Could tumour methylation help tailor radiotherapy for children with brain cancer?
A peer-reviewed model trained on molecular profiles produced individual survival curves and separated risk in independent medulloblastoma trials. Its retrospective treatment analysis is promising, but it cannot yet tell clinicians to reduce a child's radiotherapy dose.
The short version
- What happened
- MANTIS used genome-wide methylation, inferred copy-number changes and follow-up data from 2,786 patients with medulloblastoma or ependymoma to generate individual survival curves over ten years.
- Who it affects
- Children and families face a real trade-off between preventing recurrence and limiting long-term treatment effects; a model error in either direction could carry irreversible consequences.
- What remains unproven
- All model development and evaluation were retrospective analyses of previously collected molecular and outcome data.
Keep in mind
All model development and evaluation were retrospective analyses of previously collected molecular and outcome data.
The treatment-effect subgroup was derived after the ACNS0331 trial rather than prespecified and tested prospectively.
The neurocognitive analysis included only 61 patients in the low-predicted-effect subgroup and cannot establish a treatment policy's cognitive benefit.
Performance varied by risk group and tumour biology; MANTIS did not outperform existing schemes in the analysed SHH subgroup.
Methylation assays, clinical protocols and patient populations may differ from the historical cohorts used here.
The model estimates probabilities and counterfactual outcomes; it cannot observe both radiation outcomes for the same child.
Source analysis · 1 linked source. Primary analysed source date: 3 November 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 10 min The evidence, methods, implications and limitations
The direct answer: stronger prognosis evidence, not a licence to lower treatment
A peer-reviewed Nature Genetics study reports that a neural survival model built from childhood brain-tumour methylation profiles separated patients with different outcomes in independent clinical-trial cohorts. The model, called MANTIS, produced a survival probability curve for each patient rather than a single yes-or-no risk label. For medulloblastoma, its highest reported concordance index was 0.80 ± 0.07 in an average-risk validation group. A concordance index measures how often the model orders two patients' survival correctly; it is useful evidence of discrimination, but it is not the proportion of children whose outcome was predicted correctly.
The paper also explored whether the same molecular representation could identify children who might receive a lower craniospinal irradiation dose without losing survival benefit. That analysis used an earlier randomized trial, which gives the treatment comparison a stronger foundation than an ordinary observational dataset. However, the AI-defined subgroup was created retrospectively. No child was prospectively assigned a radiation dose using MANTIS, and the study did not demonstrate that an AI-guided treatment policy improves survival, cognition or quality of life. The clinically responsible conclusion is therefore bounded: the model improves molecular risk stratification in historical trial data and generates a testable de-escalation hypothesis, not a treatment recommendation for today's patients.[1]
What was trained, calibrated and held back for evaluation
The full study used genome-wide DNA methylation array data from 2,786 patients with medulloblastoma or ependymoma. For medulloblastoma, researchers trained MANTIS on 820 patients drawn from public and previously published cohorts and selected settings on a separate calibration set of 100. They then reserved 898 patients aged three to 21 from three clinical trials—SJMB03, ACNS0331 and ACNS0332—for independent evaluation. Those trial cohorts were not used to develop the model. Keeping them separate reduces the risk that reported performance merely reflects memorising the development data or tuning repeatedly against the final test set.
MANTIS combined more than 350,000 methylation measurements with inferred gene-level copy-number changes and, for the best medulloblastoma version, metastatic status. A biologically structured network compressed the measurements through 16,067 gene-linked nodes. Multi-task logistic regression then estimated the chance of an event across time intervals, allowing the model to use censored follow-up rather than treating every patient as if their final outcome were already known. The paper compared MANTIS with random survival forests, CoxBoost, DeepSurv and DeepHit. This is a prognostic design: it asks whether molecular patterns rank future outcomes, not whether the model itself changes them.[1]
The distinction that matters
Where the evidence stands
The study crosses several important validation steps, but the last step—testing an AI-guided treatment policy—has not happened.
