EpgntxEinstein

The blog of the Center for Epigenomics at the Albert Einstein College of Medicine in the Bronx, New York City. The Center fosters research in epigenomics, with a focus on human disease and computational epigenomics. Blog content written by John Greally.

[epgntxeinstein]

Epigenomics in precision health

A tweet by cousin Hank in the last few days brought this correspondence in Nature Biotechnology to my attention. Entitled Challenges and recommendations for epigenomics in precision health, it represented the views of researchers in NHGRI-funded Centers of Excellence in Genomic Science (CEGS). 

It’s good to see this discussion starting. However, there are a few odd elements to the document that warrant some open discussion, presented below in the hope of creating a dialogue.

1. The unwarranted jump from biomarkers to mechanism.

After talking about the robust biomarker of ageing that is DNA methylation, there’s an unfortunate jump to the following statement: “The silver lining is that the epigenetic profile is not fixed in stone; you may improve your epigenome by changes in diet, exercise, or other modifications.” I am prepared to state confidently that there has never been a study that has proven diet, exercise or anything else changes the “epigenome”, given the problems with confounding effects not addressed in studies to date. DNA methylation correlates with chronological age. However, it does not at all follow that changing DNA methylation alters the ageing process, since we don’t know why DNA methylation changes with age, and we don’t know any interventions that change DNA methylation.

2. Some of the statements of accepted wisdom are unfounded.

In addition to the epigenomic improvement statement in #1 above, the authors present an hypothetical case of monozygotic (MZ) twins discordant for autoimmune disease, and how a blood test could find “a handful of changes in DNA methylation and chromatin accessibility in the blood of the two boys that could be related to the disease” allowing insights into “how epigenomic changes lead to disease" . This is overstated, as only one confounding variable (genotype) is eliminated in this scenario, with many others remaining. In fact, the one study that did look at an autoimmune disease (T1DM) in MZ twins taking into account cell subtype effects found essentially no differences in DNA methylation. Furthermore, this again jumps from what is basically a biomarker to causation, although blood is at least reasonably likely to contain cell types mediating autoimmune diseases.

3. There is an odd focus on chromatin biomarkers.

Maybe not so odd, as some of the authors are the developers of the extremely useful ATAC-seq method. However, it reads strangely to see the justification for epigenetic biomarkers put in terms of DNA methylation (ageing, cigarette smoking) but then all of the discussion being in terms of implementing ATAC-seq. While ATAC-seq is a great assay, the problem is that it may not be quantitative enough to perform as well as the DNA methylation assays cited. The changes in DNA methylation for the epigenetic ageing and cigarette smoking biomarkers are modest in degree, involving a small subset of alleles in the population of cells. ATAC-seq has not been shown to be sufficiently quantitative to detect altered chromatin accessibility of, for example, 10-20% of the alleles in the population of cells tested, the allelic proportion reflected by beta value changes in DNA methylation-based biomarker studies. 

There’s also likely to be a difference in detectability for a small proportion of loci opening chromatin compared with closing chromatin. If the proportion of alleles with closed chromatin is 95%, and this decreases to 80% with a phenotype, that will increase the reads at that locus from none to some, perhaps generating a peak call. If, however, the proportion of alleles with closed chromatin was 5%, and this increases to 20%, the peak already present will probably still be present and not detected to have changed in this inherently non-quantitative assay. This may be why, in the T cell exhaustion example they cite, the overall trend was towards increased chromatin accessibility, possibly reflecting a detection bias. 

4. The idea of a regulatory element catalogue is at odds with the paradigms they cite.

It’s not clear why the authors make a case for an ENCODE/Roadmap-like regulatory element cataloguing exercise, for two reasons. Firstly, these consortia-led projects were of value in the era when these kinds of studies were not easily performed by individual investigators. Secondly, in examples they cite like T cell exhaustion, what is important is not the regulatory landscape of the canonical cell type, it’s the comparison how the landscape changes between normal and abnormal cells, which should really prompt these studies being done by individual labs. 

Bottom line: it’s good to see an informed and thoughtful discussion of how to use “epigenetic” (in this case meaning transcriptional regulatory molecular mechanisms) marks in clinical practice. However, we would be well-advised to adhere to the biomarker utility in the short term, and only make claims about causation or potential for intervention with much more careful and interpretable studies than hitherto performed. I’m also far from convinced that ATAC-seq or other chromatin assays have the quantitative capacity to outperform DNA methylation as biomarkers with clinical utility. And please, no assertions about how we can modify our epigenomes with “diet, exercise or other modifications”. That line of unfounded reasoning leads to epigenetic yoga, epigenetic face cream, and epigenetic anti-dandruff shampoo. 

[epgntxeinstein]

The three definitions of #epigenetics

Basically, there are three definitions of epigenetics plaguing the earth.

The big one is the biochemists’ definition: epi- (above, upon) -genetic (DNA sequence). Pioneered by outstanding scientists, but describes all transcriptional regulatory processes, and a horribly messy departure from the prior definition (below). Easy to remember, but has attracted those seeking to support genetic non-determinism (see epigenetic yoga, epigenetic face cream) and Lamarckianism.

The prior definition is number 2, cellular memory (or persistent homeostasis, as Nanney defined it). This is the one for the purists, who are a bit unsettled by the fact that they are joined by the unruly transgenerational epigenetics enthusiasts.

And then we have the third definition from the epidemiologists, biomarkers with attitude, the association of DNA methylation changes with phenotypes. Some of the biggest studies out there, generally good epidemiologically, but frequently ignore that the DNA methylation they study comes from actual cells, who would like their role to be recognised and are thinking of unionising. 

