A research team at the University of Utah has developed an artificial intelligence and machine learning method based on quantum mechanics. This will improve the prediction of cancer outcomes and allow the use of an individual patient’s comprehensive molecular background to identify therapeutic targets. The approach described in APL quantumaddresses a major hurdle in leveraging traditional AI to predict patient outcomes in clinical trials: the sheer volume of data needed to train large language models and account for the complexity of disease factors.
“It’s not just one gene; everything that’s happening inside a patient’s cells is important,” said Ory Alter, Ph.D., associate professor of biomedical engineering at the University of Utah Institute for Scientific Computing and Imaging. To take this into account, the research team developed a method that can simultaneously analyze multiple layers of molecular information, including tumor DNA, blood DNA, and tumor RNA.
While clinical trials can enroll as few as 20 to 100 patients, existing genomic datasets often contain data detailing millions to billions of molecular features. According to the researchers, many existing AI and machine learning methods require more patient samples than genetic characteristics to properly train their models. For example, they pointed to a recent large-scale language model of the 30,000-nucleotide genome of the COVID-19 virus, which required 110 million samples. Extrapolating from this, the Utah team said a complete modeling of the human genome’s 3 billion nucleotides would require 33 trillion patient samples.
To overcome this limitation, the researchers used a series of algorithms known as multi-tensor comparative spectral decomposition. This algorithm was developed by Alter based on the concepts of entanglement and superposition from quantum mechanics. The researchers say the results are similar to a prism splitting light into individual color components, yielding data on multiple layers of a patient’s molecular makeup, including tumor and blood genomes and RNA transcriptomics, and demonstrating associated patterns in cancer that can predict outcomes for individual patients.
“This model rewrites a series of multiple omic profiles from a single patient as a superposition of phenotypes, with each phenotype represented by a set of multiple intertwined patterns,” the researchers wrote. Importantly, data from one molecular profile can be approximated to analyzes from other profiles, thus maintaining predictive consistency across different types of biological data.
The researchers tested their model using an open-source dataset of the childhood cancer neuroblastoma. Their analysis found two previously unrecognized predictors of survival and treatment response. Each predictive element was found in three separate but interrelated data types: tumor genome, blood genome, and tumor transcriptome. This study found that these predictors were superior to the currently used biomarker, the MYCN gene, in predicting treatment response and outcome.
The new method is based on substantial research work by Alter and colleagues. Previous studies in this area have used related comparative spectral decomposition methods to analyze genomic and transcriptomic data from other tumor types, including glioblastoma.
The team will continue to work towards developing an approach that can be used in clinical practice. “This is the ultimate in precision medicine,” Alter said. “You have one person. Can you take data from that one person and come up with a treatment for that person? I think we can accomplish that.”
