Three new AI models from U.S. universities are tackling different challenges in predicting how cancer patients will respond to treatments—from decoding complex genetic data to modeling tumor metabolism in real time.
Genetic Profile Model for Solid Tumors
Researchers at the University of California San Diego developed an AI model named MutationProjector. The model is trained on genomic data from over 30,000 tumors across 10 types of solid cancers. It analyzes the combination of genetic alterations in a tumor to generate a representation of its biological state, with the goal of identifying disrupted molecular pathways and potentially effective treatments.
In independent patient cohorts for bladder cancer, lung cancer, and melanoma, the model matched or exceeded existing methods in predicting response to immunotherapy and chemotherapy. It also identified both known and previously unexpected biomarkers associated with treatment outcomes.
The study was published in Cancer Discovery. Funding was provided by grants from the National Institutes of Health (NIH) and the Advanced Research Projects Agency for Health (ARPA-H).
Quantum-Based Model for Small Patient Cohorts
Researchers at the University of Utah's Scientific Computing & Imaging Institute developed a technique called multitensor comparative spectral decompositions. This method uses concepts from quantum mechanics, such as entanglement and superposition, to analyze multiple layers of molecular data simultaneously from small patient groups (20–100 samples).
The algorithms were tested on open-source neuroblastoma data, identifying two predictors of patient life expectancy. The researchers state these predictors outperformed standard biomarkers across separate groups of children treated at different times and hospitals. The team previously validated predictions for adult glioblastoma outcomes and drug targets using CRISPR-Cas9 in clinical trials and preclinical studies.
The study was published on June 22 in APL Quantum. Funding was provided by the NIH, the National Science Foundation (NSF), and other organizations.
Digital Twin for Brain Tumor Metabolism
Researchers at the University of Michigan developed a machine-learning model described as a "digital twin" that maps real-time tumor metabolism in glioma patients. The model integrates patient data from blood draws, metabolic measurements of tumor tissue, and the tumor's genetic profile. It calculates the metabolic flux, or the speed at which cancer cells process nutrients.
The model’s accuracy was verified through comparisons with data from eight glioma patients and experiments on mice. Researchers reported that the digital twin predicted tumor responses to the drug mycophenolate mofetil, identifying instances where tumors could utilize a "salvage pathway" to resist the drug’s effects.
The study was published in Cell Metabolism. Primary funding was provided by the National Cancer Institute, part of the NIH, with additional support from foundations. Researchers from the University of Alabama, Birmingham, and the Mayo Clinic contributed to the work.