It is worth noting that none of the above remotely resembles Waddington’s original epigenetic landscape, which attempted to reconcile how genetic mutations could affect epigenesis (cell differentiation).

For perspective, this article by Ute Deichmann is strongly recommended, great summary of the history of the use of the word epigenetics.

[epgntxeinstein]

The EWAS score #EWASscore

If you are a journal editor, or a reviewer of a manuscript or a grant proposal, the following serves as a scoring system (maximum 10 points) for an epigenome-wide association study (EWAS), based on the points made in our recent review.

Starting hypothesis: Is it based on reality (in other words, does it start by saying something like “As everyone knows, DNA methylation is heritable in humans between generations...”)? Are cellular reprogramming and polycreodism (A systematic variability of cell fate decisions to create a distinctive repertoire of cells in a tissue) both considered? [1 point]

Cellular characterisation: Do the authors make the case that the cells have a phenotype, or that they could plausibly mediate organismal phenotype, or report exposure? [1 point]

Study design: Could results be confounded by reverse causation? Is the study powered adequately? Other design issues? [1 point]

Functional genomic assays: Can the assays report mosaic effects? Do they report the cis-regulatory landscape reasonably comprehensively? [1 point]

Verification and validation studies: Are they enough to generate confidence in results of genome-wide assays? [1 point]

Sources of functional genomic variability 1. Are cell subtype proportions measured and accounted for in interpreting results? [1 point]

Sources of functional genomic variability 2. Are functional variants in that cell type and cohort accounted for? [1 point]

Locus-specific targeting: Why have the functional genomic changes occurred in the observed genomic locations? Are they targeted by transcription factors or other mediators? [1 point]

Mechanism of changes: Has a plausible mechanism been proposed? Why are the changes happening in these cells? [1 point]

Wild card: is the study just kind of cool and provocative? [1 point]

For a manuscript, remember that well-intentioned researchers designed a study several years ago, when we didn’t realise how many issues affected the interpretability of an EWAS. The authors are unlikely to be able to address the issues above, but should acknowledge them all and avoid over-interpreting their findings.  However, editors should not send out for review the kinds of manuscripts that score 0-2 above and fail to acknowledge these issues.

When reviewing a grant proposal, it’s no longer acceptable to score less than 8-9. We should not be funding studies that will not be publishable in 2022.

Addendum: George Davey-Smith points out, perfectly reasonably, that the above apply when the EWAS is looking for mechanism. If all you’re looking for is a biomarker, many of the hurdles above disappear. My guess is that well over 90% of EWAS look for mechanism (we just can’t help ourselves) but if you do find the rare, restrained biomarker-only EWAS in the wild, numbers 3 and 5 above are still important, but the others can be waived.

[epgntxeinstein]

How much memory does it take to make it epigenetic?

It is always interesting to challenge people to define what they mean by epigenetics. Like obscenity, people know it when they see it, but have difficulty explaining it.

When you get away from the loose equivalence to all transcriptional regulatory processes, and become more stringent, people start to talk about the properties of cells, specifically the retention of memory of past events, reflected by permanent changes in molecular regulatory processes. 

That’s pretty reasonable, and fits with X chromosome inactivation or genomic imprinting reflecting events very early in development that are remembered throughout life. It’s also what we seek when we look for molecular traces of past exposures or stresses in epigenetic association studies.

So if we are now talking about memory and transcriptional reprogramming, consider the example of alcohol. If you are not a regular drinker and you have a couple of units of alcohol, it’s generally known that it will have much more of an effect on you than if you drink regularly. 

In the images below, you can see the hepatocellular induction of the CYP3A1 protein following rat exposure to kava. A similar pattern of induction of liver metabolic enzymes occurs with ethanol exposure. More cells induce the expression of the enzyme, there is a genuine change in the regulation of transcription of the genes encoding these responses, it’s not due to cells expressing the enzyme within the organ proliferating and replacing the non-expressing cells.

Credit for image: http://focusontoxpath.com/liver-enlargement-hepatic-enzyme-induction-histopathology/

Now, think about what happens when you stop drinking. The enzymes remain induced for some time, but will eventually revert back to their un-induced state.

So is this epigenetic? Is it a memory of past exposures? Or is it a normal cell signalling-mediated response to an exposure? Nobody describes the response to alcohol exposure or other acute exposures as epigenetic, it tends to get lumped into the more generic category of transcriptional regulation.

What if the hepatocytes could be shown to remain induced a year later? Or 6 months, or ten years? Do you need the reprogramming to be passed onto daughter cells after mitosis? What if the cell doesn't divide but remains reprogrammed, is that epigenetic?

And why ask all of these questions? Because we assume there is something special about epigenetic regulation of gene expression that distinguishes it from “normal” transcriptional regulation. We perform studies looking for these special properties in disease. So what separates the special epigenetic property from generic transcriptional regulation? How long does a cellular memory need to last?

[epgntxeinstein]

The bright future for “epigenetics”

If you’re fascinated by epigenetics, as I have been my entire academic career, you may be dismayed to see the recent stream of negativity about our field, to which I have been a prolific contributor. There was our cautionary review in Nature Methods in 2013, a @PLOSblogs perspective in 2015, and a recent review in @PLOSGenetics co-authored with Ewan Birney and George Davey-Smith. In addition, there was the backlash this year to Sid Mukherjee’s piece in the New Yorker, which was featured in Nature News and involved a blog post from Jerry Coyne that aggregated a lot of scientists’ negative responses, including my own. Jerry has another post today which criticises a TED-Ed talk with over 146,000 views that propagates current but inaccurate concepts about epigenetics. Myself and Steve Henikoff are about to release a further review questioning some basic assumptions in epigenetics in Current Biology, getting away from the human phenotypic associations quagmire and raising some simple questions about the basic biology in this field.

