AI study identifies potential candidates for lower-dose childhood brain tumor radiation

An AI model described in Nature Genetics on November 3 could help identify children whose brain tumors warrant less intensive treatment. Called MANTIS, it predicts survival from molecular tumor data. Researchers also used it to identify potential candidates for lower-dose radiation, aiming to reduce lasting harm without sacrificing survival.

A survival curve, not just a risk category

The Dana-Farber Cancer Institute–led team developed MANTIS for medulloblastoma and ependymoma, two childhood brain tumors. Its starting point is DNA methylation profiling, a molecular test already used to help classify brain tumors. The study uses measurements from that test to ask a different question: how might an individual patient’s survival probability change over time?

The research drew on genome-wide molecular data from 2,786 patients. MANTIS combines methylation patterns with changes in gene copy numbers and learns from those measurements alongside patients’ recorded survival outcomes. Rather than returning only a broad risk label, it generates a personalized survival probability curve covering 10 years after diagnosis.

The neural network includes an internal “gene layer,” with each node corresponding to a gene. That design groups molecular information by gene and reduces the number of model parameters researchers must fit. The best-performing medulloblastoma version also uses whether the cancer had spread at diagnosis, an established clinical risk factor.

Checking forecasts against independent trials

For medulloblastoma, researchers trained the model using 820 patients and used another 100 to tune it. They then evaluated it in three independent clinical trial cohorts. Those trial patients were not used for model development, separating the test of its forecasts from the data used to build them.

The authors report stronger predictive performance than current clinical indicators. Their main selection measure was the concordance index, which checks how well predicted survival rankings match observed rankings. It is not a percentage of patients whose outcomes the model correctly predicts. The researchers also checked measures of discrimination and overall prediction accuracy.

898Patients in independent trial cohorts

The three validation cohorts included patients aged 3–21. The authors specify that the resulting performance estimates apply to patients older than 3.

Performance varied across patient groups. In the SJMB03 trial, the concordance index was 0.80 ± 0.07 among average-risk patients receiving standard-dose radiation, versus 0.73 ± 0.09 in its high-risk group. That variation matters: the strongest headline result is not a uniform score across every child or treatment setting.

Finding whom less radiation might help

The treatment question goes beyond forecasting. Radiation can help control cancer while also causing long-term cognitive harm, with effects influenced by dose and age at treatment. Some earlier attempts to reduce doses were unsuccessful. The authors’ premise is that an approach that fails across a whole group might still suit selected patients.

To investigate that possibility, the team extended MANTIS to estimate individual treatment effects, using randomized trial data and methods designed to estimate causal effects. In the ACNS0331 analysis, it identified about 60% of average-risk medulloblastoma patients as potential candidates for less-intensive craniospinal irradiation—radiation directed at the brain and spine.

In that selected group, researchers detected no statistically significant survival disadvantage with lower-dose radiation; lower doses were associated with better cognitive outcomes. That finding does not prove the doses are equivalent or that changing treatment based on MANTIS is safe. The authors say prospective trials are needed to establish clinical usefulness and demonstrate patient benefit.