On 2015 Apr 27, Chloe Wong commented:

Thank you for your interest and comments on our manuscript. As we hope is clear from the Discussion section of our paper, we are very conservative in our conclusions and are the first to recognize the many limitations of doing this type of work. Please also see our articles highlighting the many important issues to consider when undertaking and interpreting epigenetic epidemiological analyses (Mill & Heijmans, 2013; Heijmans & Mill, 2012).

There are many reasons why the standard research approaches developed for genetic epidemiology are not necessarily appropriate for epigenetic studies of common disease. To date, no real precedents have been set about the optimal sample-sizes needed to detect epigenetic changes associated with disease. The number of ASD-discordant twin-pairs available for this study was small - these are extremely rare samples, and we were only able to recruit and characterize six discordant monozygotic (MZ) pairs. Furthermore, it is recognized that standard multiple testing parameters (as used in GWAS analyses) are not necessarily appropriate for genome-wide DNA methylation data ­ first, most of the sites on the commonly-used Illumina EWAS array are actually non-variable, and second there is considerable non-independence between DNA methylation at proximal CpG sites. With the aim of identifying real, biologically relevant within-twin and between-group DNA methylation differences, we therefore decided to use an analytic approach that incorporated both the significance (that is, t-test statistic) and magnitude (that is, absolute DNA methylation difference) of any observed differences to produce a ranked list of DMRs. Of note, we confirmed the variation at selected loci using an independent technology (bisulfite-pyrosequencing) to rule out technical artifacts in the data.

Because individual studies such as this are, by necessity, small, it is absolutely clear that findings should be treated with caution until they are replicated and/or validated using complimentary approaches. In this regard, it is noteworthy that one of the top-ranked differentially methylated regions identified in our twin study ­ located in the vicinity of the OR2L13 gene - is also a top-ranked differentially methylated locus in a more recent epigenetic study of ASD by Berko and colleagues (2014). Furthermore, OR2L13 was found to be significantly differentially expressed in post-mortem brain tissue from ASD cases compared to unaffected controls in the most systematic transcriptomic analysis of autism brain yet undertaken (Voineagu et al, 2011). Finally, this gene has also been implicated in autism by genetic studies that have identified recurrent CNVs spanning the locus in cases.

The main concern raised by Professor Bishop is that methylomic differences are also identified within concordant affected and concordant unaffected twins. We would argue this is not necessarily surprising; given that it is not feasible to directly study brain tissue from our twins, and ASD is a subtle developmental brain problem, it's plausible (perhaps likely) that these individuals are discordant for other traits/exposures that are also associated with epigenetic variation detected in blood. That doesn¹t mean that differences specific to ASD discordant twin-pairs are not interesting. What is clear from our analyses is that i) the sites identified as differentially methylated in the six ASD-discordant twin-pairs are not differentially methylated in concordant-unaffected twin-pairs, and ii) the overall distribution of average within-pair DNA methylation differences is significantly skewed in ASD-discordant twins, with a higher number of CpG sites demonstrating a larger average difference in DNA methylation.

We acknowledge that our data represent only the first step in identifying molecular variation associated with autism. For example, we cannot begin to tackle issues regarding causality in this study, and it is possible (perhaps likely) that many of the changes we identified represent consequences of the disease. As discussed, we were also limited to using DNA derived from blood, and moving forward it will be important to understand the utility of peripheral tissues as a proxy for inaccessible organs such as the brain. We are currently undertaking more systematic analyses in larger samples of twins and post-mortem brain to address many of the limitations of this study.

Berko ER, Suzuki M, Beren F, Lemetre C, Alaimo CM, Calder RB, Ballaban-Gil K, Gounder B, Kampf K, Kirschen J, Maqbool SB, Momin Z, Reynolds DM, Russo N, Shulman L, Stasiek E, Tozour J, Valicenti-McDermott M, Wang S, Abrahams BS, Hargitai J, Inbar D, Zhang Z, Buxbaum JD, Molholm S, Foxe JJ, Marion RW, Auton A, Greally JM. Mosaic epigenetic dysregulation of ectodermal cells in autism spectrum disorder. PLoS Genet. 2014 May 29;10(5):e1004402.

