1. The question and the data

Page 1 of 9. index | next: 2. Quality control

Background

Medulloblastoma is the most common malignant brain tumour of childhood. It arises in the cerebellum. Since 2016 the WHO classification splits it into four molecular subgroups: WNT, SHH, Group 3 and Group 4. The subgroups differ in cell of origin, driver pathway, outcome and treatment. WNT tumours have an excellent outcome and are candidates for therapy de-escalation. SHH tumours are driven by the Hedgehog pathway, for which SMO inhibitors exist. Group 3 tumours often carry MYC amplification and have the worst outcome. Group 4 is the largest subgroup and the least understood. Cavalli et al. 2017 split the four subgroups further into twelve subtypes with distinct outcomes.

DNA methylation. A methyl group can be attached to a cytosine that is followed by a guanine, a CpG site. Methylation at a gene promoter usually switches the gene off. Methylation patterns are set during development, are stable, and survive formalin fixation. They encode the cell of origin of a tumour. This is why methylation arrays became the clinical standard for classifying brain tumours (Capper et al. 2018). The Illumina 450k array measures about 485,000 CpG sites. Each site gets a beta value, the fraction of DNA copies that are methylated, from 0 (unmethylated) to 1 (methylated).

Gene expression is the amount of messenger RNA a gene produces. Here it was measured on an Affymetrix array in the same 763 tumours. Having both measurements in the same tumours makes it possible to test whether a methylation change actually changes a gene's output.

Questions

1. Can a classifier trained on public beta values reproduce the published subgroup and subtype calls, and does it still work on tumours from another laboratory and on archival (FFPE) material?
2. Which CpG sites carry the decision, and which genes do they sit in?
3. Does promoter methylation at those sites track gene expression in the same tumours?
4. Which pathways differ between subgroups at the level of methylation and expression, and do any of the affected genes point to existing drugs?

Questions 1 and 2 are about diagnosis. Questions 3 and 4 are about biology and therapy. The link between them is that a classifier is only useful beyond diagnosis if the sites it uses mean something.

Data

All data are public and are fetched by scripts/01_download.py, with md5 checksums.

Source	Content	Role here
GSE85212, Cavalli et al. 2017	763 primary medulloblastomas, 450k beta values	training and held-out test
GSE85217, Cavalli et al. 2017	the same 763 tumours, Affymetrix expression	methylation versus expression
GSE54880, Hovestadt et al. 2013	276 medulloblastomas, 450k, 169 FFPE and 107 frozen	external test cohort
GPL13534	Illumina 450k manifest	CpG to gene annotation

Labels are the subgroup and subtype calls deposited with the data by the original authors. They were themselves derived from methylation and expression clustering, so the classification part of this project reproduces a known structure rather than discovering one.

The cohort below is split into 610 training and 153 test tumours, stratified by subtype, with a fixed seed. The test tumours are not touched until the final evaluation.

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)

lab = pd.read_csv('../data/processed/labels.csv', index_col=0)
pd.crosstab(lab['subtype'], lab['split'], margins=True)

split	test	train	All
subtype
Group3_alpha	14	53	67
Group3_beta	7	30	37
Group3_gamma	8	32	40
Group4_alpha	20	78	98
Group4_beta	22	87	109
Group4_gamma	24	95	119
SHH_alpha	13	52	65
SHH_beta	7	28	35
SHH_delta	15	61	76
SHH_gamma	9	38	47
WNT_alpha	10	39	49
WNT_beta	4	17	21
All	153	610	763

The subtypes are unbalanced. Group 4 subtypes hold around a hundred tumours each. WNT beta holds 21 in total and 4 in the test split. This matters for every subtype result that follows.

Related work

The idea was worked out from the data and the four questions above. For completeness: a published R package, MBMethPred (Sharif Rahmani et al. 2023, MBMethPred, Frontiers in Genetics, https://doi.org/10.3389/fgene.2023.1233657), classifies the four subgroups from the same GSE85212 cohort and explains its models with Shapley values. It is cited here as prior work, not as a template. The subtype task, the gene-by-gene test against expression, calibration, sparse coverage and the drug lookup are not part of it.

Page 1 of 9. index | next: 2. Quality control