Limits...
Confounding by repetitive elements and CpG islands does not explain the association between hypomethylation and genomic instability.

Harris RA, Shaw C, Li J, Cheung SW, Coarfa C, Jeong M, Goodell MA, White LD, Patel A, Kang SH, Chinault AC, Gambin T, Gambin A, Lupski JR, Milosavljevic A - PLoS Genet. (2013)

View Article: PubMed Central - PubMed

Affiliation: Bioinformatics Research Laboratory, Epigenome Center, Baylor College of Medicine, Houston, Texas, USA.

AUTOMATICALLY GENERATED EXCERPT
Please rate it.

In our recent article, we reported an association between hypomethylation and genomic instability... Once the regions have been removed from the genome, Watson et al. claim the association between hypomethylation and genomic instability disappears... We applied the same method to all five sample sets brought up by Watson et al. —HapMap270, HapMap450, WTCCC, Schizophrenia Cases, and Schizophrenia Controls... As documented in Tables S1 and S2, in all five sample sets hypomethylation remained highly significantly associated with CNV density after correction for all of the “confounders” individually and in combination... All three models gave consistent results for all confounders in all five sample sets. (Table S2 describes regression models, output, and the input data extracted from our original paper sufficient to run a statistical program such as R to obtain the output.) We also employed zero-inflated negative binomial and Poisson regression models, and found completely concordant results... We therefore conclude that the assertions regarding “confounding” are not consistent with the data available... In this context it is surprising that Watson et al. fail to mention that our article considered, examined, and ruled out this possibility: “One could expect that if the windows with MI = 0 were due to low probing density, the windows within the higher mode would have fewer SNPs or CpGs... However, we examined potential biases in MI estimation due to variations in the number of SNPs, CpGs, read coverage (Figure S6CD), or sampling events (Figure S7BD) and found no significant difference between the two modes, ruling out the possibility that the two modes may be explained by variation in mappability or shallow sampling... Watson et al. ignore this possibility without sound justification while claiming that this pattern somehow provides evidence against any connection of hypomethylation and genomic instability... Fifth, contrary to what Watson et al. claim, our article does not state that hypomethylation plays a causative role in genomic instability... In summary, we thank Watson et al. for their efforts and further examination of our reported observations... Nevertheless, we find that the arguments put forward in the comment do not diminish the strength of our reported findings... Specifically, our analyses of the confounding factors suggested by Watson et al. do not diminish the contention that genomic correlates may provide only a partial explanation for the hotspots of genomic instability.

Show MeSH

Related in: MedlinePlus

Predictive power of methylation and other genomic factors for CNV counts.Predictive power of methylation, CpG island content, and repetitive element content (LINE, SINE, LTR, and Satellites) was measured using Akaike information criterion (AIC). For all five datasets, negative binomial regression was performed using all six factors and all six combinations of five factors (one factor being removed at a time). The y-axis represents the predictive power of a factor, as measured by the improvement of the AIC score based on all six factors relative to the AIC score without the factor. Note that this method measures predictive power of a factor after correction for any potential confounding due to other factors. (The detailed calculations and input data are in Supporting Information.)
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC3585018&req=5

pgen-1003333-g001: Predictive power of methylation and other genomic factors for CNV counts.Predictive power of methylation, CpG island content, and repetitive element content (LINE, SINE, LTR, and Satellites) was measured using Akaike information criterion (AIC). For all five datasets, negative binomial regression was performed using all six factors and all six combinations of five factors (one factor being removed at a time). The y-axis represents the predictive power of a factor, as measured by the improvement of the AIC score based on all six factors relative to the AIC score without the factor. Note that this method measures predictive power of a factor after correction for any potential confounding due to other factors. (The detailed calculations and input data are in Supporting Information.)

Mentions: We first applied the Negative Binomial regression model [4], because it is commonly used for overdispersed variables and is a well-accepted and robust method for count data such as CNV counts. We applied the same method to all five sample sets brought up by Watson et al.—HapMap270 [5], HapMap450 [6], WTCCC [7], Schizophrenia Cases, and Schizophrenia Controls [8]. As documented in Tables S1 and S2, in all five sample sets hypomethylation remained highly significantly associated with CNV density after correction for all of the “confounders” individually and in combination. As illustrated in Figure 1, as measured by AIC, methylation was more predictive of CNV counts per 100-Kbp window by an order of magnitude than any other factor.


