A novel principal component based method for identifying differentially methylated regions in Illumina Infinium MethylationEPIC BeadChip data

Figures & data

Table 1. Summary of Representative Genomic Regions Used in Power Simulation.

Region ID	Listed Region	Length in bp	# Probes	AAC	MAC	% Variance Explained		Probe with highest absolute PC loading (causal probe)		SD added	SD added
						PC1	PC2	PC1	PC2	PC1	PC2
1	Chr6:30038712–30039600	888	33	0.62	0.98	68.21%	5.08%	cg03343571	cg22184136	7.5	2
2	Chr6:31125920–31126373	453	17	0.054	0.31	10.48%	8.19%	cg01190171	cg15724113	2	4.5
3	Chr7:27183133–27184737	1604	36	0.75	0.95	77.77%	3.03%	cg17569124	cg03744763	4.5	2.5
4	Chr1:248100585–248100614	29	4	0.96	0.99	96.85%	1.88%	cg20507276	cg00785941	15	12
5	Chr19:8117875–8117966	91	3	0.45	0.95	67.30%	30.93%	cg11245297	cg21743830	5	1.5
6	Chr10:8095121–8096372	1251	31	0.085	0.31	10.81%	4.76%	cg09728012	cg23943136	4	3
7	Chr16:67312928–67313043	115	7	0.094	0.30	23.64%	15.38%	cg06297958	cg07498606	1.5	7
8	Chr15:72104228–72104417	189	4	0.13	0.30	30.06%	27.21%	cg06546820	cg13703253	2	5

Note: *MAC: maximum absolute pairwise correlation; AAC: average absolute pairwise correlation; SD added: the standard deviation of the random noise added to the methylation (M-value) at the ‘causal’ probe.

Figure 1. Absolute Correlations between probes in Region 1.

Note: *Region 1: Chr6:30038712-30039600.

*PC1, PC2: absolute values of PC loadings on PC1 and PC2.

*cg03343571 and cg22184136 were the most loaded probes on PC1 and PC2, where cg22184136 was dropped out by coMethDMR.

Table 2. Comparison of Numbers of Regions and DMRs Reported by DMR_PC and coMethdmr.

Cohort	Phenotype	DMRPC		coMethDMR				High correlation Regions (Analyzed by Both Methods)					Moderate correlation Regions (Analyzed by DMRPC Only)
		# Regions	# DMRs	# Regions	# DMRs	# Raw Regions	# Raw DMRs	# Regions	MAC Median (IQR)	# DMRs			# Regions	MAC Median (IQR)	# DMRs
										DMRPC	coMethDMR	Both
Discovery	Age	19799	8961	8423	2834	8632	2923	8423	0.75 (0.63–0.85)	4252	2834	2644	11376	0.44 (0.36–0.57)	4709
	Sex	19799	1100	8423	580	8632	583	8423	0.75 (0.63–0.85)	675	580	394	11376	0.44 (0.36–0.57)	425
	Smoking (0/1)	19704	42	8289	45	8494	45	8289	0.75 (0.63–0.85)	21	45	9	11415	0.44 (0.36–0.56)	21
Replication	Age	32251	17826	13616	5803	14014	5968	13616	0.73 (0.61–0.85)	8812	5803	5683	18635	0.46 (0.38–0.56)	9014
	Sex	32251	6282	13616	2398	14014	2416	13616	0.73 (0.61–0.85)	3419	2398	1818	18635	0.46 (0.38–0.56)	2863
	Smoking (0/1/2)	31914	17	13319	6	13719	6	13319	0.74 (0.61–0.86)	14	6	3	18595	0.45 (0.38–0.56)	3

Note: IQR= interquartile range.

*Under coMethDMR, raw regions refer to genomic regions from the original output from coMethDMR including those subregions, regions refer to genomic regions with subregions combined.

Figure 2. The Genome-wide False Positive Rates on Genic Regions using 100 Discovery Cohort Subjects.

Note: *Genome-wide FP: genome-wide false positive rate.

*Minimal MAC:minimal maximum absolute pairwise correlation of a genomic region

*80%, 90%, 95%, 99%:minimal variance explained by PCs used.

*MetaPC:meta-analysis using multiple PCs, MultiPC: multivariate regression using multiple PCs.

*MetaPC1:meta-analysis using 1^st PC only, MultiPC1: multivariate regression using 1^st PC only.

Figure 3. The Genome-wide False Positive Rates on Genic Regions using All Discovery Cohort Subjects (n = 528).

Note: *Genome-wide FP: genome-wide false positive rate.

