Pruning and thresholding approach for methylation risk scores in multi-ancestry populations

Figures & data

Figure 1. Simulation study. Prediction R2 of methylation risk scores (MRS) estimated with pruning and thresholding + Co-Methylation with genomic CpG Background ((P+T CoMeBack), P+T and thresholding (T) method in dependence of (A) the proportion of causal CpG sites in co-methylation regions (CMRs) and (B) proportion of phenotype variance explained by DNA methylation, among Indian participants. For each simulation, the discovery cohort was repeatedly and randomly split into a training set comprising 762 Indians and a testing set comprising 136 people of the same ancestry. Phenotypes were simulated without an influence of ancestry. Results are shown for (a) different proportions of causal CpGs located in CMR (30%, 50%, 70%, 100%) and (b) different proportions of phenotype variance explained by DNA methylation (10%, 30%, 50%, 80%). Each box represents the distribution of prediction accuracy across 1000 simulations, where the central mark is the median and the edges of the box are the 25^th and 75^th percentiles.

Table 1. Overview of included EWAS, their phenotypes, training sample and methods.

MRS	Training dataset publication	Training Population	Phenotype	MRS publication	P-value threshold/ Method	No. of CpG sites (joint-analysis)
P+T CoMeBack MRS	Sikdar et al. 2019 ^[Citation36]	Multi-ethnic newborns (mainly White, N = 5,648)	Most cohorts ascertained sustained smoking during pregnancy by questionnaires; two cohorts incorporated cotinine-based smoking measure	Proposed in this article	5×10^−22 (Mixed)	21 (43 passed P-value threshold and 22 excluded by pruning)
					5×10^−24 (Black)	20 (42 passed P-value threshold and 22 excluded by pruning)
					5×10^−22 (Pooled)	21 (43 passed P-value threshold and 22 excluded by pruning)
P+T MRS	Sikdar et al. 2019 ^[Citation36]	Multi-ethnic newborns (mainly White, N = 5,648)	Most cohorts ascertained sustained smoking during pregnancy by questionnaires; two cohorts incorporated cotinine-based smoking measure	Proposed in this article	5×10^−10 (Mixed)	198 (233 passed P-value threshold and 35 excluded by pruning)
					5×10^−36 (Black)	4 (26 passed P-value threshold and 22 excluded by pruning)
					5×10^−24 (Pooled)	8 (42 passed P-value threshold and 34 excluded by pruning)
T MRS	Sikdar et al. 2019 ^[Citation36]	Multi-ethnic newborns (mainly White, N = 5,648)	Most cohorts ascertained sustained smoking during pregnancy by questionnaires; two cohorts incorporated cotinine-based smoking measure	Proposed in this article	5×10^−9 (Mixed)	344
					5×10^−24 (Black)	42
					5×10^−16 (Pooled)	72
Reese MRS	Reese et al. 2017 ^[Citation37]	White newborns (N = 1,068)*	Sustained smoking during pregnancy obtained from combined information of cotinine-based and self-report based classification	Reese et al. 2017 ^[Citation37]	Logistic LASSO regression	28
Richmond 568 MRS	Joubert et al. 2016 ^[Citation38]	Multi-ethnic newborns (N = 6,685)*	Maternal smoking during pregnancy via questionnaires	Richmond et al. 2018 ^[Citation39]	Robust linear regression; Bonferroni corrected P-value<0.05	568
Richmond 19 MRS	Joubert et al. 2016 ^[Citation38]	Multi-ethnic older children (average age = 6.8 years) (N = 3,187)	Maternal smoking during pregnancy via questionnaires	Richmond et al. 2018	Robust linear regression; Bonferroni corrected P-value<0.05	19

Note: * These training populations overlapped with training population for summary statistics used for P+T MRS.

Abbreviations: Methylation risk scores (MRS); Pruning and thresholding (P+T); Co-Methylation with genomic CpG Background (CoMeBack); Thresholding (T).