When someone like Carl Zimmer writes a piece for the New York Times on epigenetics, the newsworthy aspect to the current discussions is the negative angle -- witness how these scientists are revealing a controversy in what looked like a really attractive field externally. It’s like a family squabble in a soap opera on a trivial level, or a real-time witnessing of a momentous occasion as a field changes direction, on a more profound level. Either way, the story is the negativity.

The responses I have been getting to these recent publications are actually overall very positive. The negative responses focus on impracticality (the epigenetics ship has sailed, too late to do anything about it) (Peripheral blood is the only sample possible at scale from human subjects) (Testing more than just DNA methylation will make human studies too expensive) and some degree of dismissiveness (We can fix cell subtype effects analytically, stop banging on about it). 

What’s being missed is the positive component in each of these publications --  we describe how there is massive potential for the field of epigenetics, now that we recognise our problems with study interpretability. For example, a problem is DNA sequence variability influencing DNA methylation outcomes. If we identify the DNA sequence polymorphisms associated with the phenotype and confounding our DNA methylation interpretation, that’s not just a GWAS-like positive outcome, it comes with a possible genomic regulatory mechanism for free. A cell subtype effect confounding interpretation of the DNA methylation result is another insight into disease pathogenesis that would be embraced enthusiastically by a cell biologist. If we start to regard DNA methylation as a multi-purpose tool that reads out a number of properties of cell populations and the individuals studied, and cling less tightly to the cellular reprogramming hypothesis that is usually being studied, we open up our potential for discovery to a much greater extent.

I have also become much more Ptashnean in my viewpoint, not because I adhere unquestioningly to Mark Ptashne’s transcription-factor-or-nothing worldview, but because we need to rebalance the field to incorporate the undoubtedly important role of TFs in cell and genomic physiology. If you look at the illustration used for Carl Zimmer’s New York Times piece, typical of every introductory slide in a lecture about epigenetics, the obvious omission is the transcription factor -- how do those pretty coloured blobs on histones and DNA know where to go in the genome without them? 

If we find what look like genuine cellular reprogramming events, not attributable to confounding influences, and we incorporate the idea that these are due to TFs changing their regulatory functions, we can start to use motif analyses and other studies to ascend the regulatory hierarchy and figure out what might have mediated these DNA methylation or other transcriptional regulatory changes, an extraordinary mechanistic insight. Without considering the role of TFs we’re attributing innate intelligence to DNA methylation or chromatin state modifiers, which almost certainly do not know where they need to go in the genome, or how this pattern should change as a response to a perturbation. 

We have so much potential for new discovery, for real insights into the mechanisms underlying phenotypic changes. Yes, we need to work harder to get to these insights, and expand our skill set beyond genomic assays or analyses, but these kinds of challenges are what we signed up for when we chose to be scientists.

I’m not negative about the future of epigenetics at all. I can’t wait for us to perform these new study designs, it really feels like the start of a extraordinarily exciting new era. Yes, the epigenetics ship has sailed, but we still have time to steer it away from the rocks that we see straight ahead and out to sea on a proper voyage of discovery.

[epgntxeinstein]

Epi+genetics

A few historical notes: Conrad (Hal) Waddington did not invent the word epigenetics, it was around since Aristotle over 2 millenia previously.

Waddington defined the epigenetic landscape. Contrary to common belief, he did not make up the word epigenetics to mean “epi-” (above, upon) “-genetics” (DNA sequence).

This epi+genetics definition (thanks Steve Henikoff for the + idea) is meaningless, because any transcriptional regulatory event that differs in two cells in the body with the same DNA is therefore “epigenetic.”

I tweeted recently a test of the perception of the meaning of the term “epigenetics”:

https://twitter.com/epgntxeinstein/status/730022770491428864

I made two Venn diagrams and asked whether epigenetics was better defined by (a) or (b), with (a) winning 34 to 26 as of today.

This was a bit dismaying, for a couple of reasons. It means that we have indeed strayed from the heritable regulatory information component of the definition shown for (b). 

However, this was kind of a trick question. In (a), “epigenetic” overlaps basically all of the category “transcriptional regulation” in the Venn diagram.

So we have established that the epi+genetics definition is being equated with all transcriptional regulation and is the more popular definition amongst my scientific peers represented in this Twitter survey (not a very scientifically rigourous survey, I’ll agree).

This adoption of epi+genetics as the common definition is unhelpful for our field, but worth knowing about. Look out for a few review papers that argue why this definition is wrong, unhelpful and damaging. Back to work...

[epgntxeinstein]

Epigenetics and the Gartner hype cycle

If the study of epigenetic dysregulation in human diseases was a technology startup company, we’d have a problem. 

Our angel investor confidence would have be shaken by the blowback about the results of studies to date being over-interpreted, or over-hyped. 

This is a well-known cycle in tech, dubbed the Hype Cycle by the Gartner consulting group.

The overhyped product fails to meet overblown expectations, leading to a reversal of initial optimism and a critical inflection point -- is the product sustainable, or is it destined to fail?

Realisation that substantial confounding influences exist that cause over-interpretation of DNA methylation variability has tipped us over the Peak of inflated expectations towards the Trough of disillusionment. 

Fortunately, in epigenetics there is substantial R&D at the basic science level that is going to get us through the Trough of disillusionment and is currently bringing us up the Slope of enlightenment. The Integrative genomics approach (previously described over at PLOS Blogs) means that we have to design and execute our studies more rigourously, but with the potential to harvest disease associations not limited to epigenetic changes but also DNA sequence, cell subtype and transcriptional associations.