Confounding by repetitive elements and CpG islands does not explain the association between hypomethylation and genomic instability.

Harris RA, Shaw C, Li J, Cheung SW, Coarfa C, Jeong M, Goodell MA, White LD, Patel A, Kang SH, Chinault AC, Gambin T, Gambin A, Lupski JR, Milosavljevic A - PLoS Genet. (2013)

Predictive power of methylation and other genomic factors for CNV counts.Predictive power of methylation, CpG island content, and repetitive element content (LINE, SINE, LTR, and Satellites) was measured using Akaike information criterion (AIC). For all five datasets, negative binomial regression was performed using all six factors and all six combinations of five factors (one factor being removed at a time). The y-axis represents the predictive power of a factor, as measured by the improvement of the AIC score based on all six factors relative to the AIC score without the factor. Note that this method measures predictive power of a factor after correction for any potential confounding due to other factors. (The detailed calculations and input data are in Supporting Information.)
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC3585018&req=5

pgen-1003333-g001: Predictive power of methylation and other genomic factors for CNV counts.Predictive power of methylation, CpG island content, and repetitive element content (LINE, SINE, LTR, and Satellites) was measured using Akaike information criterion (AIC). For all five datasets, negative binomial regression was performed using all six factors and all six combinations of five factors (one factor being removed at a time). The y-axis represents the predictive power of a factor, as measured by the improvement of the AIC score based on all six factors relative to the AIC score without the factor. Note that this method measures predictive power of a factor after correction for any potential confounding due to other factors. (The detailed calculations and input data are in Supporting Information.)
Mentions: We first applied the Negative Binomial regression model [4], because it is commonly used for overdispersed variables and is a well-accepted and robust method for count data such as CNV counts. We applied the same method to all five sample sets brought up by Watson et al.—HapMap270 [5], HapMap450 [6], WTCCC [7], Schizophrenia Cases, and Schizophrenia Controls [8]. As documented in Tables S1 and S2, in all five sample sets hypomethylation remained highly significantly associated with CNV density after correction for all of the “confounders” individually and in combination. As illustrated in Figure 1, as measured by AIC, methylation was more predictive of CNV counts per 100-Kbp window by an order of magnitude than any other factor.

View Article: PubMed Central - PubMed

Affiliation: Bioinformatics Research Laboratory, Epigenome Center, Baylor College of Medicine, Houston, Texas, USA.

AUTOMATICALLY GENERATED EXCERPT
Please rate it.

In our recent article, we reported an association between hypomethylation and genomic instability... Once the regions have been removed from the genome, Watson et al. claim the association between hypomethylation and genomic instability disappears... We applied the same method to all five sample sets brought up by Watson et al. —HapMap270, HapMap450, WTCCC, Schizophrenia Cases, and Schizophrenia Controls... As documented in Tables S1 and S2, in all five sample sets hypomethylation remained highly significantly associated with CNV density after correction for all of the “confounders” individually and in combination... All three models gave consistent results for all confounders in all five sample sets. (Table S2 describes regression models, output, and the input data extracted from our original paper sufficient to run a statistical program such as R to obtain the output.) We also employed zero-inflated negative binomial and Poisson regression models, and found completely concordant results... We therefore conclude that the assertions regarding “confounding” are not consistent with the data available... In this context it is surprising that Watson et al. fail to mention that our article considered, examined, and ruled out this possibility: “One could expect that if the windows with MI = 0 were due to low probing density, the windows within the higher mode would have fewer SNPs or CpGs... However, we examined potential biases in MI estimation due to variations in the number of SNPs, CpGs, read coverage (Figure S6CD), or sampling events (Figure S7BD) and found no significant difference between the two modes, ruling out the possibility that the two modes may be explained by variation in mappability or shallow sampling... Watson et al. ignore this possibility without sound justification while claiming that this pattern somehow provides evidence against any connection of hypomethylation and genomic instability... Fifth, contrary to what Watson et al. claim, our article does not state that hypomethylation plays a causative role in genomic instability... In summary, we thank Watson et al. for their efforts and further examination of our reported observations... Nevertheless, we find that the arguments put forward in the comment do not diminish the strength of our reported findings... Specifically, our analyses of the confounding factors suggested by Watson et al. do not diminish the contention that genomic correlates may provide only a partial explanation for the hotspots of genomic instability.

Show MeSH
Related in: MedlinePlus