Quantum Mechanics Principles Applied to AI Framework Identifies Neuroblastoma Predictors
A quantum mechanics–based multitensor AI and machine learning (ML) framework discovered, validated, and interpreted two predictors in neuroblastoma that, when combined, were consistently more accurate than the best standard-of-care biomarker, tumor MYCN amplification, for predicting survival in patients with neuroblastoma, according to a study published in APL Quantum. The predictors could also be used as potential drug targets for treating patients.
The study authors, led by Orly Alter, PhD, a Utah Science, Technology, and Research (USTAR) Associate Professor of Bioengineering and Human Genetics at the Scientific Computing and Imaging Institute and the Huntsman Cancer Institute at the University of Utah, Salt Lake City, wrote that each predictor had three entangled representations in the tumor and blood genomes and tumor transcriptome, where the result of measuring any one representation approximately determines the results of measuring the other two.
In their view, the framework can solve the challenges posed by the noisy, high-dimensional real-world multiomic data of small cohorts for more precise and individualized treatment tailoring, even in smaller numbers from clinical trials, while AI usually requires larger datasets for training.
Study and Model Details
The investigators introduced what they described as a unified framework that generalizes their previously developed exact and structure-preserving (ie, lossless) “comparative spectral decompositions” to multiple tensors, or multidimensional data structures for storing and processing information, for modeling real-world, interrelated data. The framework draws on the quantum mechanical concepts of superposition, in which data can reflect multiple patterns simultaneously, and entanglement, in which patterns across different data layers are treated as inherently linked (see Sidebar). With these concepts, the AI framework was able to come together into a unique, cohesive model for reproducible results.
According to the investigators, the modeling rewrites each set of one-to-one mapped samples, such as data from a single patient, as a superposition of states, with each state represented by a set of entangled patterns, one per data layer. They reported that the patterns are used to derive predictors and that entanglement means the result of measuring any one representation approximately determines the results of measuring the others, allowing prediction based on one data layer to remain consistent with prediction based on the other layers.
The discovery data set comprised 101 patient-matched tumor and blood whole-genome sequencing (WGS) profiles and, within that cohort, 71 matched tumor RNA sequencing profiles from patients with neuroblastoma in the Therapeutically Applicable Research to Generate Effective Treatments (TARGET) initiative. Validation was performed using 419 patient-matched tumor and blood target-capture sequencing (TCS) profiles from TARGET that were mutually exclusive of the 101-patient discovery cohort.
Key Findings
Among the 90 discovery patients with all labels available, the first predictor defined two groups with a 131-month Kaplan-Meier median overall survival difference (univariate Cox hazard ratio [HR] = 2.9; concordance index = 0.77; log-rank P = 6.0 × 10-4), and the second defined two groups with a 114-month difference (HR = 3.1; concordance index = 0.74; log-rank P = 2.3 × 10-4).
Because the predictors were found to be statistically independent, combining them increased the prognostic performance, yielding a greater (136-month) median survival difference plus lower log-rank P and Akaike information criterion values.
The investigators reported that the combined predictors were statistically better than and independent of all standard-of-care indicators, including International Neuroblastoma Staging System stage, age, MYCN status, histopathology, DNA ploidy, Children’s Oncology Group risk, and the Mitosis-Karyorrhexis Index.
“The predictors encompass the whole multiome, and that is one reason why, in all available examples, they outperform all other biomarkers,” Alter et al wrote.
The predictors derived from the tumor and blood WGS profiles of the discovery set appeared to perform consistently well in the TCS profiles of the validation set. “In both the tumor and blood DNA profiles, measured by both the WGS and TCS platforms, and in both the discovery and validation sets of patients,” the investigators added, “the two predictors combined are at 73% [to] 80% concordance with survival and are consistently more accurate than MYCN [status].”
Beyond prognostic performance, the investigators interpreted the predictors in the context of neuroblastoma biology. One predictor includes co-amplification of MYCN with druggable targets previously unrecognized in neuroblastoma, whereas the other includes ID2 amplification and describes an earlier stage in neuroblastoma development, the authors wrote. The analyses also identified candidate drivers including AAMDC, ADAM17, and NBPF10. Taken together, the investigators noted that the predictors describe “independent, yet complementary, known and new disease mechanisms and druggable targets….”
Alter et al concluded: “Our quantum mechanics–based multitensor AI/ML is uniquely able to discover, validate, and interpret predictors from small-cohort, noisy, and high-dimensional multiomic data.”
DISCLOSURES: The work was partially funded by the National Cancer Institute (NCI) and USTAR Initiative support to Dr. Alter; a National Science Foundation (NSF) grant to Dr. Newman.; Alex’s Lemonade Stand Foundation, the Rally Foundation, and the St. Baldrick’s Foundation in partnership with Griffin’s Guardians to Dr. Tsai.; and the Musella Foundation in partnership with StacheStrong to Drs. Tsai and Alter. Dr. Alter is a co-founder and an equity holder in Prism AI Therapeutics, Inc. For other study author disclosures, as well as data and code availability, visit pubs.aip.org.
Sidebar
Relevant Quantum Mechanics Terminology
Multitensor: A data structure that organizes multiple layers of patient molecular data—such as tumor DNA, blood DNA, and tumor RNA—into an interconnected format where each layer is linked by shared patient samples but contains different types of molecular features.
Superposition: A quantum physics principle in which a system can exist in multiple states at once. In this algorithm, it means a patient's molecular data can reflect multiple overlapping patterns simultaneously, and the method identifies which combination of patterns best explains their clinical outcome.
Entanglement: A quantum physics principle in which two systems are linked so that understanding one reveals information about the other. In this algorithm, it means patterns discovered in one data layer (eg, blood DNA) are mathematically tied to patterns in another (eg, tumor RNA), allowing the layers to inform each other.
Lossless: Refers to an analytical method that retains all of the original data rather than compressing or simplifying it. Many AI approaches reduce complex datasets to make them computationally manageable, potentially discarding information that could be clinically relevant.
Multitensor comparative spectral decompositions: The suite of algorithms at the core of this research. Like a prism splitting white light into its component colors, these algorithms break down interconnected layers of molecular data into distinct, interpretable patterns—without discarding any information—to identify predictors of patient outcomes and potential drug targets.
ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.