4. Explainability

Page 4 of 9. previous: 3. Classification | index | next: 5. Methylation versus expression

A classifier that is right is not yet useful for biology. The question here is which CpG sites carry each subgroup call, and whether those sites sit in genes that make sense.

Integrated gradients attribute a prediction to the input features. The method moves the input from a baseline (here the mean training profile) to the actual tumour in small steps and sums the gradient of the predicted class along the path. The result is one number per CpG: positive when the tumour's methylation at that site pushes the call towards the class, negative when it pushes away. Attributions were computed on the four-group MLP and averaged over the correctly classified training tumours of each subgroup.

Each CpG is annotated from the Illumina manifest with its gene, its position relative to the gene (promoter regions TSS200, TSS1500, 5'UTR and first exon; or gene body), and whether the subgroup is hyper- or hypomethylated at that site relative to the other subgroups.

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)

display(Image(FIG / 'top_cpgs_per_group.png'))

top = pd.read_csv(TAB / 'top_cpgs_per_group.csv')
top.groupby('group').head(10)[['group', 'cpg', 'gene', 'region', 'cpg_island', 'direction', 'mean_attribution', 'delta_beta_vs_rest']]

	group	cpg	gene	region	cpg_island	direction	mean_attribution	delta_beta_vs_rest
0	WNT	cg20020844	SP6	5'UTR	S_Shore	hypermethylated	0.01680	0.64205
1	WNT	cg02704158	NaN	NaN	N_Shore	hypermethylated	0.01575	0.68853
2	WNT	cg06590268	NaN	NaN	Island	hypermethylated	0.01559	0.61425
3	WNT	cg08031151	NaN	NaN	N_Shore	hypermethylated	0.01549	0.66551
4	WNT	cg15892785	GPR153	3'UTR	Island	hypermethylated	0.01489	0.60862
5	WNT	cg17769084	NaN	NaN	Island	hypermethylated	0.01480	0.66948
6	WNT	cg14174502	NaN	NaN	NaN	hypomethylated	0.01462	-0.61974
7	WNT	cg22855020	SLC20A2	5'UTR	NaN	hypomethylated	0.01451	-0.71315
8	WNT	cg07188009	NaN	NaN	S_Shelf	hypomethylated	0.01423	-0.59633
9	WNT	cg03807298	NaN	NaN	Island	hypermethylated	0.01410	0.75737
50	SHH	cg17741809	LIPA	Body	NaN	hypermethylated	0.01870	0.76322
51	SHH	cg09060654	LIPA	Body	NaN	hypermethylated	0.01599	0.67444
52	SHH	cg10959440	SASH1	Body	NaN	hypermethylated	0.01566	0.82187
53	SHH	cg20406460	NaN	NaN	Island	hypermethylated	0.01555	0.75759
54	SHH	cg14060518	NaN	NaN	S_Shore	hypermethylated	0.01550	0.85124
55	SHH	cg20608895	NaN	NaN	NaN	hypermethylated	0.01458	0.79116
56	SHH	cg09624466	OTX2OS1	TSS1500	Island	hypermethylated	0.01419	0.73754
57	SHH	cg24699146	HMGCL	TSS1500	NaN	hypermethylated	0.01393	0.67032
58	SHH	cg15416279	NaN	NaN	Island	hypermethylated	0.01384	0.68151
59	SHH	cg21039708	OTX2OS1	TSS1500	Island	hypermethylated	0.01362	0.74751
100	Group3	cg12565585	RIMS2	Body	Island	hypermethylated	0.02691	0.66683
101	Group3	cg13993183	NaN	NaN	S_Shore	hypermethylated	0.02420	0.62160
102	Group3	cg25900085	MIR124-2	TSS200	S_Shore	hypomethylated	0.02405	-0.60381
103	Group3	cg26939946	RNF112	TSS200	NaN	hypermethylated	0.02381	0.62858
104	Group3	cg06807993	MIR657	TSS1500	N_Shelf	hypermethylated	0.02117	0.56362
105	Group3	cg05437476	NaN	NaN	Island	hypermethylated	0.02113	0.58834
106	Group3	cg13899017	CADM3	Body	Island	hypermethylated	0.02093	0.51826
107	Group3	cg12880095	NaN	NaN	NaN	hypomethylated	0.02084	-0.53614
108	Group3	cg04045327	GAK	Body	N_Shore	hypermethylated	0.02045	0.64178
109	Group3	cg06925984	NaN	NaN	Island	hypermethylated	0.02043	0.55999
150	Group4	cg25624478	NFASC	Body	N_Shelf	hypomethylated	0.01870	-0.66123
151	Group4	cg11033271	CDCA7L	Body	N_Shore	hypermethylated	0.01854	0.64637
152	Group4	cg24615189	KCNIP1	TSS1500	NaN	hypermethylated	0.01580	0.55875
153	Group4	cg26947766	KCNIP1	TSS1500	NaN	hypermethylated	0.01495	0.52704
154	Group4	cg22365240	NaN	NaN	NaN	hypermethylated	0.01431	0.63464
155	Group4	cg05494467	PCDHGB5	3'UTR	N_Shore	hypomethylated	0.01383	-0.49128
156	Group4	cg23146346	RBM24	TSS200	S_Shore	hypomethylated	0.01340	-0.54943
157	Group4	cg04585185	C21orf34	Body	NaN	hypomethylated	0.01328	-0.44517
158	Group4	cg16918905	NaN	NaN	Island	hypomethylated	0.01316	-0.53323
159	Group4	cg01968133	CASZ1	Body	N_Shelf	hypomethylated	0.01310	-0.64693

The full list, 50 CpGs per subgroup, is in results/tables/top_cpgs_per_group.csv. Two things stand out. First, most top CpGs are hypomethylated in their subgroup. Loss of promoter methylation switches a gene on, so the model mostly recognises subgroups by programs that are turned on, not off. Second, many top CpGs sit in genes of neuronal and developmental identity rather than in the known driver genes. This is expected. Drivers such as CTNNB1, PTCH1 or MYC are defined by mutation or amplification. Methylation records lineage, and lineage is what separates the subgroups.

Whether these sites actually control expression is the question of the next page.

Page 4 of 9. previous: 3. Classification | index | next: 5. Methylation versus expression