Brain Tumor Biomarker Survives Tough Test: MGMT Signal Holds After Batch Effect Scrutiny

In the rarefied world of pediatric brain tumor research, a single biomarker can shape how clinicians think about an incurable disease. That is exactly why a recent statistical dispute over MGMT expression in H3 G34-mutant diffuse hemispheric glioma mattered so much. Now, a team at St. Jude Children’s Research Hospital has delivered a detailed rebuttal showing that the link between MGMT expression and patient survival is not an artifact of laboratory technique. Writing in Acta Neuropathologica, the researchers demonstrate that their original finding survives rigorous adjustment for the pre-analytic and technical variables that critics argued could have manufactured the association out of nothing.

The controversy began when independent analysts reexamined the team’s published RNA sequencing data and raised the possibility that technical batch effects, systematic differences introduced by how and when samples are processed, might confound the reported relationship between MGMT and outcome. The concern is far from academic. Cohorts of rare pediatric brain tumors are almost always retrospective, assembled across multiple institutions, collected over many years, and stored using different preservation methods. Each of those factors can leave fingerprints on gene expression measurements that have nothing to do with the underlying biology of the tumor.

To settle the question, the St. Jude group analyzed RNA sequencing data from 110 H3 G34-mutant diffuse hemispheric gliomas, all of which had been sequenced centrally on the Illumina NovaSeq 6000 platform. The dataset included every RNA-sequenced tumor from the original report. The researchers deliberately excluded four samples that had been run on an older sequencing platform, the very samples that had appeared as extreme outliers in the critics’ principal component analysis. Raw sequencing counts were then normalized using the trimmed mean of M values method from the edgeR package, converted to log2 counts per million, and corrected for processing batch effects with the limma removeBatchEffect procedure. Notably, the team acknowledged that this batch correction had been part of their original analysis all along, although they regretted not describing it explicitly in their original Methods section.

The technical analysis revealed just how powerful these artifacts can be when left uncorrected. In raw counts, the first two principal components tracked library size, the total number of sequencing reads per sample, and samples clustered sharply by the year their data were released, with Kruskal-Wallis tests yielding p values below 0.001. After normalization and batch correction, that separation vanished entirely, with p values of 0.51 and 0.81. The lesson is a familiar one in genomics: because all gene counts scale with sequencing depth, raw counts of completely unrelated genes tend to correlate with one another for purely mathematical reasons.

That principle directly undercut the critics’ central piece of evidence. The reanalysis had reported that MGMT expression correlated strongly with two housekeeping genes, CD34 and ACTB, with correlation coefficients of 0.76 and 0.75 in raw feature counts from 20 tumors, suggesting a shared technical artifact. The St. Jude team found both correlations in their own cohort but weaker in raw counts, at 0.29 and 0.42, and after proper correction the correlations disappeared altogether, dropping to negative 0.04 and 0.01. Raw ACTB counts were also 4.6-fold higher in fresh-frozen samples than in formalin-fixed paraffin-embedded samples, a depth difference that normalization cleanly removed. In other words, the smoking gun the critics had identified was itself an artifact of looking at unnormalized data.

The team then went a step further, testing whether any of four pre-analytic and technical variables could actually explain the survival association. They examined collection year, preservation method, data release year, and tumor content, the latter estimated from the variant allele frequency of the H3 G34 mutation on whole exome sequencing combined with review of stained tissue sections. A true confounder must be associated with both the biomarker and the outcome. None of the four variables was significantly associated with MGMT expression, progression-free survival, or overall survival in univariable or multivariable models. MGMT expression did not differ by preservation method, with a p value of 0.71, although only five tumors in the cohort were fresh-frozen.

The most decisive result came from the survival models themselves. Among the 25 patients from the original report, continuous MGMT expression, adjusted for all four technical variables, remained strongly associated with both progression-free survival and overall survival. Each doubling of MGMT expression carried an adjusted hazard ratio of 1.64 for progression, with a 95 percent confidence interval of 1.20 to 2.24 and a p value of 0.002, and an adjusted hazard ratio of 3.00 for death, with a confidence interval of 1.73 to 5.21 and a p value below 0.001. Critically, the association held regardless of the median cutoff used in the original study, meaning the finding does not depend on any arbitrary threshold chosen to split patients into high and low expression groups.

The rebuttal also addressed a second concern about methylation data. The critics had suggested that a shared artifact could inflate the correlation between MGMT expression and methylation at gene-body CpG sites. The St. Jude team pointed out that methylation was measured from DNA on a separate array platform, so sequencing-depth artifacts alone cannot explain the correlation. In the 34 tumors with both data types, Spearman correlations between MGMT expression and methylation at three gene-body CpG sites ranged from 0.71 to 0.75, and remained between 0.58 and 0.78 after statistically partialing out all four technical variables, with all p values below 0.001. The pattern was also biologically coherent: two promoter CpGs measured by the MGMT-STP27 assay correlated negatively with expression, a relationship reported across 31 cancer types and supported by demethylation experiments in glioblastoma cells.

Beyond defending their own result, the authors offered a practical playbook for anyone wrestling with retrospective genomic cohorts. They recommend recording pre-analytic variables from the start, normalizing data and correcting for batch before comparing samples, excluding low-complexity libraries that correction cannot rescue, and testing whether adjustment for technical variables changes the association between a marker and an outcome, with key findings confirmed by an independent assay. Five outlier samples with degraded RNA or low input remained outliers even after correction, underscoring that some damaged libraries are simply beyond statistical repair.

The team was careful not to overclaim victory. They agreed with their critics that correlation alone cannot prove that gene-body methylation regulates MGMT, and they emphasized that MGMT expression still requires validation in larger, prospective cohorts before it can guide therapy decisions for children with these aggressive tumors. But the core message stands: a candidate biomarker for a devastating pediatric cancer has now withstood a hostile audit of its technical foundations, and the field has a clearer template for telling real biology from batch-effect noise.

Subject of Research: MGMT expression as a prognostic biomarker in H3 G34-mutant diffuse hemispheric glioma and the control of technical batch effects in retrospective RNA-seq cohorts

Article Title: The association between MGMT expression and outcome in H3 G34-mutant diffuse hemispheric glioma is robust to adjustment for pre-analytic and technical variables

Article References: The association between MGMT expression and outcome in H3 G34-mutant diffuse hemispheric glioma is robust to adjustment for pre-analytic and technical variables. (n.d.). https://doi.org/10.1007/s00401-026-03105-3

Image Credits: AI Generated

DOI: 10.1007/s00401-026-03105-3

Keywords: MGMT, H3 G34-mutant glioma, diffuse hemispheric glioma, batch effects, RNA sequencing, biomarker, pediatric brain tumor, methylation, prognosis, bioinformatics, Acta Neuropathologica, St. Jude