Diagnosis Structure Explains Much of the Apparent Mutation-Burden Survival Signal in Public Pediatric Brain Tumor Atlas Data
Public pediatric brain tumor datasets now make pan-histology molecular prognostic analyses possible, but that breadth creates a hard statistical problem: tumor class, molecular burden, treatment, and outcome are not exchangeable across diagnoses. We reanalyzed public summary files from OpenPBTA release v23, merging harmonized clinical annotations with coding tumor mutation burden for 782 survival-annotated molecular profiles from 943 enrolled participants. Hypermutated samples, defined as coding tumor mutation burden at or above 10 mutations/Mb, had worse unadjusted overall survival than lower-burden samples (log-rank p=0.0036), but the hypermutated group was small (n=7) and diagnosis-mixed. Mutation burden also varied strongly across broad histology classes. After residualizing log-transformed mutation burden against age at diagnosis and broad histology, the association with observed death weakened (Mann-Whitney p=0.089). Within the three broad histology strata large enough for sensitivity checks, median-split TMB did not show a consistent high-burden survival penalty. These results do not support mutation burden as a diagnosis-independent prognostic marker in this public summary-data setting. They instead quantify a useful failure mode: pan-pediatric CNS tumor models can rediscover diagnosis composition unless diagnosis structure is modeled explicitly. OpenPBTA remains an unusually valuable resource, but survival modeling from public summaries should be framed as hypothesis generation unless treatment, molecular subtype, sampling phase, and diagnosis-specific effects are incorporated.
Reviews
This short reanalysis tackles an important and common failure mode in pan-diagnosis prognostic modeling: apparent biomarker–survival associations that are largely induced by diagnosis composition rather than biology that generalizes across tumor classes. Using OpenPBTA v23 public summary tables, the authors show a statistically significant unadjusted survival difference when dichotomizing coding TMB at a conventional “hypermutation” cutoff (≥10 mut/Mb), and they correctly highlight that the hypermutated group is extremely small and diagnosis-mixed. The main strength is the paper’s focus on confounding structure and the attempt to quantify how much of the signal remains after explicitly accounting for age and broad histology; the within-stratum sensitivity checks further support the claim that a single pan-pediatric-CNS “high TMB is worse” story is not robust in these summaries. The main weakness is methodological: the core adjustment strategy (“residualize log10(TMB+0.05) on age and broad histology, then compare residuals by death event via Mann–Whitney”) is not a survival model and does not account for censoring or time-to-event information; it effectively tests association with event occurrence rather than hazard/OS. Given the tiny hypermutation n (7) and likely diagnosis/treatment heterogeneity, power is also very limited, and the dichotomization choices (≥10; median splits within strata) can be unstable. Nonetheless, the central conclusion is appropriately cautious and largely justified: in this public, summary-data setting, the presented analyses do not support coding TMB as a diagnosis-independent prognostic marker, and they convincingly motivate diagnosis-aware modeling as a minimum requirement. Integrity summary: Claim support is generally aligned with the described analyses, and the authors appropriately flag the dominant confounding risk and the summary-data limitations. However, the evidentiary link between “diagnosis explains much of the signal” and “T