Thread by @cremieuxrecueil

Amazing!

The missing heritability issue between SNP heritability methods and traditional pedigree-based estimates has now shrunken to just 12%.

Thanks to large-scale whole-genome data and simultaneously estimated phenotypes, there's not much missing heritability left!

This analysis has several advantages compared to earlier ones.

The most obvious is the whole-genome data combined with a large sample size. All earlier whole-genome heritability estimates have been made using smaller samples, and thus had far greater uncertainty.

The next big thing is that the SNP and pedigree heritability estimates came from the same sample.

This can matter a lot.

If one sample has a heritability of 0.5 for a trait and another has a heritability of 0.4, it'd be a mistake to chalk the difference up to the method.

For a lot of these differences, despite high precision, the difference is just not significant.

You see significant differences for a surprising bundle of traits, colored in red. These include things like telomere length and height.

Why? Who knows.

The authors mapped out the next steps:

Looking into getting the last 8% of the genome into the calculations, including ultra-rare variants, and incorporating structural variants.

I have no doubt doing this will continue to shrink the missing heritability.

Onward and upward!

The point about ultra-rare variants is already something the authors looked into, but they excluded it from the main analyses to keep their results precise and theoretically understandable.

Laudable!

Nonetheless, preliminarily adding those leads to some noteworthy changes:

For the 17 traits that are not subject to an obviously incorrect model that produces negative estimates with the inclusion of ultra-rare variants, the missing heritability is curbed by 1.5pp on average, or 20.1% of the residual gap.

This will evidently be fruitful to look into!

The authors noted that they included marginally-significant rare variant effects. That imprecision could bias estimates up or down. Larger N's needed!

They also discussed including sex chromosomes, which, under these assumptions (see pic) would explain ~8% of the residual gap.

They also noted that including that last 8% of the genome might close the residual gap by 0.047 * 0.33 (WGS h^2) = roughly 1.6pp, or about 13% of it.

Better calling and annotation might boost everything else in the process too, so definitely worth pursuing!

Finally, maybe for some of the liability-scaled traits, the gap *should* grow, because the pedigree estimates are plausibly biased downwards.

Figuring out how true this is will take getting a more representative sample or being clever about phenotyping.

This is great stuff.

It seems as though 8 + 13 + 20 = 41% of the residual gap is already able to be targeted based on known sources of bias. This could be differential across traits, like it already seems to be with "Number of children", meaning it could explain more or less.

Indeed, if we do these based on the differential numbers already provided, then the number isn't 41%, it's 54%: 46% remains elsewhere! Or maybe here?

We shouldn't speculate more though, since what we have to speculate on is imprecise and possibly biased in unknown ways.

One thing I'm very curious about is the adjustments

It seems as though they overcontrolled, leading to major underestimates of the heritability of educational attainment and fluid intelligence

This looks liable to remove true heritable geographically-structured genetic variance

Something curious is that in Extended Data Fig. 2, we get different estimates from the main + suppl ones for height and educational attainment, and there's no significant missing heritability with these, plus, adjusting for assortative mating kills the HE-reg vs. GREML gap too.

The other trait they re-estimated with AM in mind was fluid intelligence.

That showed a significant baseline heritability gap with GREML, but it became nonsignificant with AM correction and in the uncorrected and corrected (gap direction reversed!) HE-reg results.

Overall, this is a very cool paper and I'm glad it's finally out. People have been looking forward to this for a long while, and it's nice to see what the state-of-the-art is, and to get hints of what's next to come.

Study's here: nature.com/articles/s4158…