Medulloblastoma subgroups from DNA methylation, with explanations
A learning project on public data. Nine pages, each a notebook executed end to end,
from the question through quality control, classification, explanation, the test against
gene expression, and the pathway and therapy angle. Code and tables:
github.com/robinson-vidva/mb-methylation-xai.
- 1. The question and the data: Why medulloblastoma, why methylation, what the four questions are, and the data.
- 2. Quality control: Probe filtering, beta distributions, failed probes in FFPE material, a sex check, and PCA.
- 3. Classification: Three models, two tasks, cross-validation, the held-out split, calibration, an external cohort, and sparse coverage.
- 4. Explainability: Integrated gradients: which CpGs carry each subgroup call and which genes they sit in.
- 5. Methylation versus expression: Does promoter methylation at those sites track expression in the same tumours? Concordant genes per subgroup.
- 6. Pathways and therapy: Pathway enrichment, then Open Targets: which concordant genes and which drivers have drugs.
- 7. Gene clusters and pathway flow: Gene clusters, a co-methylation network, pathway flow for Hedgehog, Wnt and the photoreceptor program, pathway activity per tumour.
- 8. Genes: names, cancer roles and literature: Every gene named: full name, oncogene or tumour suppressor status, medulloblastoma literature, Reactome pathways, and whether the methylation change fits the role.
- 9. Conclusions: Answers to the four questions, limits, reproducibility across operating systems, and next steps.
Data: GSE85212, GSE85217 (Cavalli et al. 2017), GSE54880 (Hovestadt et al. 2013), GPL13534.