Trudging up the Slope of enlightenment is by its very nature a hard slog. But it’s not only an absolute requirement, it’s hugely exciting and promises to yield major insights into human disease pathophysiology when we get to the Plateau of productivity. So let’s take heart as a field and embrace this challenge, we know the direction we need to take.

[epgntxeinstein]

Our new #DataViz #D3js software #Pcaso http://pcaso.io/mundo

Explore your own data in multiple dimensions, upload to http://pcaso.io/

@genomenathan @robaboukhalil @nygenome @einsteinmed

[epgntxeinstein]

Over-interpreted epigenetics study of the week (2)

UPDATE: take a look at the response of the lead author at his site, kudos to him for engaging in this constructive discussion.

Epigenetics studies of interesting questions sometimes get great traction, whether it has to do with transmitting your memory of surviving the Holocaust or your sexual orientation. 

At the current ASHG meeting an abstract was presented (see the end of this posting) in which it was claimed that testing epigenetic markers can allow you to predict sexual orientation. 

There are many reasons why this study is uninterpretable, some of which are described in general terms here. What is not described in the abstract is that the cells used were from saliva samples, which includes a variable mixture of buccal epithelium with a majority of leukocytes. The presence of microbial DNA also has the potential to cross-hybridise to human probes on a microarray, so the screening approach used could be criticised for several potential technical flaws. Also not described is the marginal, uncorrected significance for this underpowered study. These only came to light when the presentation happened.

However, there were some warning flags in the abstract. Why use a new algorithm to identify these predictive markers, did current approaches not yield any results? Where is the mention of performing some sort of locus-specific, orthogonal, more quantitative assay to verify and validate the DNA methylation changes predicted by the microarray studies? 

Then they fall into the rabbit hole. Up to the mention of “...9 regions” they had adhered to a description of biomarker discovery, but like everyone else in the history of epigenetics studies they could not resist trying to interpret the findings mechanistically. So they go there: they talk about the genes implicated. Now you need to invoke all of the issues to do with mechanism, including understanding why these loci are of relevance to the phenotype in the cells of saliva. Ewan Birney also made the point in a Tweet yesterday that cross-sectional studies like this are subject to the influence of reverse causation, so study design issues also need to be taken into account.

Some poor young lad gets up on stage at #ASHG15 having worked hard to generate this story and is now being eviscerated by people like me. It’s not personal about him or his colleagues, but we can no longer allow poor epigenetics studies to be given credibility if this field is to survive. By ‘poor,’ I mean uninterpretable. We should only present biomarker studies when they are shown to perform robustly as biomarkers. We should only present mechanistic studies when we have excluded the many biological and technical sources of variability that can mislead us. If we have an intriguing preliminary observation, we present it as such and do not claim that we have generated “...strong support to the hypothesis that epigenetics is involved in sexual orientation.”

Who is culpable in the current situation? The first news report publicising this study came from @NatureNews, but more concerning is the abstract review process that permitted this study to get accepted for presentation, compounded by the press release issued by the American Society for Human Genetics. Both organisations should know better -- they need to be substantially more rigourous and not blindly accept that the numbers generated from DNA methylation studies are inherently meaningful. 

As a field, we need something like a consensus checklist to guide scientists and  reviewers in epigenetics studies. It would be unusually prescriptive an approach, but appears necessary to counterbalance the current over-interpretation of epigenomics studies of human phenotypes. The historical lessons of GWAS self-correcting to account for population effects and to move towards adequately powered studies is what we need to learn. The epigenome-wide association study is at a critical juncture, it is time for epigenetics researchers to be much more self-critical and rigourous.

A novel predictive model of sexual orientation using epigenetic markers.

Authors: T. C. Ngun [1]; W. Guo [2]; N. M. Ghahramani [3]; K. Purkayastha [1]; D. Conn [4]; F. J. Sanchez [5]; S. Bocklandt [1]; M. Zhang [2,6]; C. M. Ramirez [4]; M. Pellegrini [7]; E. Vilain [1]

1) Department of Human Genetics, David Geffen School of Medicine at University of California Los Angeles (UCLA), Los Angeles, CA, USA; 2) Bioinformatics Division and Center for Synthetic & Systems Biology, TNLIST, Tsinghua University, Beijing 100084, China; 3) Department of Pathology and Laboratory Medicine, David Geffen School of Medicine at UCLA, Los Angeles, CA, USA; 4) Fielding School of Public Health, UCLA, Los Angeles, CA, USA; 5) Department of Counseling Psychology, The University of Wisconsin-Madison, WI, USA; 6) Department of Molecular and Cell Biology, Center for Systems Biology, The University of Texas at Dallas, Richardson, TX 75080, USA; 7) Department of Molecular, Cellular, and Developmental Biology, UCLA, Los Angeles, CA, USA.

