Is there a good recent review about epistasis in human genetics - both rare and common? (@leonidkruglyak @jcbarret @dgmacarthur)
2016-02-10 17:27:00Or, alternatively, about review about epistasis in general (Drosophila, Mouse etc?)
2016-02-10 17:27:45I can see both sides of the argument about the importance of epistasis. Case For: we know it exists (mendelian genetics, model organisms)>>
2016-02-10 17:29:40but we are poorly powered to untangle epistasis in common variation of natural phenotypes, as well as this being expected to be complex.
2016-02-10 17:30:55Case against. Although epistasis does exist for severe variation, in common variation one expects smoothing of interaction terms>>
2016-02-10 17:31:44This is due to both environmental "noise" and evolutionary arguments. In practice we don't need invoke epistasis to have good models
2016-02-10 17:32:33Hence estimates of h2 (narrow sense heritability - no epistasis) is not so different from H2 (broad sense heritability) in practice.
2016-02-10 17:33:18However, I find the variance decomposition models weirdly fragile, and the out-of-sample prediction is way worse than % var explained
2016-02-10 17:33:59So - I'm in the "epistasis is important, we can't find/model it well" camp I think. Love to hear other arguments.
2016-02-10 17:35:22.@ewanbirney Just back from a really interesting meeting on epistais, with @moorejh. Clearly important, but systematically under-estimated
2016-02-10 17:43:59.@ewanbirney Strain of human exceptionalism claims epistatic interactions not impt in human genetics, tho' seen everywhere else
2016-02-10 17:45:08.@ewanbirney Only because ignoring epistasis inflates h2. But then find most of it "missing"! Hmmm....
2016-02-10 17:46:54.@ewanbirney For "complex" disorders, epistasis implicit in liability-threshold model. (Otherwise, how to explain the threshold?)
2016-02-10 17:48:25.@ewanbirney For "complex" disorders, epistasis implicit in liability-threshold model. (Otherwise, how to explain the threshold?)
2016-02-10 17:48:25@felbalazard @BioMickWatson oh. I'd be wary of bringing out a high dimensional trainer (random forests) on case / control data
2016-02-10 18:04:44@felbalazard @BioMickWatson lots of likely data quality / DNA quality / subtle population stratification for the random forest to use
2016-02-10 18:05:05@felbalazard @BioMickWatson in a pure quantitative trait in one cohort ... I'd be happier ...
2016-02-10 18:05:34@felbalazard the right thing is to do it on a quantitative trait ... And we've tried random forests ... I don't trust them in case / control
2016-02-10 18:10:02@felbalazard @BioMickWatson far better randomisation of trait measurement to genotype measurement. Even then, got to be careful ..
2016-02-10 18:11:15.@Abebab Absolutely agree on that point! High heritability *within* groups does NOT imply mean diffs b/w groups are due to genetics
2016-02-10 19:08:26.@Abebab Absolutely agree on that point! High heritability *within* groups does NOT imply mean diffs b/w groups are due to genetics
2016-02-10 19:08:26.@Abebab e.g. body mass index highly heritable w/in populations, yet varies massively (thank you) across populations for non-genetic reasons
2016-02-10 19:09:29.@Abebab e.g. body mass index highly heritable w/in populations, yet varies massively (thank you) across populations for non-genetic reasons
2016-02-10 19:09:29.@ewanbirney No - model just says when have N alleles you're fine, but N+1 (or small number) and you're diseased. Purely genetic threshold
2016-02-10 19:11:27