Glioma Subtype Classification & Survival Modeling from DNA Methylation
Classifying brain-tumor subtypes and predicting survival from DNA methylation.
Classification arm
Survival arm (exploratory)
Select a step to see what it does and the numbers behind it.
01The problem
Methylation-based glioma classifiers often act as black boxes, and many prognostic models come from a single cohort and overfit, so clinicians can't easily trust or reuse them.
02Approach & architecture
Built a leakage-controlled pipeline on Illumina 450k and EPIC data: an XGBoost classifier reduced from 403,954 probes to 70 CpGs, plus penalized Cox models on interpretable gene-region features under nested cross-validation.
03Results explorer
Accuracy of the frozen three-class XGBoost classifier, with 95% confidence intervals where reported. External cohorts were never seen in training; EPIC sits lower than 450k, consistent with platform-specific probe differences.
Every filter is fitted on the 462 training samples only. On a log scale, each bar is one step: 403,954 probes become 70 with no detectable cost in held-out accuracy.
Held-out accuracy as the feature set shrinks
Nested cross-validation: 5 outer × 5 inner folds, repeated 5 times. Points are means, whiskers ± 1 SD across outer folds. The pooled row ranks patients across histologies, so it is shown but not compared with the rows below it.
Hazard ratio between high- and low-risk halves (median split of each model's risk score), log scale with 95% CI. These are in-sample: the split is made on the data the model was fitted to, so they overstate separation in new patients.
The pooled survival cohort after QC: 711 patients, 256 deaths. Glioblastoma has the highest event rate and roughly a third of the median survival of the IDH-mutant tumours.
Median overall survival
Leading gene-region features of the final models. Elastic-Net Cox and CoxBoost converged on largely the same genes within each histology. These are candidates for validation, not established markers.
Pooled model
4 non-zero coefficients; DIAPH1 (gene body) dominates
- DIAPH1
- LOC100133893
- ATXN7L1
- LRRC14B
Astrocytoma
Promoter features; POMT1 and HSPG2 lead in both models
- POMT1
- HSPG2
- C8orf4
- PANK3
- GGT1
- EDNRA
- ZNF160
Oligodendroglioma
Top three shared by both models
- TRMT61B
- LHFPL5
- PRKAR2A
- C20orf95
- HOXC9
- HOXA7
- RDM1
- KIF17
- MIR1260
- PDGFRL
Glioblastoma
Dominant promoter signals
- HMGA2
- RHOBTB3
- CNOT8
- POFUT1
- SPP1
- TRIM4
- RNF115
Figure 2. Results explorer. All values from the preprint (Abulizi & Mamatjan, 2026): Section 3.1 and Tables 1–4.
04Results & impact
Measured What the work shows
97.4% accuracy on a held-out test set and 93–100% on external 450k and EPIC cohorts, from a 70-probe signature selected out of 404K CpG probes. First-author preprint with Dr. Yasin Mamatjan, posted on SSRN and under peer review.
Expected Possible impact, not yet measured
Exploratory survival models surfaced candidate prognostic loci (HMGA2, RHOBTB3, HSPG2, TRMT61B). These are hypotheses: they need validation in independent cohorts before they could inform prognosis. A compact probe panel could, if validated prospectively, make methylation subtyping cheaper to run.
05Lessons
Strict leakage control is the difference between optimistic-looking results and honest ones. SIS feature screening, scaling, and hyperparameter tuning must all be re-fit inside every training fold — otherwise the held-out folds aren’t really held out. The largest practical win came from gene-region aggregation: collapsing hundreds of thousands of CpGs to per-gene promoter and body summaries improved interpretability without sacrificing performance, and produced feature lists that biologists could actually engage with. Pooled-stratified modelling captured a single dominant cross-cancer risk axis with very few features, while per-cancer models surfaced richer histology-specific biology — different questions, different right tools.