A conventional search for marginal gene (G) effects influencing risk to CP in the total sample of 550 trios showed no markers achieved significance at a strict genome-wide level (p≤10−7
). Supporting Figure 1
presents a Manhattan plot for all autosomal and X-linked SNPs. There were 7 SNPs with asymptotic p~10−6
, and two of these mapped to the DSC3
gene on chr. 18, which contained several additional SNPs giving p<0.001. Four other genes (TPP2, HTR1B, BCL6
) had a single SNP with p~10−6
, plus at least one other SNP yielding p<0.001. Thus, while the initial GWAS failed to yield evidence of individual genes controlling risk to CP in these 550 case-parent trios (see QQ plot in Supporting Figure 2
), several SNPs did yield suggestive evidence.
A genome-wide screen for GxE interaction was carried out using PBAT under the strategy described above, where both the 2 df test for gene (G) and gene-environment (GxE) interaction and the 1 df test for (GxE) interaction alone were examined. This screening process revealed several markers achieving genome-wide significance, especially for the 1 df test of GxE interaction (Supporting Figure 3
). To further investigate this evidence, presents “double Manhattan plots” to summarize evidence for G and GxE interaction effects for each of the three maternal exposures ( for GxAlcohol; for GxSmoking and for GxVitamin). In these plots, the bottom half shows the log10
(p-value) for the conventional family based test of SNP effects ignoring exposure
(where more significant results fall farther below the mid-line). In the top half, the −log10
(p-values) are presented for each autosomal SNP from both the 2 df test of G and GxE interaction together (red dots) and the 1 df test for GxE interaction alone (blue dots). Only SNPs yielding asymptotic p<0.0001 in either of these two tests were included to minimize clutter. Dashed lines connect p-values from the marginal test ignoring exposures (below mid-line) to those models considering GxE interaction in either the 1 df test or the 2 df test (above mid-line). Here the very strongest signals for G effects ignoring exposures were omitted (i.e. p-value<0.00001 in the conventional TDT) to highlight those SNPs showing evidence of GxE interaction, which favors identifying genes suggesting GxE interaction. We focused on SNPs in genes yielding p<10−6
in one or another test for GxE interaction for further analysis. As seen in , 8 markers gave p<10−6
in the 1 df test for GxAlcohol interaction among autosomal SNPs, including 3 in MLLT3
on chr. 9q22.
Figure 1 Double Manhattan plots for SNP effects ignoring maternal exposures (lower half) and considering G and G×E interaction for three maternal exposures. Blue dots represent –log10(p-value) from 1 df test of G×E interaction alone; red (more ...)
lists all genes (including 3 pseudo-genes and 1 open reading frame) showing evidence of GxE interaction for any of the three maternal exposures at p≤10−6, along with a total count of SNPs mapping to this gene and the count of additional SNPs yielding p<0.01 in either the 1 df test for GxE alone or the 2 df test for G and GxE interaction together. We dropped 5 genes with one SNP each yielding p<10−6 in either test for GxE interaction but ≤3 additional SNPs showing p<10−2 (AGXT2, HMP19, PRDM14, BTN2A and ETV6). While we examined all genes listed in (including pseudo-genes), here we focus on recognized genes showing evidence of GxE interaction (noted in bold in and labeled in ).
Table III Genes yielding p-values <10−6 in either the 2 df test for G and GxE or the 1 df test for GxE with one or more maternal exposures in genome wide screen using PBAT on 550 CP case-parent trios. Total counts of SNPs mapping to genes and numbers (more ...)
When considering maternal alcohol exposure, MLLT3
on chr. 9 yielded evidence of GxAlcohol interaction. A total of 144 SNPs mapped to MLLT3
, and 3 of these gave strong evidence of GxAlcohol interaction in the 1 df test (rs4621895, p=1.9*10−7
; rs668703, p=6.6*10−7
; and rs4977433, p=1.7*10−6
). A cluster of 7 adjacent SNPs showed evidence of G and GxE interaction, even though none were significant when exposure to maternal alcohol was ignored (see column 3 in Supporting Table I
). Six of these 7 SNPs yielded nominal significance (p<0.05) in both the 1 df test for GxE interaction alone and in the 2 df test for combined effects of G and GxE interaction (last 2 columns of Supporting Table I
). Pairwise linkage disequilibrium (LD) as measured by r2
in Asian and European parents separately could not account for this pattern alone (Supporting Figure 4
Conditional logistic regression models were used to estimate the odds ratio of having CP given the infant carried one risk allele in the absence of exposure [OR(CP|G no E)] and in its presence [OR(CP|G and E)], along with their 95%CI. Estimated OR(CP|G no E) and OR(CP|C and E) for these 7 SNPs in MLLT3
are presented in along with p-values from a LRT for the 1 df test in the conditional logistic regression framework. Here, an additive model was used and the apparent `high-risk allele' became the target allele (which was the minor allele for rs4621895, rs4977433, rs648703 and rs2780841, but the major allele for rs10757142 and rs6475464 -- see Supporting Table I
). Estimated OR(CP|G and E) and their 95%CI for a heterozygous child exposed to maternal alcohol consumption were distinctly higher (open circles) compared to a similar unexposed child (solid circles).
Figure 2 Estimated OR(CP|G no E) and OR(CP|G and E) for maternal alcohol consumption from logistic regression on 550 case-parent trios. P-values from 1 df LRT for G×E interaction are shown along the X axis. Panel A: Estimated OR(CP|G no E) and OR(CP|G (more ...)