*Minimal MAC:minimal maximum absolute pairwise correlation of a genomic region × 80%, 90%, 95%, 99%: minimal variance explained by PCs used.

*MetaPC:meta-analysis using multiple PCs, MultiPC: multivariate regression using multiple PCs.

*MetaPC1:meta-analysis using 1st PC only, MultiPC1: multivariate regression using 1st PC only.

Figure 4. The True Positive Rates for Continuous Signals on Representative Regions: a.Region 1; b. Region 2; c. Region 3.

Note: Legend: *80%, 90%, 95%, 99%:minimal variance explained by PCs used.

*MetaPC:meta-analysis using multiple PCs, MultiPC: multivariate regression using multiple PCs.

*MetaPC1:meta-analysis using 1st PC only, MultiPC1: multivariate regression using 1st PC only.

*CT_PC1,CT_PC2,CT_PC1+PC2: continuous true positive signals simulated associated with PC1, PC2, and PC1+PC2.

Table 3. Replicate ‘Novel’ DMRs without EWAS Hits in the Discovery cohort that were only identified by DMR_PC or coMethdmr.

Phenotype	Novel DMRs Uniquely Identify by DMRPC										Novel DMRs Uniquely Identify by coMethDMR
	High Correlation Regions					Moderate Correlation Regions					High Correlation Regions
	# All	# with EWAS Replication	# with DMR Replication	with EWAS/DMR Replication		# All	# with EWAS Replication	# with DMR Replication	with EWAS/DMR Replication		# All	# with EWAS Replication	# with DMR Replication	with EWAS/DMR Replication
				n	%				n	%				n	%
Age	58	51	43	52	89.66	49	44	36	44	89.80	29	22	14	22	75.86
Sex	52	32	29	32	61.54	62	41	21	41	66.13	88	63	39	64	72.73
Smoking	1	0	0	0	0	2	0	0	0	0	14	1	0	1	7.14

Table 4. Comparison of Relative Computational Burden between DMR_PC and coMethdmr.

Cohort	Phenotype	# Subjects	DMRPC		coMethDMR
			Peak Memory in GB	Time in hours	Peak Memory in GB	Time in hours
Discovery	Age	528	31.93	9.69	29.11	7.76
	Sex	528	31.93	9.46	29.11	7.11
	Smoking (0/1)	400	31.69	7.77	29.11	7.06
Replication	Age	647	38.60	12.03	34.87	8.94
	Sex	647	38.60	11.99	34.87	8.55
	Smoking (0,1,2)	461	38.49	9.41	34.96	7.60

Note: *Results in coMethDMR were based on two functions: CoMethAllRegions() and lmmTestAllRegions().

*Results should be interpreted relatively, as they may vary depending on the computer system used. The estimates of peak memory and running time were computed by the shared compute nodes at Boston University Shared Computing Cluster.

Figure 5. Visualization of Two Age-related DMRs in the Discovery Cohort a.DMR: Chr6:11044877 –11,044,974 (${\mathbf{P}}_{\mathbf{F}\mathbf{D}\mathbf{R}}=3.31\times {10}^{-64}$). b.DMR: Chr3:147125712 –147,127,193 (${\mathbf{P}}_{\mathbf{F}\mathbf{D}\mathbf{R}}=4.69\times {10}^{-32}$).

Note: Legend: *The DMR Chr6:11044877-11044974 was the most significant age-related DMR in the Discovery cohort ${\mathrm{P}}_{\mathrm{F}\mathrm{D}\mathrm{R}}=3.31\times {10}^{-64}$ with 4 probes available and a MAC of 0.77. In this region, 2 PCs were adopted in the DMR_PC analysis to explain 86.08% total variance of methylation residuals.

*The DMR Chr3:147125712-147127193 (${\mathrm{P}}_{\mathrm{F}\mathrm{D}\mathrm{R}}=4.69\times {10}^{-32}$) had 28 probes available within the region and a MAC of 0.65. In the region, 10 PCs were adopted in the DMRPC analysis to explain 64.85% total variance of methylation residuals.

*Only PCs with nominal p-values<0.05 and probes (at least 2) on each PC with absolute PC loadings above the estimated 50% quantile were plotted. Note, weight signs (±) are arbitrary in PCA.

Data availability statement

The datasets analysed during the current study are not publicly available. Qualified investigators can apply to the PTSD Genetics and TRACTS data repositories to gain access to these data via a Data Use Agreement. Please contact Dr. MW Miller regarding access to methylation data from PTSD Genetics and TRACTS data repositories.