The Pith: A great deal of important medical genetic differences between people may be due to the nature of interactions of genetic variants.
If you’ve been reading this blog for a while you know that there is a question in genomics right now as to “missing heritability.” The issue is basically that there are traits where patterns of inheritance within the population strongly imply that most of the variation is due to genes, but attempts to ascertain which specific genetic variants are responsible for this variation have failed to yield much. For example, with height you have a trait which is ~80-90 percent heritable in Western populations, which means that the substantial majority of the population wide variation is attributable to genes. But geneticists feel very lucky if they detect a variant which can account for 1 percent of the variance.
One simple explanation, which gains some genomic support, is that variation on these traits is due to innumerable variants widely distributed across the genome. Therefore, a variant of one percent effect may be a rather large one. There are also those who argue that it may be that there are even more very rare, but somewhat larger effect, alleles at work.
Another model is that the “missing heritability” can be solved by reconceptualizing the “genetic architecture” of the trait. This means that currently a major assumption of many models for putatively polygenic traits is that the variation is due to many genes of small effect which modify the trait value in an additive and independent manner. In other words, the genetic architecture in this sense is a linear system. A clear alternative, or complementary, possibility is that there are genetic interactions which are generating deviations from linearity. This would be epistasis, which has different implications depending on the sort of biology you’re talking about (e.g., molecular vs. evolutionary).
A new paper in PNAS makes the case that a lot of the “missing heritability” has to do with the assumption of additivity in many of the models attempting to smoke out associations. But first, let me point to a press release from from GeneWatch:
The study supports earlier findings by GeneWatch UK that much of the heritability of common diseases, calculated using twin studies, may not exist (2). Scientists have been puzzled by the failure of large genetic studies to find genes which explain the “missing heritability” of common diseases such as heart disease and cancer and traits such as height. Typically, 85 to 95 per cent of the expected heritability has not been found. Today’s new study confirms that one explanation may be that interactions between multiple genes would reduce the predicted heritability. These interactions were not properly accounted for by the eugenicist Ronald Fisher who developed the original twin studies method in 1918, and later analysis has not corrected Fisher’s error.
“Claims of a genetic revolution in healthcare have long been based on false assumptions” said Dr Helen Wallace, Director of GeneWatch UK, “If heritability is much lower than expected this means that genetic differences play only a small role in explaining why some individuals get a disease which others do not. Genetic testing can help people with rare disorders but will never be useful to predict and prevent the common diseases that most people get.”
My attention was brought to this by Hellen Wallace herself, who sent a rather bombastic email stating that Eric Lander has come around to her model, where gene-gene interactions loom large. But from reading the paper I think one of the issue that the authors highlight is that there is often a conflation between heritability in the narrow sense, h2, and heritability in the broad sense, H2. h2 accounts for additive genetic variation. The authors seem to be making the case that you may have to focus on heritability in the broad sense. They state: “Broad-sense heritability H2 measures the full contribution of genes… H2 is the relevant quantity for clinical risk assessment, because it measures our ultimate ability to predict phenotype from genotype.”
I have characterized GeneWatch as “Genetic Creationists” before, and that is because of their misrepresentations and exaggerations. A close reading of this paper does not seem to align at all with their agenda, though it does imply that attempts to map genotype to phenotype are going to be hard. Let me jump to the paper’s conclusion:
Finally, notwithstanding our focus here, we believe that concerns about missing heritability should not distract from the fundamental goals of medical genetics. Human genetic studies to discover variants associated with common traits should primarily be regarded as the analog to mutant hunts in model organisms, with the primary purpose being to identify the underlying pathways and processes. The key focus should be to study the biological role of the variants discovered so far. The proportion of phenotypic variance explained by a variant in the human population is a notoriously poor predictor of the importance of the gene for biology or medicine. [A classic example is the gene encodingHMGCoA reductase, which explains only a tiny fraction of the variance in cholesterol levels but is a powerful target for cholesterol-lowering drugs (1).] Ultimately, the most important goal for biomedical research is not explaining heritability—that is, predicting personalized patient risk—but understanding pathways underlying disease and using that knowledge to develop strategies for therapy and prevention.