Sexual orientation is one of the most pronounced sex differences in the animal kingdom. Although upwards of 95% of the general population is heterosexual, a small but significant proportion of individuals (3-5%) is homosexual. Male sexual orientation has been linked to several genomic loci, with Xq28 and 8p12 being the most replicated. As with other complex traits, environmental factors may also play an important role. Firstly, monozygotic twins show substantial levels of discordance for this trait. Secondly, each male pregnancy a woman has increases the chance that her next son will be homosexual by 33% (the fraternal birth order effect). Thirdly, early life androgen exposure in women is associated with increased rates of non-heterosexual identity. Taken together, the evidence suggests a role for non-genetic and, possibly, epigenetic influences on sexual orientation. Our aim in this study was to create a predictive model for sexual orientation using epigenetic markers. We created our model based on genome-wide DNA methylation patterns in 37 monozygotic male twin pairs that were discordant for sexual orientation. 10 monozygotic twin pairs concordant for homosexuality were included as a control population. Genomic sites where methylation occurred were consolidated into short regions based on proximity and correlation of their methylation patterns to increase the signal to noise ratio. We then applied the FuzzyForest algorithm to our dataset. Briefly, regions were clustered into modules using Weighted Gene Coexpression Network Analysis and recursive feature elimination was performed with the random forest algorithm (RF) to identify regions most relevant to sexual orientation. The highest prediction accuracy was achieved using information from just 9 regions. Some of these regions were associated with the regulatory domains of two genes, CIITA and KIF1A. The former is a transcriptional regulator that is sometimes referred to as the master control factor of class II major histocompatibility complex genes. The latter is a neuron-specific transport protein that is important for movement of synaptic vesicle precursors along axons. Our results demonstrate that studies of the epigenome can yield new insights into the biological underpinnings of sexual orientation and provide strong support to the hypothesis that epigenetics is involved in sexual orientation. To our knowledge, this is the first example of a biomarker-based predictive model for sexual orientation.

[epgntxeinstein]

Over-interpreted epigenetics study of the week

This weekend Twitter is buzzing about an article in The Guardian:

Study of Holocaust survivors finds trauma passed on to children's genes http://gu.com/p/4byz9/stw

The gist of the Guardian article was that “life experience can affect the genes of subsequent generations,” with a tone enthusiastically selling trans-generational epigenetic inheritance, helped by quotations from the study authors (or, more likely, the public relations department of their institution), including “The gene changes in the children could only be attributed to Holocaust exposure in the parents,” and “To our knowledge, this provides the first demonstration of transmission of pre-conception stress effects resulting in epigenetic changes in both the exposed parents and their offspring in humans.”

So what were these conclusions based on? The cohort was 32 Holocaust survivors and 8 Jewish control individuals, and 22 offspring of the Holocaust survivors and 9 Jewish controls, matched for age, sex, a pertinent FKBP5 genotype and other phenotypic parameters, the design was cross-sectional, the biological sample used was peripheral blood leukocytes, the molecular assay was bisulphite pyrosequencing of 6 CG dinucleotides at an intronic enhancer of the FKBP5 gene, combining the observations in three ‘bins’ of 2, 3 and 1 CG and averaging the DNA methylation values observed in the bins with ≥1 CG before making comparisons. The analysis involved modelling any potential contributors that suggested themselves to the authors, including the FKBP5 rs1360780 genotype, and a number of psychological measures, looking for DNA methylation changes that could be attributed to the direct or parental Holocaust exposure alone.

I really don’t like criticising colleagues, a tendency I’m guessing we all share in science. Giving negative feedback on what is clearly in this case an over-interpreted study should always cause us to think sympathetically about the good intentions of the authors. However, if a study is over-interpreted, and is gaining media traction, and influencing other scientists to pursue similarly poor studies, there needs to be some sort of critical response. So, with apologies to the study authors, I offer the following feedback.

The starting hypothesis is not explicitly stated, but can be inferred to be as follows -- people who have survived the trauma of the Holocaust are distinctive for having cells in their bodies that change DNA methylation, and their offspring, not having been exposed to the Holocaust, also have changes in DNA methylation in their cells, making them distinctive compared with other people who are not offspring of Holocaust survivors. What is assumed based on the shared change in DNA methylation is that the parent changed DNA methylation, that this occurred in somatic and germ line cells, and that the offspring inherited the DNA methylation change, a trans-generational inheritance of this epigenetic regulator.

So what exactly did they find in this study? Fundamentally, they found that while the first two bins of DNA methylation values measured for >1 CG showed no difference between groups, the DNA methylation values of the single CG in the third bin were different in the Holocaust survivors themselves compared with controls, and in their offspring compared with controls.

So does this prove the starting hypothesis? The answer is no, the study is pretty typical of all epigenetics studies today for being uninterpretable, for the following reasons:

a. Confounding molecular processes are not studied. I describe these molecular processes in a recent PLOS Genetics Biologue, listing transcription through the locus being tested and the cis effects of DNA sequence variability between individuals upon DNA methylation as major influences that need to be tested. The control group was matched for ethnicity (’Jewish’) but there is clearly enough genetic diversity within an ethnic group to drive DNA methylation differences through methylation quantitative trait loci (mQTLs). All of the effects observed could be secondary to mQTLs or, at this intronic locus, polymorphic transcription of the FKBP5 gene between the individuals tested. Without knowing the effects of these influences, you can’t assume the changes of DNA methylation are independent of these influences in a way that addresses the starting hypothesis.

b. The influence of cell subtype composition is unstudied. Since Andy Houseman demonstrated the effects of leukocyte subtypes on DNA methylation in 2012, the epigenetics field has started to pay more attention to this issue. The peripheral blood tested in the Holocaust study is undoubtedly variable in terms of cell subtypes between individuals; the question is whether the locus tested is differentially DNA methylated between leukocyte subtypes and that there is enough of a systematic shift in proportion of one of these differentially DNA methylated cell subtypes into the test or control group to influence the DNA methylation patterns observed.

In fact, a critical look at the results strongly suggest this to be the major influence occurring in the Holocaust study. The DNA methylation values at the discriminatory CG dinucleotide are between 50-60%. Such intermediate DNA methylation can only be due to 50-60% of the alleles (probably cells) in the population studied being methylated at that locus, and 40-50% being unmethylated. The presence of intermediate DNA methylation indicates that the testing has been performed on what we call a meta-epigenome, or a collection of epigenomes, with mosaicism for the individual specific cell types and their characteristic DNA methylation patterns. To shift from 50-60% DNA methylation has to be due to an altered proportion of cells within the population having DNA methylation at that locus. While this could indeed reflect a specific cell type changing its epigenetic pattern, it has first to be excluded that the change is not merely reflective of cell subtype proportional changes, and that no cell type within the population has, in fact, changed its DNA methylation pattern at all.