Heijmans BT, Mill J. Commentary: The seven plagues of epigenetic epidemiology. Int J Epidemiol. 2012 Feb;41(1):74-8.

Mill J, Heijmans BT. From promises to practical strategies in epigenetic epidemiology. Nat Rev Genet. 2013 Aug;14(8):585-94.

Voineagu I, Wang X, Johnston P, Lowe JK, Tian Y, et al. (2011) Transcriptomic analysis of autistic brain reveals convergent molecular pathology. Nature 474: 380­384.

On 2015 Apr 25, Dorothy V M Bishop commented:

This is a pioneering study using a clever design with a unique dataset. I am grateful to the authors for making their raw data available. A colleague had suggested that I might be able to use the methods described in this paper with some of my own data, and it was good to have the opportunity to work through the analyses and gain more understanding of what was done.

Unfortunately, having done so, I became dubious as to whether the results show differentially methylated regions associated with ASD, as claimed. Over 23,000 sites were examined for methylation differences, and the numbers where discordant twins differed was not high. Most tellingly, when I used the authors' data to analyse concordant groups in similar fashion, the number of methylation differences was similar. This was true both for twins concordant for ASD (where there were 341 between-twin differences with p < .01, compared with 203 such differences in the discordant twins), and for twins who were concordant for low symptom scores on CAST (where there were 188 between-twin differences with p < .01 – here I selected the first 6 twins from this group to give an equivalent sample size to the discordant twins).

In addition, the findings of correlations between CAST scales and levels of methylation at given site is not impressive, given that the number of correlations computed was over 90,000 (4 per site), so some would be expected to achieve low p-values by chance.

I appreciate this is a new area, and exploratory work needs to be done, but given that the field of molecular genetics has learned the importance for controlling for chance findings when looking for associations with SNPs, I am wondering perhaps the same lesson will prove necessary when examining methylation data. With a twin data set such as this one, I would argue it is very useful to have concordant twins as a comparison group, as they can be used to give an indication of the amount of discordance between twins in methylation that is to be expected regardless of phenotype.

I have uploaded the analysis file I used to compare methylation patterns in different groups here: osf.io/z6w92 so that others can check my working.

On 2015 Apr 25, Dorothy V M Bishop commented:

This is a pioneering study using a clever design with a unique dataset. I am grateful to the authors for making their raw data available. A colleague had suggested that I might be able to use the methods described in this paper with some of my own data, and it was good to have the opportunity to work through the analyses and gain more understanding of what was done.

Unfortunately, having done so, I became dubious as to whether the results show differentially methylated regions associated with ASD, as claimed. Over 23,000 sites were examined for methylation differences, and the numbers where discordant twins differed was not high. Most tellingly, when I used the authors' data to analyse concordant groups in similar fashion, the number of methylation differences was similar. This was true both for twins concordant for ASD (where there were 341 between-twin differences with p < .01, compared with 203 such differences in the discordant twins), and for twins who were concordant for low symptom scores on CAST (where there were 188 between-twin differences with p < .01 – here I selected the first 6 twins from this group to give an equivalent sample size to the discordant twins).

In addition, the findings of correlations between CAST scales and levels of methylation at given site is not impressive, given that the number of correlations computed was over 90,000 (4 per site), so some would be expected to achieve low p-values by chance.

I appreciate this is a new area, and exploratory work needs to be done, but given that the field of molecular genetics has learned the importance for controlling for chance findings when looking for associations with SNPs, I am wondering perhaps the same lesson will prove necessary when examining methylation data. With a twin data set such as this one, I would argue it is very useful to have concordant twins as a comparison group, as they can be used to give an indication of the amount of discordance between twins in methylation that is to be expected regardless of phenotype.

I have uploaded the analysis file I used to compare methylation patterns in different groups here: osf.io/z6w92 so that others can check my working.

On 2015 Apr 27, Chloe Wong commented:

Thank you for your interest and comments on our manuscript. As we hope is clear from the Discussion section of our paper, we are very conservative in our conclusions and are the first to recognize the many limitations of doing this type of work. Please also see our articles highlighting the many important issues to consider when undertaking and interpreting epigenetic epidemiological analyses (Mill & Heijmans, 2013; Heijmans & Mill, 2012).

