitly observational: the trial registers what the platform does, it does not randomize patients against physician choice.¹

The second is a peer-reviewed cohort study from the same Singapore ecosystem, published in npj Precision Oncology in 2025, which applied the QPOP protocol to 45 primary soft tissue sarcoma samples and reported a concordance analysis between QPOP-defined responders and actual clinical outcomes.² That paper is the strongest evidence behind the trial, and it is feasibility evidence, not efficacy evidence.

How it works

QPOP stands for quadratic phenotypic optimization platform. The term of art matters: the platform is mechanism-agnostic. It does not sequence the tumor looking for driver mutations, because most soft tissue sarcomas have no actionable mutation to find. Instead it measures phenotype directly: how many cells survive each drug exposure.²

The experimental design is the clever part. A full combination screen of even a dozen drugs at several doses is combinatorially impossible on a small biopsy, so the group uses an orthogonal array composite design: a predesigned set of 155 test combinations, run at three low dose levels, that is sufficient to fit a second-order regression model of cell viability over the whole drug landscape. The model then interpolates the response of every possible combination from the drug set, including ones never tested, and ranks them. In the sarcoma study this meant a 12-drug panel run in technical duplicates, with viability read out 48 hours after treatment, and a personalized report delivered to the treating clinician within about a week of sample collection.²

The trial adds the three-dimensional piece: patient-derived organoids and other 3D models alongside 2D cultures, screened against the institutional panel, with the model-generation success rate per tumor subtype as the first primary outcome. That is a platform metric, not a biology result. It asks the question any vendor or hospital must answer before this can be a service: what fraction of patients who walk in actually get a usable answer?¹

Where a skeptic should push

Push first on the size of the clinical concordance claim. Of 51 samples collected in the sarcoma study, 6 were excluded, three for insufficient material and three for failing quality control on Z-prime scores, leaving 45 samples with reports, a success rate of 88.2 percent. But the concordance analysis rests on 14 evaluable patients with 27 treatment outcomes. On that base, the receiver-operating curve had an area of 0.769 with a 95 percent confidence interval running from 0.585 to 0.954, which is wide enough to include genuinely poor discrimination; the total predictive value was 77.8 percent, and the odds ratio for partial response or stable disease in QPOP-defined responders was 13.5 with a confidence interval of 2.07 to 73.3.² Every one of those intervals is doing heavy lifting, and none of them comes from a randomized comparison.

Push second on what the assay omits. A 48-hour viability readout on dissociated cells says nothing about drug penetration in a human tumor, nothing about the immune system's contribution, nothing about metabolism or toxicity at tolerated doses, and nothing about the stromal microenvironment that sarcomas in particular depend on. The paper's own strongest case is retrospective concordance and a handful of striking anecdotal responses, including a solitary fibrous tumor patient who outlived a one-year prognosis after QPOP redirected treatment twice.² Anecdotes of that quality are worth reporting. They are not a denominator.

The single most load-bearing assumption is that a second-order surface fitted to 155 wells predicts in vivo combination behavior. The paper itself found that single-drug sensitivity correlated with combination ranking for some agents but not for others, meaning the interaction term, the part the model is really selling, is exactly the part least constrained by data.²

Finally, push on reproducibility. The paper's data availability statement is blunt: the QPOP analysis code is unavailable because the software, Optim.AI, is proprietary to KYAN Technologies.² An independent laboratory can repeat the wet-lab protocol, and indeed the drug panel, doses and array design are all published. What it cannot do is rerun, audit or stress-test the ranking itself, which is the clinical product.

When the treatment ranker is a trade secret

For platform access, Q-SAM is a working template of the next access fight. The wet-lab half of the stack, tissue handling, drug dispensing, viability readout, is fully specified in the open literature. The decision half, the ranker that converts measurements into a treatment order, is closed. That inversion matters because it reverses where the leverage sits. A hospital can build the assay and still be wholly dependent on a vendor relationship for the thing the patient actually receives, a ranked list of options. Access to the platform becomes access to a license, with pricing, service levels and upgrade cycles that the published science cannot discipline. The 88 percent model-generation rate is the honest access metric here: about one patient in eight fails to get any answer at all, before any question of who can afford one is asked.²

For vendor capability, KYAN Technologies' position shows how a thin layer becomes a moat. The mathematics underneath, second-order regression over a designed array, is textbook. The moat is not the math but the accumulated calibration: the choice of dose levels, the cutoff values, the interface habits, and the clinical credibility earned one report at a time. Any incumbent in organoid platforms should study this, because the same shape is coming to their market. Read-out hardware, electrodes, perfusion chips and culture reagents will commoditize exactly the way the QPOP wet-lab half did. The durable vendor asset will be the closed model that interprets the readout, and the regulatory goodwill around it.

The ethics and governance question is accountability diffusion. Q-SAM is observational by design: QPOP reports are generated alongside standard-of-care treatment, and the registry language frames them as support for physician selection, not commands.¹ That is a sensible clinical-safety posture and a liability structure worth watching at the same time. When a ranked regimen fails, where does responsibility sit? The physician who followed the ranking, the vendor whose model produced it, the trial that validated the concordance, or the institution that bought the service? The proprietary code makes the question sharper: no external party can examine whether a failure was an assay artifact, a model miss, or a correct prediction of ex vivo behavior that simply does not transfer to the patient. Regulators have not yet had to answer this for a closed algorithm directing therapy on living tissue, and Q-SAM's observational wrapper means they may not have to for years.

The parallel to organoid intelligence platforms is direct, even though this trial concerns tumor tissue. The architecture is identical: living tissue as the assay substrate, a standardized perturbation panel as the input alphabet, a closed optimization layer that converts response into a ranked decision, and a clinician or operator who is formally free to ignore it. Every organoid-computing vendor proposing to measure what neural tissue has learned, and to steer its training, will ship this exact stack. Q-SAM is the dress rehearsal of its governance: who audits the ranker, who owns the tissue-derived data that tunes it, and what recourse exists when the ranking is wrong. The field should want those answered for sarcoma before they arrive for brain tissue, where the substrate cannot be discarded after a failed recommendation.

The bottom line

Established: the ex vivo QPOP workflow is feasible at cohort scale, with 45 of 51 sarcoma samples yielding reports in about a week, published concordance between rankings and outcomes in a small evaluable set, and a registered trial extending the approach to organoid models and to prospective melanoma and sarcoma cohorts.¹²

Not established: that QPOP-directed therapy improves survival. There is no randomized or prospectively controlled evidence of clinical benefit as of September 2026, the concordance intervals are wide, and the ranking software is closed to independent verification.

What would confirm the claim: Q-SAM's own primary outcomes, model-generation rates per subtype and radiological response rates, maturing toward the estimated 2028 completion, ideally followed by a randomized QPOP-guided versus physician-choice comparison with survival endpoints. What would break it: a maturing cohort in which the predictive value collapses toward chance, or a regulatory finding that the proprietary ranker cannot be validated as a decision-support device. Either outcome will set terms the organoid platform market inherits.