Knowing what cell subtype has changed in proportion to cause the DNA methylation difference would be quite interesting in terms of the question being asked by the authors of the study, but would not address the starting hypothesis.

Without wishing to beat this study to death, the fact that the DNA methylation difference that characterises the exposed parents is an increase while that characterising the offspring is a decrease is difficult to reconcile with any sort of heritability mechanism. 

The story being told by the Holocaust study is indeed fascinating as a scientific possibility, and will no doubt prompt others to pursue similar studies. Unfortunately, the story is typical of many in the field of epigenetics, with conclusions drawn based on uninterpretable studies. While there have been high-profile reviews that emphasise the weakness of the evidence for trans-generational epigenetic heritability, especially in humans, these have not been enough to steer people towards the rigourous studies needed to test this attractive but poorly-founded idea. Every week there are uninterpretable epigenetics studies published, the Holocaust study is merely one of many, these authors are merely following prevailing beliefs in over-interpreting their data. However, every such study damages the ‘brand’ of epigenetics a little more. If we want human disease epigenetics to be sustainable as a field of research, we have got to start to do substantially better in designing, executing, interpreting, reviewing and funding these studies. 

[epgntxeinstein]

An open letter to my Pathology colleagues

We need your help explaining something. Why are you still using a 121 year old technique?

Blum F. 1894 Notiz über die Anwendung des Formaldehyds (Formol) als Hartungs- und Konservierungsmittel. Anatomischer Anzeiger 9, 224

The technique of formalin fixation and embedding in paraffin has served you well for over a century, but it’s time to move on. The problem is that the histology services offered by the pathologists are no longer the end point for the diagnostic processes but part of a supply chain that extends downstream to genomics, and formalin-fixed, paraffin-embedded (FFPE) samples are the enemy of genomics, destroying RNA and profoundly damaging DNA.

When I talk to my pathology colleagues, they tell me that they’re not oblivious to the problem, and that there are new fixation and preservation techniques being developed. And why am I complaining, aren't there kits that allow us to work with FFPE-derived DNA?

Well yes, but that’s like saying that the Gowanus Canal here in NYC is less polluted than it used to be and shouldn’t we consider catching fish from it for our dinner. Why fish from the Gowanus Canal in the first place? Why use FFPE these days at all? 

My father is a pathologist, and I put this question to him. He asked what we’d use in the lab to preserve nucleic acids in tissues. I suggested RNAlater from Qiagen, so he told me to give him a tube of it. He used it to fix a discard sample of tonsil, performed histology and immunohistochemistry for the CD3 surface marker, and got results identical to FFPE (picture below). When pressed, he said the one problem was that the samples were not quite as rigid as FFPE, and needed more careful cutting for sections. 

It looks as if there is room for re-thinking the use of FFPE processing of samples. I can’t think of another technique from the 1800s still in mainstream use today, perhaps it’s time to retire what has been an extraordinarily useful approach so that the pathology department does not compromise the supply chain of modern molecular diagnostics. 

As a genomics-focused researcher, I would appreciate the insights from my pathology colleagues. Can this be accomplished?

[epgntxeinstein]

Fabricated DNA methylation data

The British Journal of Psychiatry has, laudably, retracted a paper on the request of co-authors who found one of the group to have fabricated DNA methylation data.

There's a lot of scandal associated with this story, as documented by @ivanoransky at Retraction Watch. The text of the retraction in BJPsych is paywalled, but I'm going to quote it here as this is in the public interest:

"An investigation carried out at the request of the Dean of the Faculty of Medicine of the University of Geneva has concluded that one of the authors (Alain Malafosse) fabricated methylation data. A reanalysis of the DNA reveals no significant correlation between childhood trauma and methylation of the NR3C1 gene. The original conclusions therefore no longer hold true and we wish to retract the paper."

So here's the question -- what was the "reanalysis of the DNA"? Going through the original paper, these were the essential components of the story:

They studied "Ninety-nine adult participants receiving out-patient treatment for bipolar disorder type 1 or 2", scoring childhood trauma using the "Childhood Trauma Questionnaire (CTQ), a self-report questionnaire that examines five types of trauma (sexual abuse, physical abuse, physical neglect, emotional abuse and emotional neglect), was used to assess childhood traumatic experience"

Samples of DNA were extracted from peripheral blood leucocytes

They performed bisulphite pyrosequencing on "a portion of the exon 1FNR3C1promoter...containing eight CpG sites"

"Samples were processed and analysed masked to demographic variables, psychiatric diagnosis and childhood maltreatment history."

The statistical analysis: "Linear or logistic regression with adjustment for age and gender and age at onset where applicable (number of mood episodes, for instance) was used to assess the effect of childhood maltreatment on continuous and categorical clinical variables. The sum score of abuse and neglect was calculated, as described in our earlier study, by summing the different maltreatments, each treated as a binary outcome (yes or no).17 Linear regression with adjustment for age and gender was used to assess the effect of childhood maltreatment on NR3C1 methylation status. Mean methylation percentage of the eight NR3C1 CpG sites was used as the dependent variable. For the association with NR3C1 methylation status, severity of each type of trauma as described by Bernstein & Fink, based on four levels of maltreatment (none, low, moderate and severe), was also used. In a secondary analysis all CpGs were analysed separately. The results of regression models are presented as standardised regression coefficients (β) with 95% confidence intervals which can be interpreted as effect size. The threshold for significance was P = 0.05."