There are many reasons why the standard research approaches developed for genetic epidemiology are not necessarily appropriate for epigenetic studies of common disease. To date, no real precedents have been set about the optimal sample-sizes needed to detect epigenetic changes associated with disease. The number of ASD-discordant twin-pairs available for this study was small - these are extremely rare samples, and we were only able to recruit and characterize six discordant monozygotic (MZ) pairs. Furthermore, it is recognized that standard multiple testing parameters (as used in GWAS analyses) are not necessarily appropriate for genome-wide DNA methylation data ­ first, most of the sites on the commonly-used Illumina EWAS array are actually non-variable, and second there is considerable non-independence between DNA methylation at proximal CpG sites. With the aim of identifying real, biologically relevant within-twin and between-group DNA methylation differences, we therefore decided to use an analytic approach that incorporated both the significance (that is, t-test statistic) and magnitude (that is, absolute DNA methylation difference) of any observed differences to produce a ranked list of DMRs. Of note, we confirmed the variation at selected loci using an independent technology (bisulfite-pyrosequencing) to rule out technical artifacts in the data.

Because individual studies such as this are, by necessity, small, it is absolutely clear that findings should be treated with caution until they are replicated and/or validated using complimentary approaches. In this regard, it is noteworthy that one of the top-ranked differentially methylated regions identified in our twin study ­ located in the vicinity of the OR2L13 gene - is also a top-ranked differentially methylated locus in a more recent epigenetic study of ASD by Berko and colleagues (2014). Furthermore, OR2L13 was found to be significantly differentially expressed in post-mortem brain tissue from ASD cases compared to unaffected controls in the most systematic transcriptomic analysis of autism brain yet undertaken (Voineagu et al, 2011). Finally, this gene has also been implicated in autism by genetic studies that have identified recurrent CNVs spanning the locus in cases.

The main concern raised by Professor Bishop is that methylomic differences are also identified within concordant affected and concordant unaffected twins. We would argue this is not necessarily surprising; given that it is not feasible to directly study brain tissue from our twins, and ASD is a subtle developmental brain problem, it's plausible (perhaps likely) that these individuals are discordant for other traits/exposures that are also associated with epigenetic variation detected in blood. That doesn¹t mean that differences specific to ASD discordant twin-pairs are not interesting. What is clear from our analyses is that i) the sites identified as differentially methylated in the six ASD-discordant twin-pairs are not differentially methylated in concordant-unaffected twin-pairs, and ii) the overall distribution of average within-pair DNA methylation differences is significantly skewed in ASD-discordant twins, with a higher number of CpG sites demonstrating a larger average difference in DNA methylation.

We acknowledge that our data represent only the first step in identifying molecular variation associated with autism. For example, we cannot begin to tackle issues regarding causality in this study, and it is possible (perhaps likely) that many of the changes we identified represent consequences of the disease. As discussed, we were also limited to using DNA derived from blood, and moving forward it will be important to understand the utility of peripheral tissues as a proxy for inaccessible organs such as the brain. We are currently undertaking more systematic analyses in larger samples of twins and post-mortem brain to address many of the limitations of this study.

Berko ER, Suzuki M, Beren F, Lemetre C, Alaimo CM, Calder RB, Ballaban-Gil K, Gounder B, Kampf K, Kirschen J, Maqbool SB, Momin Z, Reynolds DM, Russo N, Shulman L, Stasiek E, Tozour J, Valicenti-McDermott M, Wang S, Abrahams BS, Hargitai J, Inbar D, Zhang Z, Buxbaum JD, Molholm S, Foxe JJ, Marion RW, Auton A, Greally JM. Mosaic epigenetic dysregulation of ectodermal cells in autism spectrum disorder. PLoS Genet. 2014 May 29;10(5):e1004402.

Heijmans BT, Mill J. Commentary: The seven plagues of epigenetic epidemiology. Int J Epidemiol. 2012 Feb;41(1):74-8.

Mill J, Heijmans BT. From promises to practical strategies in epigenetic epidemiology. Nat Rev Genet. 2013 Aug;14(8):585-94.

Voineagu I, Wang X, Johnston P, Lowe JK, Tian Y, et al. (2011) Transcriptomic analysis of autistic brain reveals convergent molecular pathology. Nature 474: 380­384.