You can read the full paper at PNAS, it’s open access. But really you have to go through the supplements, and I’ve only read a few sections of that. Do I believe this? I think the model works out. Frankly, I wanted to check the acknowledgments, and the people listed there give me confidence that the theory here is legitimate, even if you don’t work out the equations yourself. But is this empirically the case? That’s a different issue altogether. The authors note that there will be follow up papers soon. What I will be curious about is the extent of differences in interaction effects by trait. A supplementary table gives us a taste. You see correlations for monozygotic and dizygotic twins. In an additive model the third row should be ~ 0. Look at birth weight and voting behavior, and contrast it with height and IQ.
Citaion: The mystery of missing heritability: Genetic interactions create phantom heritability, doi:10.1073/pnas.1119675109
Razib Khan’s degrees are in biochemistry and biology. He has blogged about genetics since 2002, previously worked in software development, is an Unz Foundation Junior Fellow and lives in the western US. He loves habaneros.
January 6th, 2012 at 1:52 pm
Haven’t looked at this in a while, but check out the research on recombinant inbred lines in mice. They suggest lots of GxG, but the biological models may not fit wild outbred populations.
January 6th, 2012 at 2:16 pm
Yeah, that’s the real problem–we can’t inbreed the humans enough for the models. Although I had a genetics professor in college that said that’s what Ivy League schools were for. (kidding….)
But I agree on your take on this so far. But I’m still looking at more too.
I call GeneWatch “deniers” rather than “creationists” though.
January 6th, 2012 at 2:46 pm
I can see how a strict additive model would result in missing heritability for diseases in humans, which can be estimated in very few ways. But what about missing heritability for a trait like height? Estimates show it’s high regardless of the relationship (twin studies, parent-offspring etc.), and body size is ubiquitously high across organisms which can be subjected to more elaborate breeding designs, but there’s still substantial missing heritability. Wouldn’t it generate predictable patterns of over- and under-estimation depending on the familial relationship if epistasis played a major role in missing heritability for human height?
January 6th, 2012 at 3:00 pm
Making RILs with any organism reveals crazy patterns of linkage disequilibrium across chromosomes. The fact that it’s incredibly hard to tease epistasis apart even in flies and worms should scare the shit out of us with ever getting a grip on q-genetics in humans. I think model system work made us too optimistic about the linearity of geno-pheno relationships because there was so much selection bias in what was studied and published — any mutant that didn’t “behave” in map crosses (i.e. Mendelian, scoreable) was chucked, and thus the simple (and less representative) cases dominate the literature. Well, now the easy stuff is done. Good f–ing luck.
January 6th, 2012 at 3:39 pm
i just want to make clear in the comments: the paper is very qualified. i didn’t emphasize that cuz you can all read it. but they aren’t offering the solution, but a possibility.
January 6th, 2012 at 6:24 pm
For height and IQ, epistatic effects and “phantom heritability” cannot be very large, because the studies by Visscher et al. where overall genetic similarity and phenotypic similarity were found to be highly correlated establish the lower bounds of narrow heritability (Zuk et al. explicitly agree with this).
I think non-zero cells in the third column could also be due to assortative mating (which inflates DZ correlations), shared environmental effects, or specific prenatal effects that can lower MZ correlations (for example).
January 6th, 2012 at 6:34 pm
Not an expert here but when I looked rather attentively at the literature at some point, a pattern emerged (unless it was my own bias playing): every time people work with model organisms and look at things that are relatively easy to control and interpret, they find epistasis everywhere and dominant. In contrast, human genomics is dominated by computer scientists and former physicists who are just sure that a simple additive model is the only one that matters and that it explains everything (even when it does not, as is the case with height).
Myself, I’d be very surprised if vast majority of lowly bench scientists who work on metabolic pathways or complex protein interactions would not favor major role of epistasis in the strongest possible manner. It just makes so much sense – other genes are the “environment” with which a particular allele interacts to produce the exact phenotype. Some interactions are strong, some weak. (Typical case: single known mutation in a key protein produces a wide range of disease phenotypes, from deadly to mild).
January 7th, 2012 at 3:17 am
This may seem like a stupid question to all of you well versed in genetics but how do I interpret negative figures in column 3? One theory I have is that positive figures imply interactions with positive feedback and negative figures interactions with negative feedback. Another theory is that environmental effects are implicated – perhaps even in some interactive fashion. I just don’t know and would very much like to understand this.
