AI Reads Tumor Methylation to Forecast Childhood Brain Tumor Outcomes—and Flag Who Can Safely Get Less Radiation

Children with the two most common malignant brain tumors, medulloblastoma and ependymoma, face a painful trade-off. Radiation and chemotherapy save lives, but craniospinal irradiation (CSI) in particular can cause lasting cognitive decline in a developing brain. A new artificial intelligence tool, reported in Nature Genetics, uses a tumor's DNA methylation profile to predict an individual child's survival odds over ten years—and to estimate who might do just as well with less intensive radiation.

The work was led by Volker Hovestadt at the Dana-Farber Cancer Institute, with first author Sabina Stefan, and involved collaborators across several institutions and clinical trial groups.

The tool, called MANTIS (methylation profiling and artificial neural networks for time-dependent individualized survival predictions), was built and evaluated on genome-wide DNA methylation array data from 2,786 patients. Instead of a single risk label, it produces a personalized survival curve spanning a decade after diagnosis.

What the study showed

MANTIS was trained on public datasets and then tested on patients from three independent clinical trials (SJMB03, ACNS0331 and ACNS0332), which were held back from model development. In average-risk medulloblastoma patients who received standard-dose CSI, the model reached a concordance index—a measure of how well predicted rankings match observed outcomes—as high as 0.80. It also discriminated outcomes better than existing subgroup-based and clinico-molecular risk schemes, particularly in the heterogeneous Group 3 and Group 4 medulloblastoma subgroups, where it significantly outperformed the SJMB03, clinico-molecular and revised clinico-molecular approaches. Within the SHH subgroup, however, the model did not outperform existing risk stratification schemes, where established genetic risk factors already provide strong prognostic accuracy.

An extension called MANTIS-ITE went a step further, using data from the randomized ACNS0331 trial to estimate how each patient would fare under lower- versus standard-dose radiation. About 60% of average-risk patients were predicted to have no significant survival disadvantage with the lower dose (P = 0.683), whereas those flagged as treatment-sensitive fared worse with reduced radiation. Among children flagged as candidates for lower radiation who underwent cognitive testing (a subset of 61 patients), those who received standard-dose CSI showed declining IQ scores over follow-up, while those given the lower dose stayed relatively stable.

Applying explainable-AI methods, the team found that the model's risk signals in Group 3 medulloblastoma centered on MYC activation, including cases where the gene was switched on epigenetically rather than by high-level amplification—a nuance that standard tests can miss. The same framework was adapted to ependymoma with similar performance, but models trained on one tumor type performed at chance on the other, suggesting methylation encodes largely tumor-specific prognostic information.

Why it matters

Methylation profiling is already used clinically to classify brain tumors. This study suggests the same single assay could also carry prognostic information, potentially consolidating several tests into one and offering a more granular picture than broad risk categories. For families, that could eventually mean fewer children are overtreated—and fewer are undertreated.

Caveats

The work is retrospective: the models were trained and validated on archived datasets, not tested prospectively in the clinic. Whether acting on these predictions actually improves patient outcomes remains unproven. Treatment-effect modeling was limited to CSI dose, and the model does not account for late effects such as secondary cancers. The authors also note that the tool adds little in groups that already do very well, such as WNT tumors.

What to watch

The immediate next step is prospective validation, where predictions are made before treatment decisions and their impact on survival and cognition is measured. If that holds up, the blueprint could extend beyond childhood brain tumors to other cancers where outcome variability and overtreatment remain unresolved.

Reference: Stefan, S., Amann, A.S., Smith, K.S. et al. AI-based precision prognostics and therapy personalization for childhood brain tumors. Nature Genetics (2026). https://doi.org/10.1038/s41588-026-02815-4