Although none 141 SNPs mapping to SMC2
on chr. 9q31.1 were significant at the p<0.0001 level when maternal exposures were ignored, some did achieve nominal significance. When maternal alcohol consumption was considered, SNP rs1536895 yielded p=1.53*10−8
in the 1 df test for GxE interaction from PBAT and an adjacent SNP approached genome wide significance (rs10125685, p=9.83*10−6
). These SNPs identified a region spanning 11kb (and including 6 SNPs) where evidence of GxAlcohol interaction was apparent. When conditional logistic regression models were used to estimate exposure specific ORs, all 6 of these SNPs suggested modest G effects ignoring exposure (see column 3 of Supporting Table II
for p-values from the LRT), and 5 of these 6 were also significant in the 2 df test for combined G and GxAlcohol interaction. shows estimated OR(CP|G no E) and OR(CP|G and E), plus their 95%CI, for these 6 SNPs under an additive model. For 4 separate SNPs, the putative high risk allele was associated with a doubling of risk when the fetus was exposed to maternal alcohol consumption, although some confidence intervals were quite wide.
Eighteen SNPs mapped to TBK1
on chromosome 12q14.2, but only one was nominally significant when exposure to maternal smoking was ignored (rs2141765; p=0.0095). However, 4 SNPs were significant at p<0.01 in the 2 df test for G and G×Smoking interaction, and 6 were significant at this level in the 1 df test for G×Smoking interaction alone (including rs7969932 with p=7.86*10−8
), forming a cluster of 9 SNPs spanning 30 kb. In a conditional logistic regression model, 6 of these 9 SNPs were nominally significant in the 2 df test for combined effects of G and G×E interaction, and 5 were significant in the 1 df test for G×E interaction alone ( which shows p-values for both the 1 df and 2 df tests; also Supporting Table III
Figure 3 Estimated OR(CP|G no E) and OR(CP|G and E) considering G×E interaction with maternal smoking in logistic regression on 550 case-parent trios. Panel A: Estimated OR(CP|G no E) and OR(CP|G and E) for 9 SNPs in TBK1. P-values from 2 df LRT for G (more ...) ZNF236
on chromosome 18q22–q23 encompassed 39 SNPs, one of which yielded evidence of influencing risk when exposures were ignored (rs470337, p=0.015). However, rs372075 gave p=6.75*10−8
in the 1 df test for G×E interaction and rs470563 gave p=6.91*10−6
in the 2 df test. A block of 10 SNPs (spanning 57 kb) was examined using conditional logistic regression models, and again rs470337 was significant when maternal smoking was ignored (p-value=0.016; see Supporting Table IV
). However, when maternal smoking was considered, 7 of these 10 SNPs were significant in the 2 df test for G and G×E interaction combined (Supporting Table IV
). These SNPs showed no evidence of influencing risk for unexposed infants, i.e. the 95%CI of the OR(CP|G no E) always overlapped the null hypothesis value of one. shows estimated OR(CP|G and E) were distinctly higher for 6 of these SNPs.
Figure 4 Estimated OR(CP|G no E) and OR(CP|G and E) for 11 SNPs in BAALC considering G×E interaction with maternal vitamin supplementation in logistic regression on 550 CP case-parent trios. P-values from 1 df LRT for G×E interaction are shown (more ...)
Among the 61 SNPs mapping to BAALC
on chr. 8q22.3, a block of 11 SNPs (spanning 34kb) yielded one SNP with significant evidence in the 1 df test for G×E interaction (rs6468862; p=2.03*10−7
), plus 4 additional SNPs yielding nominal significance from PBAT. Only rs6468862 showed strong evidence of influencing risk in conditional logistic regression when maternal exposures were ignored (see column 3 of Supporting Table V
). However, when G×E interaction was included in the model, 6 SNPs became significant in either the 1 or 2 df test (Supporting Table V
), and the estimated OR(CP|G no E) and OR(CP|G and E) were distinct (see ).
Among the genes in , OBSCN
on chr. 1q42.13 and ACOXL on chr. 2q13 deserve additional mention. The 27 SNPs mapping to OBSCN
included 22 SNPs spanning 114 kb, 18 of which were nominally significant for G effects ignoring maternal exposures. When G×Smoking interaction was included, 17 of these 22 showed significant evidence in the 2 df test for G and G×Smoking interaction combined. Thus, OBSCN
may represent “quantitative interaction” where exposure to maternal smoking enhances G effects. Among all the genes considered, only OBSCN showed any evidence of heterogeneity between trios of Asian and European ancestry in formal tests of heterogeneity considering G×Smoking interaction (data not shown). Because exposure to smoking is much lower among Asian mothers, however, it is difficult to confirm the absence of G×Smoking interaction in this group. Supporting Figure 5
shows estimated OR(CP|G no E) and OR(CP|G and E) for these 22 SNPs. A total of 126 SNPs mapped to the ACOXL
gene, and these included a block of 24 adjacent SNPs (spanning 64 kb) of which 11 were significant in the 1 df test for G×Vitamin interaction (including rs7602030, p=3.13*10−7
). When conditional logistic regression models were used to estimate OR(CP|G no E) and OR(CP|G and E), 10 SNPs in ACOXL
showed significantly lower risk to the child if the mother used multivitamin supplements (see Supporting Figure 6
Supporting Figures 7–10
illustrate estimated effects as OR(CP|G no E) and OR(CP|G and E) for 3 pseudo-genes (LOC645762 on chr. 4; LOC391828 on chr. 5; and LOC392027 on chr. 7) and for the open reading frame c6orf105 on chr. 6. Each of these putative genes met the criteria used to select genes giving evidence of G×E interaction, but 2 of these 3 pseudo-genes (LOC645762 and LOC391828) have been dropped from the latest version of the human genome (Build 37) and the inferred status of the remaining pseudo-genes makes it difficult to assess their true relevance.