January 7th, 2012 at 7:36 am
“This means that currently a major assumption of many models for putatively polygenic traits is that the variation is due to many genes of small effect which modify the trait value in an additive and independent manner”
I would say this is stronger than an assumption: Data and Theory Point to Mainly Additive Genetic Variance for Complex Traits
Previous discussion from 08 (including James Crow dialogue, appropriately enough).
January 7th, 2012 at 9:31 am
Epigenetic processes such as methylation also seem to be a good place to look for explaining much of the variance in complex/multigenic traits. Epigenetic events in development (and throughout life) can have complex, non-linear effects on patterning of expression, so small differences in genotype along many genes could potentially explain complex inherited traits and/or complex multigenic diseases.
January 7th, 2012 at 9:47 am
But this is the point — once you inbreed a model organism enough to look at the effects of one or two genes then you find epistasis. Otherwise there is so much variation in a typical human and human development is so plastic that (as a statistician rather than a computer scientist or physicist) the law of large numbers comes into play and the effect of lots of genes of small effect on complex traits gives a nice gaussian distribution, which is indistinguishable from what would be expected with many genes of small effect.
January 7th, 2012 at 11:10 am
gives a nice gaussian distribution, which is indistinguishable from what would be expected with many genes of small effect.
Indistinguishable, maybe (or even yes). But to *explain* a polygenic trait, it matters tremendously whether it is a sum of individual independent genes or a complex non-linear function of gene-gene interactions.
January 7th, 2012 at 11:53 am
Is it poor form if I just say “told you so”?
More seriously, I thought there were independent ways to evaluate the additivity of gene contributions – i.e. to determine whether epistasis plays a role or not. When Richard Plomin came to give a talk in Birmingham about IQ and heritability, I asked if the remarkable heritability of IQ (estimated from the much higher similarity between MZ than DZ twins) might not equivalently be caused by epistasis. I got a quite confident answer that IQ was pretty much entirely additive. I didn’t ask how this was arrived at.
I guess I missed something.
January 7th, 2012 at 12:06 pm
#13, the authors above probably wouldn’t disagree with plomin i bet. i think the chart i post is a hint as to how they’ll partition traits in terms of candidates for epistasis vs. not.
January 8th, 2012 at 6:19 pm
Don’t work in, didn’t study, barely understand gene stat. So from the evol side what interests me is: knowability; time to knowability. To my mind viruses are the most relavent comparison.
Viruses have had eons of time; humans a few decades (of precision, longer vague study).
Virus structure is (partly) gene tools. Humans have hands, which work the (whatever), which works the (other thing), which was made (wherever), by (whoever else), etc.
Viruses have huge numbers. Humans think we have overpopulation.
Virus has nothing better to do, get it right or die. Humans have survived without genomics.
Humans are building a better HD every day. Virus carries its knowledge on its back.
Human intellect. Viruses don’t have the biggest heads.
Human communication. I’m not going to speculate here.
Knowability? In the long run we’ll know everything, I’m sure. But while time stays finite, the stories may be more “This is going to be hard, y’see, the thing is…” than popular science/medical genetics/drug development would like.
January 8th, 2012 at 6:51 pm
(And harder than sci fi would like, how did I forget science fiction? Never boring, you’ll never have to slog through pages/hours of setbacks. Whatever you want it’s doable instantly. Maybe it’s expensive. Really though, I’ll never get purple skin. Sad face.)
January 9th, 2012 at 12:57 pm
Missing heritability with regards to height? I would argue that what you eat is part of ones heritage and can haven an affect on ones height.
January 9th, 2012 at 4:17 pm
Here is Steve Hsu’s take.
“1. The non-additive model analyzed in the paper requires significant shared environment correlations to mask the non-additivity and make it consistent with data that (at face value) support additivity. See Table 7 in the Supplement. This level of environmental effect is, in the cases of height and g, probably excluded by adoption studies, although it may still be allowed for many disease traits.
2. The criticisms in section 11 of Hill, Goddard, and Visscher (2008; also discussed previously here) are, to my mind, rather weak. To quote a string theorist friend: “It is nothing more than the calculus of words”
In particular, I flat out disagree with the following (p.46 of the Supplement):…
http://infoproc.blogspot.co.nz/2012/01/phantom-heritability.html