Figure 2. Simulation study. Prediction R² of methylation risk scores (MRS) estimated with pruning and thresholding + Co-Methylation with genomic CpG Background (P+T CoMeBack), P+T and thresholding (T) approach across different racial groups and among multi-ancestry populations. For each simulation, the discovery cohort was repeatedly and randomly split into a training set comprising 762 Indians and a testing set comprising 136 people of each ancestry group. The proportion of causal CpGs located in co-methylation regions (CMR) is 70% and the proportion of phenotype variance explained by DNA methylation (and ancestry) is 80%. Results are shown for the prediction of simulated phenotypes (a) without an influence of ancestry and (b) influenced by ancestry. Joint-analysis refers to MRS analyses of all participants pooled from all ancestry groups and standardization refers to standardizing MRS within each ancestry group and then merging all participants before analyses. Each box represents the distribution of prediction accuracy across 1000 simulations, where the central mark is the median and the edges of the box are the 25^th and 75^th percentiles.

Figure 3. Real data application. Methylation risk scores (MRS) for the prediction of maternal smoking during pregnancy using cord blood DNA methylation data from newborns in the South African Drakenstein Child Health Study (DCHS). Prediction R² of maternal smoking status is shown stratified for a. Mixed infants. b. Black infants. c. joint-analysis (all subjects pooled from all ancestries) d. Standardization (standardizing MRS within each ancestry and merging all subjects before analyses.

Figure 4. Real data application. Comparison of MRS estimated with pruning and thresholding+ Co-Methylation with genomic CpG Background (P+T CoMeBack), P+T, Thresholding (T) and 3 other published methylation risk scores (MRS) for predicting maternal smoking status in the South African Drakenstein Child Health Study (DCHS). a. A Prediction R² of all 6 MRS methods for Mixed infants, Black infants and pooled samples (joint-analysis). A receiver operating characteristic (ROC) curve comparing prediction performance of all 6 MRS among (b) Mixed infants, (c) Black infants and (d) pooled samples (joint-analysis).

Table 2. Association between maternal smoking status and MRS in DCHS.

MRS	Mixed			Black			Pooled (Joint-analysis)
	Beta-coefficient*	Standard Error	P-value	Beta-coefficient	Standard Error	P-value	Beta-coefficient	Standard Error	P-value
P+T CoMeBack MRS	0.88	0.12	5.65 ×10^−11	0.76	0.18	5.74 ×10^−5	0.82	0.10	3.01×10^−14
P+T MRS	0.71	0.11	1.05 ×10^−8	0.52	0.14	3.29 ×10^−4	0.69	0.10	1.32×10^−10
T MRS	0.80	0.13	7.35 ×10^−9	0.70	0.19	2.82 ×10^−4	0.64	0.09	2.55 ×10^−11
Reese MRS	0.97	0.13	3.77 ×10^−11	0.73	0.16	2.06 ×10^−5	0.77	0.09	2.50 ×10^−15
Richmond 568 MRS	0.85	0.14	9.84 ×10^−9	0.54	0.16	7.82 ×10^−4	0.65	0.10	1.57 ×10^−10
Richmond 19 MRS	0.76	0.14	5.37E x10^−7	0.70	0.19	3.78 ×10^−4	0.70	0.11	2.05 ×10^−9

*Beta-coefficients indicate change in interquartile range (IQR) of MRS between smokers and non-smokers.

Abbreviations: Methylation risk scores (MRS); South African Drakenstein Child Health Study (DCHS); Pruning and thresholding (P+T); Co-Methylation with genomic CpG Background (CoMeBack); Thresholding (T).

Sikdar S, Joehanes R, Joubert BR, et al. Comparison of smoking-related DNA methylation between newborns from prenatal exposure and adults from personal smoking. Epigenomics. 2019 Oct;11(13):1487–1500.

 PubMed Web of Science ®Google Scholar

Reese SE, Zhao S, Wu MC, et al. DNA methylation score as a biomarker in newborns for sustained maternal smoking during pregnancy. Environ Health Perspect. 2017 Apr;125(4):760–766.

 PubMed Web of Science ®Google Scholar

Joubert BR, Felix Janine F, Yousefi P, et al. DNA methylation in newborns and maternal smoking in pregnancy: genome-wide consortium meta-analysis. Am J Hum Genet. 2016;98(4):680–696.

 PubMed Web of Science ®Google Scholar

Richmond RC, Suderman M, Langdon R, et al. DNA methylation as a marker for prenatal smoke exposure in adults. Int J Epidemiol. 2018;47(4):1120–1130.

 PubMed Web of Science ®Google Scholar

Data availability statement

The data for simulation studies were derived from the following dataset (GSE55763, GSE84727, GSE80417, GSE111629 and GSE72680) from NCBI Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo/). The data for real data application are available from the corresponding author, [AH], upon reasonable request.