For a child diagnosed with neuroblastoma, one of the most common cancers of infancy, the hardest part can begin after the diagnosis: determining what treatment is most likely to work.
Neuroblastoma develops when immature nerve cells grow uncontrollably. But the disease does not behave the same way in every child. Some tumors disappear without treatment, while others spread rapidly and require intensive therapy. Doctors have long sought ways to distinguish between those paths and match each patient with the treatment most likely to help.
One approach has been to look for mutations in individual genes. But cancer is rarely governed by a single genetic switch. A patient’s response to treatment can be shaped by an enormous molecular landscape, including DNA and RNA found in tumors, blood and other tissues.
“It’s much more than just one gene—everything that’s happening in the cells of the patient matters,” says Orly Alter, an associate professor of biomedical engineering at the University of Utah’s Scientific Computing & Imaging Institute and a member of the Huntsman Cancer Institute’s Cancer Control & Population Sciences.
That complexity presents a problem for a technology increasingly being used to analyze biological data: artificial intelligence.
Modern AI and machine-learning systems generally perform best when they can learn from vast numbers of examples. In medicine, however, clinical trials often involve only dozens of patients, while the amount of molecular information collected from each patient can run into the millions or billions of measurements.
The imbalance is particularly stark in genomics. A recent large language model trained on the 30,000-nucleotide genome of the virus that causes Covid-19 required about 110 million samples, according to the researchers. Scaling the same approach to the roughly three-billion-nucleotide human genome would theoretically require about 33 trillion patients.
No such dataset exists, of course. And it underscores a fundamental challenge in applying conventional AI to medicine: There may be vastly more information about each patient than there are patients from whom an algorithm can learn.
Alter and her collaborators have developed a different approach. Drawing on mathematical concepts associated with quantum mechanics, they say they have created an AI and machine-learning technique that can extract useful patterns from extremely high-dimensional biological data even when the number of patients is small.
Their findings, published in Applied Physics Letters Quantum, suggest a possible new route for selecting treatments, identifying drug targets and understanding why patients respond differently to the same therapy.
Looking across the molecular landscape
The researchers’ method is based on a group of algorithms called multitensor comparative spectral decompositions. Alter developed the approach using mathematical concepts related to quantum mechanical entanglement and superposition.
The basic idea is to examine several layers of biological information simultaneously rather than isolating one genetic feature at a time.
A patient’s tumor, for example, contains information in its DNA. Its RNA provides another layer, showing which genetic instructions are being actively expressed. Blood can provide still another molecular snapshot of what is happening in the body.
Alter’s approach attempts to identify relationships among those layers.
“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 expounds. “Even for very few patients, we can still take everything in—their millions to billions of molecular features—and make sense of them. We can, therefore, understand the disease mechanisms and predict drug targets to improve patients’ outcomes. We also validate our AI/ML predictions of targets and outcomes experimentally, which is widely considered a biotechnology holy grail.”
To illustrate the concept, the researchers compare the method to a prism. Rather than treating a patient’s molecular data as an undifferentiated mass, the algorithms separate it into linked patterns that can reveal relationships between different biological layers.
The team tested the technique using publicly available data from children with neuroblastoma.
The algorithms identified two previously unrecognized predictors of life expectancy following treatment. According to the researchers, those predictors consistently performed better than standard biomarkers when tested against tumor and blood DNA as well as tumor RNA.
The findings also held across separate groups of children treated at different hospitals and at different times. That consistency is important because a model that works only in the dataset used to develop it may have little value in actual medicine.
The results suggest that the method could potentially be used beyond a single group of patients, although further validation would be needed before such an approach could guide routine clinical care.
From prediction to treatment
For Alter, one of the potential advantages of the approach is that it does not merely produce a prediction. It can also provide clues about the biology behind it.
“Neural network models are black boxes, but our predictors are interpretable; they point to disease mechanisms and suggest genes to target to sensitize tumors to treatment,” Alter notes.
Her group has also tested predictions involving adult glioblastoma, a particularly aggressive form of brain cancer. In those studies, the researchers used clinical trials and preclinical experiments, including CRISPR-Cas9 gene editing, to investigate predicted patient outcomes and potential drug targets.
That distinction matters in drug development. An algorithm that identifies a statistical correlation may be useful, but researchers ultimately need to establish whether the underlying biological mechanism is real and whether changing it can alter the course of disease.
Alter, who specializes in computational medicine and also holds an adjunct appointment in the University of Utah’s Department of Human Genetics, has taken the work beyond academic research.
Her university spinoff, Prism AI Therapeutics, uses the algorithms and predictors to help biotechnology and pharmaceutical companies identify which patients may be most likely to benefit from clinical trials and which genes could potentially be targeted to improve treatment.
The broader promise is precision medicine that does not stop at dividing patients into increasingly refined categories.
Instead, Alter envisions a system capable of making predictions for a single patient.
“That’s the ultimate precision medicine,” she says. “You have a single person. Can you take the data from just that one person and come up with a treatment for them? I think we can get there.”
For cancer researchers, that would represent a significant shift. Much of modern medicine still relies on evidence drawn from groups of patients, with doctors using those averages to make decisions about individuals whose biology may differ substantially from the study population.
The possibility of analyzing an individual’s full molecular profile could eventually allow treatment decisions to be based on that person’s disease rather than on what tends to work for patients who resemble them.
That future remains a research goal, not a clinical reality. But Alter sees applications beyond cancer and medicine as well.
“The algorithms are completely data agnostic, and there could be endless applications also outside of medicine,” she adds, highlighting sustainable energy as one possibility.
Stay ahead of the Stories shaping our world. Subscribe to Impact Newswire and join our
WhatsApp Channel for updates on global tech, business, and innovation—all in one place.
Dive deeper into the future with the Cause Effect 4.0 Podcast, where we explore the ideas, trends, and technologies driving the global AI conversation.
Got a story to share? Contact Us to reach a global audience with Impact Newswire.
Discover more from Impact AI News
Subscribe to get the latest posts sent to your email.