Development
Train on 820 medulloblastoma patients and calibrate on 100
A locked molecular survival model
Independent validation
Evaluate in 898 patients from three trials
Useful but uneven prognostic discrimination
Retrospective treatment analysis
Reanalyse 170 randomized patients
A 104-patient de-escalation hypothesis
Still required
Prespecify and test AI-guided dosing prospectively
Clinical benefit and safety remain unproven
Independent trials showed useful discrimination—and important unevenness
In 206 average-risk SJMB03 patients treated with standard-dose irradiation, MANTIS reported a concordance index of 0.80 ± 0.07, a time-dependent area under the receiver operating characteristic curve of 0.85 ± 0.04 and an integrated Brier score of 0.08 ± 0.01. In 276 average-risk ACNS0331 patients receiving standard-dose treatment, the corresponding values were 0.77 ± 0.08, 0.81 ± 0.07 and 0.10 ± 0.01. Performance was lower in high-risk cohorts: the concordance index was 0.73 ± 0.09 among 99 high-risk SJMB03 patients and 0.68 ± 0.07 among 231 high-risk ACNS0332 patients. The model was informative across several settings, but not equally strong everywhere.
That variation matters more than a headline best score. Among 604 patients with Group 3 or Group 4 medulloblastoma, MANTIS achieved a concordance index of 0.73 ± 0.04, compared with 0.66 ± 0.04 for the SJMB03 scheme, 0.68 ± 0.04 for a clinico-molecular scheme and 0.70 ± 0.04 for a revised version. Yet in 143 patients with SHH tumours it did not outperform existing schemes: 0.77 ± 0.07 for MANTIS versus 0.74 ± 0.05 and 0.79 ± 0.05 for the comparators. A future clinical service would need to specify where the model adds information and where established genetics already performs as well.[1]
The radiotherapy analysis used randomization, but the subgroup was model-defined later
The treatment-effect experiment focused on 170 children aged three to seven with average-risk medulloblastoma in ACNS0331. The original trial randomized 86 to lower-dose craniospinal irradiation of 18 Gy and 84 to the standard 23.4 Gy. Across the trial as a whole, lower-dose treatment had previously produced inferior outcomes, so any proposal to reduce radiation must identify a subgroup with exceptional care. Researchers first pretrained a compact molecular representation using 1,524 additional methylation profiles without survival outcomes, then fine-tuned a dual-head survival model to estimate each patient's five-year survival under both doses.
MANTIS classified 104 patients—about 61%—as having a small predicted treatment effect. In this retrospective subgroup, lower-dose treatment was not associated with worse overall survival; the paper reports P = 0.683. The remaining 66 were classed as likely to benefit more from standard-dose treatment, and observed outcomes were worse for those who actually received the lower dose. Randomization supports comparison within the original trial, but it does not make a subgroup discovered by the model automatically confirmatory. The cut-off, model and selection procedure all need testing in a new, prespecified trial before clinicians can rely on the subgroup.[1]
Why the human stakes are unusually high
Craniospinal irradiation can help prevent a deadly tumour from returning, but exposure of the developing brain and spine can contribute to long-term cognitive and other health effects. Families and clinicians are therefore balancing two different forms of harm: undertreatment that permits recurrence and treatment intensity that affects a child's later development. A model that reliably identifies children for whom a lower dose preserves survival could be valuable. A model that is wrong in either direction could cause irreversible harm—either avoidable relapse or avoidable late effects.
The study examined intelligence scores for 61 patients in the low-predicted-effect subgroup: 27 received the lower dose and 34 the standard dose. Scores were relatively stable in the lower-dose group and declined over follow-up in the standard-dose group, but the subset was small and incomplete. The article's mixed-effects comparison reports P = 0.85 for the described trajectory analysis, so this evidence should not be turned into a claim that MANTIS has proven cognitive benefit. A prospective study would need prespecified neurocognitive endpoints, longer follow-up and careful accounting for missing assessments, schooling, age, supportive care and other influences.[1]
Ependymoma results broaden the method but do not remove the validation gap
The researchers also built a version for posterior fossa ependymoma. It used 709 patients for training, 140 for calibration and 159 for validation. The best ependymoma model used methylation values without the copy-number input that helped in medulloblastoma. Its calibration-set concordance index was 0.75 ± 0.14. Analyses of established risk factors, multiple regions from the same tumour and matched primary and recurrent samples suggested that the predictions tracked known biology and some within-tumour variation.
