Interesting article from @WiringTheBrain that develops variational auto encoders (VAEs) as a metaphor for the relationship between genome and organism, with the genome corresponding to the compressed representation. I like the 'generative model' metaphor better than the 'blueprint' metaphor, although the genome also specifies part of the 'decoder', i.e. developmental program, so can't be reduced to the compressed representation. Genomes also encode different developmental programs for different cell types so the biological picture for a complex multicellular organism is much messier. But I like how this metaphor grasps at the mystery of how something as complex as an animal with a brain that produces specific behaviors can be reliably developed from a linear code of limited complexity.

RT @Dr_delaFuente: 🧬 New Feature in Genomic SEM! 🧬 The new rgmodel() function is now available in Genomic SEM, providing a more precise way…

RT @QuantaMagazine: All hash tables help a computer store and access data. Andrew Krapivin’s new version takes less time and fewer steps. h…

RT @pnatarajanmd: So fantastic that @NPR featured the unexplained troublingly high rates of heart disease among S Asian individuals, and ou…

RT @SpencrGreenberg: If this were true, then the number of adverse childhood experiences people report (ACEs) would be extremely correlated…

@unboxpolitics @lu_sichu @krichard1212 The problem is the UKB doesn't have enough twins or siblings for this

RT @ed_hagen: Cannibalism is fairly common in the Upper Paleolithic but its connection to warfare is uncertain. This study suggests a famil…

@lu_sichu @krichard1212 It's better than them not having it. We have to give them some credit for putting anything like this in when most biobanks don't have anything or don't let you access it. But yeah people overinterpret results derived from it.

@lu_sichu @krichard1212 Yeah we shouldn't be using the UKB 'fluid intelligence' test to make strong claims about properly measured IQ. It's also got serious ascertainment issues in that only a more educated subset of the UKB, already an ascertained sample, took the test.

UKB added a bunch of additional cognitive test data that has improved things, but this hasn't been integrated into much work yet. It's not been a priority for biomedical research. If they're interested in cognitive data at all, it's usually because of studying age related cognitive decline. The behaviour genetics people generally collected better cognitive phenotypes because they're from a psychology background. What phenotype collection makes sense depends a bit on what question you're interested in. If you just want a good heritability estimate, you don't need huge samples, so could collect high quality data if motivated for a reasonable cost. However if you want samples in the millions, it's going to be hard to get great cognitive test data. I think we should move towards collecting SAT scores or exam marks/grades in UK as a better proxy than education level. Linkage to military test data also a potentially promising avenue. But you need buy in from the people who control these things and it's often not there.

"What I'm pretty certain of is that math heritability is much, much lower than 50%". I'm not sure I agree with this. Twin studies typically find that specific ability heritabilities are similar to the general factor. If we measured the heritability of SAT-math I expect it wouldn't be that different from IQ. But I wonder whether we may be talking past each other somewhat. Genetics almost certainly plays an important role in whether someone has the ability to be proficient at higher mathematics at all, but within the set of individuals who have that level of ability, genetics may play a lesser role in who actually achieves things. Other factors like interest, determination, creativity, health, luck, etc - some also influenced by genetics - are also important.

No method is perfect. Twin estimates might be upward biased for various reasons, but they're not useless or easily dismissed. I've written before on the missing heritability problem: Since then, genomic analyses of siblings have given heritability estimates closer to twin estimates than I expected. So the method that I developed, RDR, may be underestimating heritability due to ultra rare variants and non additive effects. Both twin and siblings/RDR underestimate heritability when there's assortative mating so that needs to be factored in too. My sense is that heritability of IQ is in the range of 30-70% with very high confidence. It would be fairly straightforward to settle these debates if we had the right data: large scale genotyped family data with good quality cognitive and educational outcomes. But for various reasons this is not yet readily available.

You did trash twin studies of IQ, though, so it's fair to respond to that. If you're making a point that IQ and genetics don't explain everything when it comes to real world mathematical or scientific achievement then I would agree. @erikphoel has written well on this:

RT @RonDePinho: If we want to thoughtfully reduce NIH indirect costs, we need to reduce the federal regulatory burden that causes universit…

@SashaGusevPosts @iamChalzz @AnaerobicLagoon @BronskiJoseph @davidbessis Yeah the noise in the h2 estimates and potential bias in the PGS values are the main concerns. We're still some way off having all the relevant parameters estimated well in the same dataset...

@mbeisen Plausible that this change leads to more non-research costs getting billed through direct costs, meaning less money going to research - unless they increase how much one can bill in direct costs.

@phil49472744 @phl43 @lu_sichu Poorly defined and conceptualized. And often invalid arguments, i.e. they don't show what they think they show.

RT @MariosGeorgakis: Most genetic studies test associations with disease incidence — there are few genome-wide explorations for disease pro…