So the mystery remains. It is tempting to conclude that biological epistasis is widespread but mostly gets mapped to statistical additivity. But that does not explain the deviations from additivity often observed in twins. /x
Replies
Nice blog and good to see this also from the twins/shared environment side. We (with my colleagues in @wittbrodtlab.bsky.social) have tried to tackle the non-additive in experimental settings (in medaka fish) which we can map to human (as the medaka fish are "wild") www.biorxiv.org/content/10.1...
aug 28, 2025, 3:10 pm • 12 5 • view
We see lots of non-additivity - both GxE and GxG and we see third order terms - GxGxE and for loci we can experimental test (via CRISPR) that our statistical inference was right - that indeed we have the causal allele, it does have GxE; we can show genetic background effects as well - GxG
aug 28, 2025, 3:10 pm • 1 1 • view
This is all in the context of wild-derived haplotypes (same "trick" of inbreeding from the wild Trudy Mackay does) and we can show that the alleles we find in our experimental crosses are present in the wild (and so would be part of a GWAS etc if we did this in Medaka)
aug 28, 2025, 3:10 pm • 0 0 • view
We then took all the effect size estimates of addivity, dominance, GxE and GxE to a simulation in a fully outbred setting - as we have in human cohorts - to ask with these genetic parameters what would we expect to discover / estimate
aug 28, 2025, 3:10 pm • 0 0 • view
Similar to the work you cite here (Hill etc) and your own work, we also agree that most non-additive containing loci are adequately discovered using an additive model. Two provisos though - having the causal variant typed *really* helps+you have the measure the relevant environments reasonably well
aug 28, 2025, 3:10 pm • 3 0 • view
Unsurprisingly to actually discover and estimate GxE and dominance / GxG interactions you need either way more samples (10 million or more, not 0.5 million) or far higher allele frequencies.
aug 28, 2025, 3:10 pm • 3 0 • view
As @mikeinouye.bsky.social says, this really shows that the "mainly additive" observation in human is completely consistent with the "pervasive non additivity" seen in model organisms and farm animals - it is allele frequence and model estimation of small proportion of variance things
aug 28, 2025, 3:10 pm • 5 1 • view
However, the fact that additivity is... fine... for discovery, and probably also out of sample population assessment (PRS etc) doesn't mean it is good for everything we want to do.
aug 28, 2025, 3:10 pm • 1 0 • view
Sort of obviously in rare disease diagnosis - where you want v v low false positives and you are conditioning on the alleles in one person, you don't just say "oh it will be fine to use an additive model" - obviously you take into account dominance/recessive if present;
aug 28, 2025, 3:10 pm • 0 0 • view
Furthermore you will take into account GxE, though this is complex because we rarely know these terms - this looks like variable penetrance to us.
aug 28, 2025, 3:10 pm • 1 0 • view
I still maintain that local epistasis (here defined within some cis region) wouldn't be at the genotypic level but at haplotype level. GxG for two her sites is 1. But if those hets represent diff haplotypes then it's model mismatch
aug 28, 2025, 2:05 am • 2 0 • view
It's definitely not my favored explanation for "missing" heritability but it still seems to me like the methods available in humans are ill-suited for estimating epistatic effects compared to what is possible in Drosophila, yeast, or plants. I just find it hard to believe epistasis in humans is ~ 0
aug 28, 2025, 1:42 pm • 3 0 • view
One question I have: do we know if statistical additivity that's actually from biological epistasis behaves the same way (e.g. response to selection) as statistical additivity from biological additivity?
aug 28, 2025, 2:05 pm • 3 0 • view
Yeah I was thinking about this initially and unable to find any examples in the literature where epistasis tagged by additivity versus "pure" additivity mattered. I think the response really is just about narrow-sense h2 (as in the Breeder's Equation).
aug 29, 2025, 12:35 am • 0 0 • view
I learned only recently is that various VCs are defined sequentially as a residual of the lower order terms. So higher order terms only takes the remnants their lower terms, progressively getting obscure as the order grows. This makes me think if variance is a good measure for non-linearity.
aug 28, 2025, 1:32 pm • 3 0 • view
Yeah, in general multi component models will only take the part of the interaction that is not captured by additivity. The twin ADE model partitions them correctly but ONLY if there's no shared environment (and other assumptions hold).
aug 28, 2025, 1:41 pm • 2 0 • view
If we assume gene-level epistasis, like we find all the time in interventional laboratory studies, could this come down to lack of effective methodology for turning locus-level GWAS into gene-level variables?
aug 29, 2025, 1:34 am • 0 0 • view