The differences in DNA methylation observed associated with differing degrees of emotional abuse during childhood were of the order of 5-10%, reaching statistical significance. 

This kind of study, based on the use of peripheral blood leukocytes and ignoring the effects of DNA sequence variability, is already loaded with potential confounding influences. In the hands of the most honest and best-intentioned researchers, misleading results are being published monthly. So in an open request to the co-authors of this paper, could they please divulge what was fabricated in the retracted study, so that as a community of researchers we are alerted to the forensic considerations that we need to think about when reviewing these manuscripts? I realise that the co-authors probably want nothing more to do with this particular publication, but it would be doing the broader field a great service were they to address this issue.

[epgntxeinstein]

Don't blame the epigeneticists

In a Nature Comment piece a few weeks ago, an interesting perspective was raised about recent discussions of heritable, non-genetic risk for disease, making the point that this often involves disproportionate finger-pointing at mothers and their behaviours causing risk for offspring.

The lede is put as follows "Careless discussion of epigenetic research on how early life affects health across generations could harm women." The examples of headlines provided by the authors do indeed emphasise maternal behaviours, and assert that "Factors such as the paternal contribution, family life and social environment receive less attention."

However, by pointing the finger explicitly at epigenetics research, the authors miss the mark. Yes, epigenetics research is the stimulus that has prompted many recent press releases, and the misguided discussions that the authors rightly decry. However, the studies that indicate heritable, non-genetic risk of disease are epidemiological, not epigenetic. Once the associations have been identified by epidemiologists, we epigeneticists go in trying to find out whether DNA methylation or another epigenetic regulator is also associated, bringing us a step closer to mechanism. However, we do not make the original association, that's from our epidemiology colleagues.

So when the foetal alcohol syndrome example (FAS) is raised as an example of over-reaction to early epidemiological associations, the authors are doing exactly what they warn against in their legitimate point about using overly broad brush strokes -- epigenetics research cannot be blamed for FAS, crack baby or refrigerator mother hysteria, but the lede and the juxtaposition of these discussions lead to the blame being focused on epigenetics research. The point is then emphasised in a Nature podcast just to rub it in further.

If you have had the experience of having to give a talk about your work on something associated with maternal risk, and you are looking at an audience of scientists and clinicians, with at least half the room women, you learn to tread sensitively. As a paediatrician and a clinical geneticist, I'm well aware that a parent of a child with a problem doesn't need to be loaded with any more guilt than they will generate on their own, usually without any foundation in fact. 

There are, however, genuine biological differences between the influence that can be exerted by father and mother on the developing foetus, with, for example, ongoing exposures during pregnancy for mothers having more potential effects on the developing foetus than exposures to the father. We should be careful not to take measures that stifle debate or research of genuine differences in exposures between fathers and mothers.

It could easily be argued that epigenetics research, instead of increasing the emphasis on the burden of responsibility of mothers, has in fact opened up the possibility that the extraordinarily limited contribution of the paternal gamete can confer non-genetic risk. The sperm contributes no mitochondrial DNA, and even lacks canonical histones in over 90% of the genome, making it difficult in the past to see how it could pass on any information that wasn't purely genetic. However, we now have examples of paternal diet having effects on offspring associated with DNA methylation differences, so epigenetics research has, in fact, opened up ways to blame Dad. 

The points made by the authors of the Nature Comment are very valuable and necessary. There needs to be more care in how press releases and popular communications are crafted, especially if the societal, legal or other implications raised by the authors begin to emerge as real issues. I would, however, ask the authors to reconsider where they are placing the blame for this miscommunication. Epigeneticists are not the problem, and our research has more potential for good than harm in clarifying risk and allowing fact-based guidance for both parents.

[epgntxeinstein]

Jerzy Jurka 1950-2014

I was sorry to see the news today that Jerzy Jurka has passed away.

For everyone who has ever used RepeatMasker, or studied transposable elements in any way, Jerzy impacted your career. He was a pioneer in identifying and cataloguing repetitive elements in genomes long before we had full genome assemblies. He was interested in these genomic elements when they were the first thing discarded from analysis by other researchers. 

I interacted with him a number of years ago when I was trying to understand why SINEs are depleted in imprinted regions. He was very self-effacing but profoundly knowledgeable about transposable element biology, it was a very rewarding experience for me.

His legacy is embedded in RepeatMasker. The Alu SINEs in the human genome are subdivided into AluJ, AluS and AluY families. When I asked him about the significance of the names, he said that J was for Jurka, S for his colleague Dr. Smith, and Y was for Young Alu SINEs. 

So next time you see AluJ elements pop up in your analysis, tip your hat to Jerzy Jurka, and be inspired by his dedication to an area of science that defined "junk" DNA, but ultimately proved to be a rich source of information. We need these brave pioneers.

[epgntxeinstein]

Reviewable Research: software publications

We really need to change how we publish in the field of genomics research. Everyone knows that, but it's one of those problems that is simultaneously daunting and nobody's clear responsibility. 

Enter John Quackenbush, who has taken on the responsibility to lift the journal Genomics from the citation index doldrums. I am lending a hand with a new section on Genomics Analysis and Tools. 

So here's the radical idea. Make software publications in the field of genomics meet the standards of rigour of those in Computer Science. We enlisted colleague Aaron Golden to help craft appropriate guidelines, shown below.

So why would an author want to go to these extra steps of stringency when publishing a new algorithm? Because we're trying to differentiate high-quality, serious software that would merit exploration by colleagues and funding support. If the reader knows the software has been tested by reviewers, is available through a Reproducible Research repository, and has pre-defined computing requirements, they are much more likely to consider testing or adopting the software.

