Biomedical Research · global
Which Children With Brain Tumors Could Receive Less Radiotherapy? An AI Model Looks for Clues in Tumor Molecular Features
MANTIS analyzed data from 2,786 patients to try to identify those with medulloblastoma who might be suitable for lower radiation doses. The predictions offer hope of reducing long-term cognitive damage, but whether treatment can safely be adjusted on this basis still requires confirmation in prospective trials.
For children with malignant brain tumors, the costs of treatment may persist after they return to school: radiation helps control cancer but can also affect learning and cognitive development. Identifying patients who can receive less intensive treatment without compromising their chances of survival has long been a challenge in pediatric oncology. MANTIS, an artificial intelligence model newly published in *Nature Genetics*, seeks to provide a more detailed basis for this decision using tumors’ molecular features.
The study was conducted by an international team led by Dana-Farber Cancer Institute in the United States, with Sabina Stefan as first author and Volker Hovestadt as senior author. The team analyzed data from 2,786 patients to develop a survival prediction framework for medulloblastoma and ependymoma. Rather than assigning patients to a few risk groups, MANTIS aims to map how each individual’s probability of survival changes over the next 10 years.
The model uses clues including DNA methylation patterns and gains or losses in DNA copy number. The former reflect the distribution of chemical marks across the genome, while the latter show whether portions of genetic material have increased or are missing. The researchers combined these molecular data with long-term follow-up results to learn the relationship between tumor biology and patient outcomes. Dana-Farber noted that relevant methylation testing is already used to diagnose brain tumors at some medical centers, giving this approach the potential to extract more information from existing tests.
The study validated predictive performance in independent clinical trial cohorts, with a concordance index reaching 0.80 ± 0.07, outperforming the current clinical measures included in the comparison. This value measures whether the model can correctly rank patients’ survival risks; it cannot be interpreted directly as meaning that “the prognosis of 80% of patients can be accurately predicted.” It supports the model’s ability to distinguish risk, but does not yet answer whether changing treatment according to its predictions would lead to better patient outcomes.
The exploration with the greatest clinical significance was extending MANTIS to estimate the treatment effects of different radiation doses. According to the paper’s abstract published on the Broad Institute website, the analysis identified approximately 60% of average-risk medulloblastoma patients as showing no statistically significant survival disadvantage when receiving lower-dose craniospinal irradiation (P = 0.683). This type of irradiation covers the brain and spinal cord; in the data analyzed, lower doses were also associated with better neurocognitive outcomes. This finding suggesting the potential for dose reduction concerns medulloblastoma and cannot be directly extended into treatment recommendations for ependymoma.
However, “no significant difference detected” does not mean that the two doses have been proven equally safe. This work modeled treatment effects using existing data; it was not a clinical trial in which AI first selected patients and a dose-reduction strategy was then prospectively validated. The model identifies a population worth testing further. Whether cognitive damage can be reduced while preserving chances of survival still requires prospective clinical studies with model-based stratification and long-term follow-up.
The significance of MANTIS lies in its attempt to make differences within the same risk group visible, creating an opportunity to align treatment intensity more closely with the biology of individual tumors. In moving from a prediction tool to clinical decision-making, the crucial next step will be validation: determining which children can safely receive less radiation, and how much that reduction can improve their future learning and daily lives.