Cancer Research · global
Reading Prognosis from Tumor Methylation: AI Seeks Clues to Reducing Radiotherapy Doses for Children with Brain Cancer
MANTIS uses molecular and follow-up data from 2,786 patients to generate personalized survival predictions and retrospectively examine which children might tolerate lower-dose radiotherapy. Independent data support its predictions, but whether they can safely guide treatment adjustments still requires prospective trials.
For children with malignant brain tumors, the costs of treatment can extend into adulthood. Radiation helps control cancer but can also damage the developing brain, affecting learning and cognition. Reducing the risk of recurrence while avoiding unnecessary treatment burdens has long been a challenge in pediatric oncology. A new study attempts to find more refined grounds for these decisions in the molecular signals of tumors themselves.
Published on November 3 in *Nature Genetics*, the study integrated DNA methylation and clinical outcome data from 2,786 patients with medulloblastoma or ependymoma to develop an artificial intelligence model called MANTIS. The research was a collaboration among Dana-Farber Cancer Institute and other institutions, with Sabina Stefan as first author and Volker Hovestadt as senior author. The model aims to move beyond broad risk groups to map changes in each patient's probability of survival over the ten years following diagnosis.
DNA methylation consists of chemical marks attached to DNA. Their distribution can reflect a tumor's biological characteristics and is already used to help classify brain tumors. MANTIS extracts additional prognostic clues from these test data; the medulloblastoma version also incorporates gene copy number changes and metastatic status at diagnosis. An accompanying research briefing in *Nature Genetics* also explains this integration. Dana-Farber noted that reusing information generated by existing diagnostic tests could allow the same test to provide more clinical value in the future.
The reliability of predictions cannot be assessed solely from the data used to build a model. The researchers also validated MANTIS using data from 898 patients with medulloblastoma across three independent clinical trials and tested the corresponding model in an independent ependymoma cohort. The concordance index reported in the paper and by *The ASCO Post* reached as high as 0.80 ± 0.07. This metric measures how closely predicted survival rankings match actual rankings; it cannot be interpreted as meaning that each patient's predicted probability of survival is 80% accurate.
An exploration more closely related to treatment decisions drew on the randomized ACNS0331 trial. The researchers reanalyzed data from 170 patients aged 3 to 7 with average-risk medulloblastoma, comparing craniospinal irradiation at 18 Gy and 23.4 Gy. The original trial's overall results showed worse disease outcomes in the lower-dose group, so doses cannot be reduced across the board for these patients. The team extended MANTIS into a treatment-effect model to try to identify how individual children respond to changes in dose.
Among the 104 patients selected by the model, representing approximately 60% of those analyzed, the existing data did not show significantly worse survival in the lower-dose group; the P value for the comparison between groups was 0.683. However, “no statistically significant difference” does not establish that the two doses are equally safe, nor is it sufficient to establish that all these children can receive reduced doses. This was an exploratory analysis of past trial data. No prospective trial has yet assigned treatment according to the model to validate its judgments.
MANTIS currently offers clues for designing the next trials. The researchers noted that it remains necessary to determine how its predictions should be interpreted clinically and whether using the model actually improves patient outcomes. It also does not yet account for late risks such as second cancers after treatment. For molecular testing to advance from identifying tumors to helping select treatment intensity, the key will be to demonstrate that reducing harm does not come at the cost of more recurrences or deaths.