Six million molecular features per patient. Seventy-one children. Two survival predictors that beat the biomarkers oncologists currently rely on. Hold those numbers next to each other for a second, because the gap between them is what a team at the University of Utah just spent a paper closing, and it is a much stranger achievement than the word stamped on the press release.
The word is “quantum,” and it is doing more marketing than science. The study, published this month in APL Quantum, comes from a group led by Orly Alter, a biomedical engineer at Utah’s Scientific Computing and Imaging Institute. Her algorithms, borrowed from the mathematical formalism of quantum mechanics, pulled two new predictors of survival out of one of the nastiest data problems in medicine. The work holds up. But before anyone pictures a refrigerated quantum computer humming in a Utah basement, it helps to know that the math here is older than the headline, and that the absence of any sci-fi machine is the whole point.
Quantum mechanics, underneath the cats and the wave functions, is a particular kind of linear algebra: vectors in an enormous abstract space, operators that split a tangled system into clean, separable patterns. It is a toolkit for finding structure in something that looks hopelessly complicated, and it does not care whether the complicated thing is an electron or a tumor. That indifference is what Alter has been exploiting for two decades.
Neuroblastoma is the most common cancer in infants. It begins when early nerve cells fail to mature and start multiplying, and it behaves like a cruel coin-flip: some tumors regress on their own, others kill despite everything medicine throws at them. The clinical problem is not a shortage of treatment so much as a shortage of certainty about who needs the aggressive version and who can be spared it. Get the prediction right and a child avoids brutal chemotherapy they never needed, or an aggressive tumor gets caught while there is still time.
Here is what makes that prediction so hard. Modern oncology can read millions of molecular features off a single patient: DNA and RNA from the tumor, DNA from the blood, layer on layer of genomic detail. Alter’s team took in roughly six million such features per analysis. The trouble is that conventional machine learning, the kind powering the AI everyone is breathless about, is starving for examples when the features outnumber the patients this badly. Alter’s own comparison is vivid: to learn responsibly from that many features the usual way, you would need an estimated 33 trillion patients. Nobody has 33 trillion of anything. The team had 71.
The math is what bridges that distance. Alter’s algorithms, which she calls “multitensor comparative spectral decompositions,” lean on the quantum ideas of superposition and entanglement, the principle that information lives in the relationships between layers of a system rather than in any single layer. “Our quantum approach allows us to find the relevant information in every layer of the data, for example, from the patients’ blood in addition to their tumors,” Alter said. In plainer terms: instead of searching the blood DNA, tumor DNA, and tumor RNA as three separate piles, the method treats them as one entangled object and surfaces the patterns that only appear when you look across all of them at once. That cross-layer view is what lets 71 children carry an analysis that would otherwise drown a black-box model.
The two predictors it found outperformed standard biomarkers across tumor and blood DNA and tumor RNA. They also did the thing that separates a finding from a fluke: they held up on entirely separate groups of children treated at different hospitals in different years. A model that only works on the data it was born from is a parlor trick. One that survives a fresh cohort is a result, and this one cleared that bar.
It also did something most medical AI cannot. The predictors are interpretable: they do not just emit a risk score, they point to specific disease mechanisms and name genes you might target to make a tumor more sensitive to treatment. Alter’s group then went further and tested some of those predictions in the lab, in adult glioblastoma, using CRISPR-Cas9 to check whether the genes the math flagged actually mattered. That is the reverse of the reflex dominating the field right now, where a model hands down a verdict and nobody, including its builders, can explain it. A model that hands you a mechanism and a drug target is worth more than ten that hand you a number and a shrug.
So why dwell on the branding? Because the branding is where the honesty gets tested, and this is not quantum computing. There are no qubits, no exotic processor, no quantum hardware anywhere in the pipeline. It is classical linear algebra, run on ordinary machines, that happens to use the same spectral mathematics physicists use to describe quantum systems. Alter has been applying this exact family of matrix and tensor decompositions to genomics since long before “quantum” became the word that loosens grant committees’ and investors’ wallets. The substance came first; the label arrived later. Calling it a “quantum approach” is accurate in the narrow technical sense and quietly misleading to anyone who reads the headline and pictures a machine from a movie curing cancer. The actual lesson runs the other way. You do not need the machine. You need clever math and a workable number of real patients.
A few things keep this in the territory of the promising rather than the proven, and they are the reason the next step matters, not fine print to bury. The cohort is 71 children studied after the fact, not a prospective trial that changed anyone’s treatment while it was happening, and retrospective predictors have a long record of looking sharper on the bench than at the bedside. The funding is the clean kind, public money from the NIH and NSF stacked with childhood-cancer charities like Alex’s Lemonade Stand and St. Baldrick’s, not pharma paying for a result it ordered in advance. But Alter has also founded a spinoff, Prism AI Therapeutics, that sells these algorithms to biotech and drug companies, and that is a commercial stake worth naming plainly even when the underlying work looks solid.
What to watch is straightforward. A predictor proves itself by surviving outside the lab that built it, in a prospective setting where a doctor acts on it and a child does better than they otherwise would have. The math has earned that trial. The job now is to run it, and to keep the word “quantum” from doing work the data still has to do itself.
Sources
- APL Quantum – Alter et al., “Quantum mechanics-based multitensor AI/ML uniquely able to discover, validate, and interpret predictors from small-cohort noisy high-dimensional multiomic data” (2026)
- News-Medical – “Using the mathematics of quantum mechanics to improve neuroblastoma outcomes”
- Scilight (AIP Publishing) – “A quantum mechanics approach to artificial intelligence can improve cancer outcomes”