AI Model Predicts Childhood Brain Tumour Survival From Tumour DNA

Doctors treating children with medulloblastoma or ependymoma, two types of childhood brain tumour, sort patients into broad risk groups and set treatment intensity accordingly. A study published on 3 November in Nature Genetics describes a childhood brain tumour AI that instead gives each child their own survival curve over ten years, built from the chemical tags on their tumour’s DNA.

The tool, called MANTIS, was trained on DNA methylation array data from 2,786 patients. Its authors, led by Volker Hovestadt of Dana-Farber Cancer Institute, Boston Children’s Hospital and the Broad Institute, with St. Jude Children’s Research Hospital and German partners including LMU Munich and the German Cancer Research Center (DKFZ), say it predicts outcomes better than the clinical indicators used today. The study is retrospective: it looked back at children already treated, and the authors say prospective trials are still needed before it could guide care.

How the childhood brain tumour AI reads tumour DNA

MANTIS stands for “methylation profiling and artificial neural networks for time-dependent individualized survival predictions”. DNA methylation, the pattern of chemical marks that switch genes on and off, is already used to classify brain tumours into molecular subtypes. MANTIS uses the same genome-wide data, plus clinical information, to predict the probability of survival at each point over ten years, rather than assigning a child to a fixed risk band.

The team built two models. The medulloblastoma model was trained on 820 patients with a separate calibration set of 100. The ependymoma model was trained on 709 patients, with 140 for calibration. The authors have released the code and model on GitHub and Zenodo under the open MIT licence.

The accuracy figures and where they fall short

The paper measures accuracy with the concordance index, or c-index, which scores how well a model ranks patients by outcome: 0.5 is chance and 1.0 is perfect. In the SJMB03 trial cohort, MANTIS scored 0.80 ± 0.07 for 206 average-risk medulloblastoma patients and 0.79 across all 305 patients. In data from the ACNS0331 trial it scored 0.77 for 276 average-risk patients, and 0.76 across 507 patients from ACNS0331 and ACNS0332 combined. The ependymoma model scored 0.75 ± 0.09 on a validation set of 100.

Among SJMB03 patients that current schemes class as average risk, the 27 that MANTIS called high risk were 8.4 times more likely to have a recorded death than the 97 it called low risk. For the group 3 and group 4 medulloblastoma subtypes, MANTIS scored 0.73, against 0.66 to 0.70 for three existing risk schemes. It is not better everywhere. For the SHH subtype, the authors report that it did not outperform existing schemes.

Flagging children who might manage on less radiotherapy

The most clinically striking part of the study concerns radiotherapy to the brain and spine, which saves lives but can damage a child’s developing brain. The ACNS0331 trial randomly assigned average-risk medulloblastoma patients to a standard dose of 23.4 Gy or a lower 18 Gy. The researchers fine-tuned MANTIS on 170 of those patients to estimate how much each child’s survival depended on the higher dose.

For 104 children, about 60%, the model predicted a small treatment effect. In that group, those who had received the lower dose showed no significant survival disadvantage (P = 0.683). For the other 66, those given the lower dose fared significantly worse (P = 0.00245). Nearly all SHH tumours (92.9%) fell into the low-effect group, against 53.7% of group 3 and group 4 tumours.

In a subset of 61 children with IQ data, scores tended to fall over time on the standard dose and stay relatively stable on the lower dose. The paper describes this only as “suggesting” an association and reports P = 0.85, so it is not a statistically significant finding. An earlier conference abstract from June had reported a significant IQ benefit; the final peer-reviewed paper does not.

No significant difference is not the same as no harm

The radiotherapy result is where the study will draw most attention, and where it most needs careful reading. “No significant survival disadvantage” in 104 children, split roughly evenly between doses, is not proof that the lower dose is equally safe for them. A small group can hide a real difference, and the authors’ own wording is that patients were identified “without a significant survival disadvantage”, not that their survival was shown to be equal. The model was also fine-tuned and tested on the same trial’s patients, which is why the authors call for prospective trials “to establish clinical utility, ensure predictions are clinically interpretable and demonstrate patient benefit”.

The limits are spelt out. Treatment-effect modelling covered only the radiotherapy dose, not chemotherapy or surgery, and late effects such as hearing loss or hormone problems are not modelled. The IQ signal, which would be the main reason to cut a dose, did not reach significance. Even so, the approach fits a pattern TSN has followed in other cancers: an FDA-cleared MRI model that scores 10-year breast cancer return risk and a tool that picks out aggressive prostate cancer from biopsies both aim to move patients out of crude risk bands. In children, the stakes run in both directions, because overtreatment leaves lifelong damage while undertreatment risks relapse. Releasing the code openly means other groups, including those working on digital twins for hard-to-treat breast cancer and AI-ranked drug combinations, can test the method on their own cohorts rather than take the reported accuracy on trust.

A prospective trial is the test that counts

MANTIS relies on DNA methylation profiling, the same kind of data already used to sort brain tumours into molecular subtypes. The next step the authors set out is a prospective trial in which predictions are made before treatment and children are followed forward. Until one reports, MANTIS is a well-validated research model, and doses for children with medulloblastoma will continue to be set by existing risk groups.

Sources

• Nature Genetics: Stefan, S. et al., “AI-based precision prognostics and therapy personalization for childhood brain tumors”, 3 November 2026 (peer-reviewed study)
• Zenodo: MANTIS v1.0.0 code and model release (authors’ repository)
• GitHub: sstefan01/methylation-survival (authors’ code)
• Neuro-Oncology Pediatrics: conference abstract, 23 June 2026 (earlier abstract)

Related stories

• FDA Clears MRI-Based AI That Scores 10-Year Breast Cancer Return Risk
• AI Tool Picks Out Aggressive Prostate Cancer in a 63-Biopsy Test
• UCL Digital Twins Suggest Extra Drugs for Hard-to-Treat Breast Cancer
• UCLA’s MAESTRO AI Ranked 3,592 Compounds to Pair With a Cancer Drug