Those results show that the architecture is not tied to one tumour type, but they remain retrospective. Different hospitals can use different tissue handling, arrays, laboratory pipelines, treatment protocols and follow-up practices. A model that works across archived cohorts may still drift when used prospectively or when a patient belongs to a group that was rare in training. The paper makes processed data and code available, which supports scrutiny and replication. Clinical adoption would still require locked software, quality control for the assay, external testing in contemporary patients and a clear response when data are missing or a prediction conflicts with established risk factors.[1]
Funding, interests and what would change the assessment
The work was supported by the St. Jude Research Collaboratives programme, the Charles H. Hood Foundation, the Children's Cancer Research Fund, the V Foundation and US National Cancer Institute grants to the Children's Oncology Group. The article says funders were not directly involved in study design, data collection, analysis, interpretation or writing. The authors declared no competing interests. The evidence comes from previously collected, de-identified datasets and clinical trials; this was not new prospective recruitment and did not expose children to an AI-directed treatment decision.
The assessment would strengthen if an external team reproduced performance from the released code, laboratories showed that predictions are stable across assay platforms, and a prospective multicentre study locked the model and subgroup threshold before enrolment. The decisive test would randomize model-eligible children to an AI-informed de-escalation strategy or current care, with survival, relapse, neurocognition, quality of life and adverse effects followed for years. Confidence would fall if calibration shifted in new populations, if apparent treatment effects disappeared under prespecified analysis, or if clinicians could not resolve disagreement between MANTIS and established molecular risk schemes. Until then, this is high-value validation research, not routine decision support.[1]
Watch the background, then check the evidence
Illustrative photo · not a video frame
The Royal Society · 44 min 28 sec · Background explainer
Is this the AI future we want? | The Royal Society
YouTube loads only when you choose; playback starts when you press play.
A Royal Society discussion with Madhumita Murgia and biomedical engineer Alison Noble on AI’s potential in healthcare, science and culture, recorded at the 30 June 2026 event.
Evidence check: A discussion of opportunities and risks, including future possibilities. It is not evidence that every application described is available or clinically proven.
Published 2026-07-01. Source checked 14 October 2026. Check captions on the publisher’s video page.
Check the healthcare evidencePhoto: Daniel Sone / National Cancer Institute / Unsplash · Licence
Continue the story
Read next
Continue exploring clinical validation: compare the finding and its evidence limits.
A face photo ranked cancer mortality risk—should it guide treatment?
A peer-reviewed model separated higher- and lower-risk patients in two retrospective radiotherapy cohorts, but it has not shown that face-based scores improve clinical decisions or outcomes.
Read the next reportExplore the full source and evidence record
What this means for people
- Children and families face a real trade-off between preventing recurrence and limiting long-term treatment effects; a model error in either direction could carry irreversible consequences.
- Clinicians would need calibrated probabilities, assay quality controls and a defined way to handle disagreement with established molecular and clinical risk factors.
- The study did not assign treatment using AI, so no family should interpret its model-defined subgroup as advice to reduce radiotherapy now.
Global context
The cohorts combined North American and European research networks and internationally used trial datasets, but access to methylation profiling, paediatric radiotherapy and long-term follow-up is uneven. A molecular tool could streamline information where the assay is already available, while widening disparities if validation and laboratory capacity remain concentrated in specialist centres.
What the evidence does not yet show
- All model development and evaluation were retrospective analyses of previously collected molecular and outcome data.
- The treatment-effect subgroup was derived after the ACNS0331 trial rather than prespecified and tested prospectively.
- The neurocognitive analysis included only 61 patients in the low-predicted-effect subgroup and cannot establish a treatment policy's cognitive benefit.
- Performance varied by risk group and tumour biology; MANTIS did not outperform existing schemes in the analysed SHH subgroup.
- Methylation assays, clinical protocols and patient populations may differ from the historical cohorts used here.
- The model estimates probabilities and counterfactual outcomes; it cannot observe both radiation outcomes for the same child.
What to watch next
- Independent replication using the released code and contemporary external cohorts
- Prospective multicentre calibration across laboratories and methylation platforms
- A prespecified trial of AI-guided radiotherapy de-escalation with survival and neurocognition endpoints
- Performance by age, tumour subgroup, geography and underrepresented populations
- Regulatory status, version control and procedures for clinician–model disagreement
Living evidence record
Impact record IAI-0VWESET
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
3 November 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Nature Genetics published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 3 November 2026
Research question, topics and regions
Whether a neural survival model built from tumour methylation profiles can improve childhood brain-tumour prognosis and identify patients who might safely receive less craniospinal irradiation
Topics: Childhood cancer · Medulloblastoma · Ependymoma · Radiotherapy · Precision medicine · Clinical validation
Regions: United States · Germany · North America · Europe
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.