9. Conclusions

Page 9 of 9. previous: 8. Genes: names, cancer roles and literature | index

Answers to the four questions

1. Classification. Public beta values and a fixed pipeline reproduce the four subgroups at balanced accuracy 0.97 on the held-out split and 0.98 on an external cohort, with FFPE as good as frozen. The twelve subtypes reach 0.86 at best. The subtype model is overconfident. Trained with random input masking, the same network still classifies subgroups from a few hundred CpGs.
2. Explanations. The CpGs that drive the four-group calls are mostly hypomethylated in their subgroup and sit in genes of neuronal and developmental identity, not in the known drivers.
3. Methylation and expression. Across all genes, promoter methylation barely tracks expression. At the model's top promoter sites it does, with a clearly negative correlation for most genes. The classifier's explanations point at functional sites. Split by layer, expression-only changes outnumber concordant ones five to seven times and carry the drivers and 15 of the 22 NanoString signature genes; methylation-only changes are larger still and are dominated by silent olfactory receptor clusters in partially methylated domains.
4. Pathways and therapy. Concordant genes recover the Group 3 photoreceptor program, Wnt signalling in WNT and cell fate commitment in SHH. Few concordant genes have drugs, and those that do are mostly channels and stromal genes. The approved subgroup-specific drugs act on mutated drivers (SMO, CDK6) whose promoters do not change. The pathway flow analysis makes the distinction explicit: Hedgehog and Wnt targets are over-expressed in SHH and WNT tumours with flat promoter methylation, the signature of activation by mutation, while the Group 3 photoreceptor program is hypomethylated and expressed along the whole chain, the signature of activation by demethylation. Methylation marks lineage programs and microenvironment; mutation marks the current drug targets.

Literature check on the three candidate observations

Three observations looked as if they might be new. Each was checked against PubMed on 2026-10-18. The verdicts are limited to what the searches returned.

1. Two modes of pathway activation. Hedgehog and Wnt targets are over-expressed in SHH and WNT tumours with flat promoter methylation; the Group 3 photoreceptor program is hypomethylated and expressed along the whole chain. The Group 3 half is anticipated. Hovestadt et al. 2014 showed from whole-genome bisulfite sequencing that hypomethylation tracks expression in medulloblastoma and that focal low-methylation regions mark subgroup-specific transcription factor networks (https://doi.org/10.1038/nature13268). Chen et al. 2025 showed that enhancer hypomethylation, in particular of the OTX2 super-enhancer through TET3, sustains Group 3 tumours (https://doi.org/10.1016/j.xcrm.2025.102474). Gudenas et al. 2026 described a tumour-associated photoreceptor signature shared by Group 3 medulloblastoma, pineoblastoma and retinoblastoma, whose constituents are selective dependencies (https://doi.org/10.1016/j.ccell.2026.02.010). What the searches did not return is the side-by-side contrast: the same tumours, promoter level, three pathways, mutation-driven targets with flat promoters against a demethylation-driven program. That contrast is consistent with the literature and, as far as these searches go, not described in it. An independent check supports the approach: Lindsey et al. 2014 reported regional hypomethylation of VAV1 with elevated expression as a conserved feature of SHH tumours (https://doi.org/10.1038/onc.2014.405). In this cohort VAV1 promoter beta is 0.50 lower in SHH tumours with expression modestly up, below the two-fold threshold used here, so the event is present in the data and was missed only by the threshold.
2. Explanations landing on functional sites. No paper was found that validates the attributions of a methylation classifier against expression in the same tumours. The search was narrow and this is a methods observation; it is reported here without a novelty claim.
3. The WNT feedback genes. The expression half is known. Vriend et al. 2019 and 2020 found that ZNRF3 over-expression alone predicts the WNT subgroup in the Pfister and Cavalli cohorts (https://doi.org/10.1007/s12311-019-1009-y, https://doi.org/10.1016/j.bbadis.2020.165839), and Desterke et al. 2025 listed Wnt pathway regulators among the LEF1 targets active in WNT medulloblastoma (https://doi.org/10.3390/curroncol32010035). The promoter methylation state of these genes, RNF43, ZNRF3, APCDD1 and SP5 hypomethylated in WNT tumours while AXIN2, NKD1, NOTUM and DKK1 are expressed with no promoter change, was not found in the searched literature. It is a small observation with a textbook mechanism behind it, and it would need an independent cohort before it is more than that.

One directly related study appeared in the search: Tuerxun et al. 2026 integrated the same two GEO series, GSE85212 and GSE85217, for a prognostic model with limma and ChAMP (https://doi.org/10.1177/20503121251386139). Its focus is survival, which this project does not analyse, and it is cited as related work.

What this project does not show

• Labels came from clustering of the same data types, so the classification is a reproduction.
• Beta values were taken as deposited. The clinical reference classifier starts from raw IDAT files with its own normalisation.
• Variance-based feature selection and a 10,000-CpG cap are simple choices.
• Correlation between methylation and expression is not causation.
• The sparse-coverage simulation removes CpGs uniformly. Real nanopore coverage is not uniform.
• 153 test tumours, 4 of them WNT beta. Subtype numbers are noisy.
• Open Targets tractability says a drug exists against a protein, not that the drug helps in medulloblastoma or crosses the blood-brain barrier.

Reproducibility

Seeds are fixed. The pipeline was run twice from the same code and GEO files, on Linux and on Windows. Every result from scikit-learn, scipy and statsmodels was identical. Every result that passes through PyTorch (the MLP, the autoencoder, the integrated gradients) differed, because the two CPU builds sum floating point numbers in a different order and 80 training epochs amplify that. The largest shift was the twelve-subtype MLP, 0.859 against 0.816 balanced accuracy. The README lists every difference. A stronger version would train several seeds and report the spread.

import json
import pandas as pd
from IPython.display import Image, display
from pathlib import Path
TAB = Path('../results/tables')
FIG = Path('../results/figures')
pd.set_option('display.width', 160)
pd.set_option('display.max_columns', 30)
pd.set_option('display.max_colwidth', 90)

pd.read_csv(TAB / 'test_metrics.csv')

	task	model	balanced_accuracy	macro_f1	accuracy
0	group	logreg	0.970	0.968	0.974
1	group	rf	0.974	0.974	0.980
2	group	mlp	0.970	0.968	0.974
3	subtype	logreg	0.810	0.817	0.869
4	subtype	rf	0.859	0.867	0.908
5	subtype	mlp	0.859	0.840	0.863

What would come next

• Start from IDAT files with a standard normalisation, and compare with the deposited beta values.
• Train several seeds and report means with spread for every PyTorch result.
• Temperature scaling for the subtype model, then check calibration on the external cohort.
• Survival: the Cavalli subtypes differ in outcome, and the same cohort allows CpG-level markers to be tested against survival.
• Nanopore: replace the uniform simulation with real shallow-sequencing coverage patterns.
• Therapy: for the Group 3 photoreceptor program and the SHH stromal genes, read the preclinical literature and the DepMap dependency data, and test whether the concordant genes separate responders in any published drug screen on medulloblastoma lines.

Page 9 of 9. previous: 8. Genes: names, cancer roles and literature | index