Feedback about the guidelines below will be welcomed. 

Guidelines for Software Publications in Genomics

Format

Articles should include Abstract, Introduction, Results, Discussion, Materials and Methods, Acknowledgments, and References.

Length

Maximum of 8 illustrations (Figures & Tables).  No limit on references.  Manuscripts should be no more than 8 published pages or 56,000 characters (excluding references).  One published page is ~7,000 characters, each table/figure is ~2,450 characters.

Software publication guidelines

Definition of genomics ‘software’

New and original software algorithms or systems not previously published elsewhere, pertinent to genomics research.

Novel and innovative refactoring or engineering of existing, established software algorithms or systems that significantly enhance genomics research.

Software publication requirements

Genomics application placed in context with respect to required software solution.

Algorithm/workflow clearly delineated, time/memory complexity assessed.

Justification of software infrastructure used.

Evidence of performance and stress-testing with associated metrics.

Results relevant to genomics context.

Fostering reproducible research

The software version for the publication must always be available for download & implementation, as well as any relevant data used in the publication for a given time period.

Authors must ensure automated scripts or fully documented instructions available to reproduce results reported.

Authors encouraged to release newer versions of software BUT must clarify how the updated software will impact the results reported in the original journal article at the repository hosting the software.

Authors are encouraged to use Reproducible Research repositories such as bitbucket or github to host software, or indeed to use novel supporting frameworks such as ActivePapers.

Reviewer guidelines

All software must be tested as part of the reviewing process.

Design of a specific checklist to be used to facilitate this process.

Metrics from this to be reported on publication (anonymized).

All Genomics articles that are rejected will be given the option of transfer either to Genomics Data or Gene.

[obsidianobelisk-deactivated2012]

[scientificillustration]

C. H. Waddington’s developmental landscape (top) and its underpinnings (bottom)

"The ‘developmental landscape’ is a representation of development as a system whose parameters are genetic loci and whose state space is a set of phenotypic states. The state space is depicted as a surface, each point of which represents a phenotype. The genetic parameters are depicted as pegs that pull on the surface and thus determine its contours. Epistatic interactions between genetic loci are represented by links between the strings by which those loci pull on the surface. The development of the organism is represented by the trajectory over the surface of a ball which passes through a series of phenotypic states as it rolls ‘downhill’ from conception to death. Waddington intended this diagram to make vivid the idea that the effect of a change at one genetic locus depends upon the states of all the other genetic loci, since it is all the loci together which determine the shape of the landscape. Some genetic changes, such as those which affect the tops of inaccessible ‘hills,’ will have no effect on development. Other changes of the same intrinsic genomic magnitude which affect the entrance of a valley or ‘canal’ will have a massive effect on development. The phenotypic impact of a genetic change is not proportional to the magnitude of the genomic change, but depends on the structure of the developmental system. Furthermore, the phenotypic difference produced by a genetic difference is not explained by that genetic difference in itself, but by how that change interacts with the rest of the developmental system. This picture retains considerably validity in the light of contemporary developmental genetics.

If we suppose that some of the ‘pegs’ in Waddington’s model are environmental factors, rather than genetic loci, then we can define separate notions of ‘environmental canalisation’ and ‘genetic canalisation’. A phenotypic outcome is environmentally canalised if those features of the surface which direct development to that endpoint are relatively insensitive to the manipulation of the environmental parameters. A phenotypic outcome is genetically canalised if those features of the surface which direct development to that endpoint are relatively insensitive to the manipulation of the genetic parameters. Ariew proposes to identity innateness with environmental canalization. Innateness-as-canalization is a matter of degree. A trait is more innate the more environmental parameters its development is buffered against and the wider the range of variation in those parameters against which it is buffered.”

From: Waddington, C. H. (1957). The Strategy of the Genes: A Discussion of Some Aspects of Theoretical Biology, London: Ruskin House/George Allen and Unwin Ltd.

[epgntxeinstein]

The study that will never be performed

It has been a frustrating weekend, but not for scientific reasons. Maybe that's the motivation behind this jaundiced posting. 

The epigenome-wide association study (EWAS) is a tricky beast, much more so than the genome-wide equivalent looking for sequence variants. The guidelines for such studies are being updated, which should get us to a new level of stringency and likelihood of positive outcomes.

A key issue is that we pay attention to the cell type we study. What should probably be called The Houseman Effect is now widely appreciated, that shifts of cell subpopulations within a sample can influence epigenetic outcomes, without the cells themselves necessarily having altered their epigenomes. The obvious lesson is that we should be using purified cells in EWAS, that's easy enough to appreciate.

The problem is that purified cells, especially from human subjects, can be really tricky to acquire. So when Andy Houseman and others applied statistical tricks to account for cell subpopulation proportions, this seemed like a nice strategy to rescue these difficult situations.

But here's the problem -- you might suspect that this kind of adjustment is an approximation, and not necessarily as good as a cell purification approach. How do you test how well the approximation works? By doing the study in parallel, with purified cell subtype(s) and the entire population. Who is going to have the incentive and funds to do this? Nobody. The people invested in being able to test mixed populations have a vested interest in not wanting their assumptions to be challenged, realistically speaking. Plus this kind of work is really expensive.

So this is a study that will never be performed, but is essential if we are to define really solid EWAS guidelines. It is discouraging to think of the amount of time and money that will be spent based on the possibly erroneous assumption that statistical modeling can really account for cell subtype effects without introducing noise. We will then be forced to review and read the resulting papers, and will not know whether to believe the results. Yes, this week is starting on a grouchy note, but this has to represent extenuating circumstances.