3 ms·

I thought I'd register to add a few interesting points to this thread. 'Classical' GWAS are a top-down approach, where we try to find strong associations betwe

I thought I'd register to add a few interesting points to this thread. 'Classical' GWAS are a top-down approach, where we try to find strong associations betweens a genetic locus and a trait (in this case intelligence), and they rely on very stringent p-value corrections due to the number of tests performed. This is why although height is >0.6 heritable (meaning genetics explain more than 60% of the trait variance in a given population), only 20% of the variance can be attributed to GWAS finds [1], such as the 52 genes in this article. Bear in mind that the height results were obtained with the largest GWAS ever with more than 250,000 samples and 300 institutes [2]. So where is the missing heritability? This is the starting point of a bottom-up approach, where we try to predict a trait from the genotype. With some clever but simple calculation, we know that most heritability can be found in the data (even if it does not pass the stringent significant thresholds). This is what is called 'chip heritability', the heritability explained by all the common genetic variants present on commercial chips taken together. With height, the chip heritability reaches 62.5% [1]. The bottom-up approach is especially looked at for psychiatric disorders, and already delivered some interesting results such as 0.8 genetic correlation between schizophrenia and bipolar disorder [3]. But useful clinical diagnostics based on genetic data are still a long way off. This should please the HN crowd to know that we still have one tough prediction problem to crack. Current results can be improved either with more sophisticated machine learning methods or by integrating various biological annotations to try to unearth the missing heritability that GWAS can't seem to find with the current sample sizes. Here are a few more links for those interested [4] [5]. [1] https://www.ncbi.nlm.nih.gov/pubmed/25282103?dopt=Citation https://www.ncbi.nlm.nih.gov/pubmed/25282103?dopt=Citation [2] https://www.broadinstitute.org/news/giant-study-reveals-giant-number-genes-linked-height https://www.broadinstitute.org/news/giant-study-reveals-gian... [3] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4797329/ https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4797329/ [4] http://onlinelibrary.wiley.com/doi/10.1002/gepi.21966/full http://onlinelibrary.wiley.com/doi/10.1002/gepi.21966/full [5] https://academic.oup.com/bioinformatics/article/29/2/206/202786/A-Lasso-multi-marker-mixed-model-for-association https://academic.oup.com/bioinformatics/article/29/2/206/202...

nonbel 9y ago
>"This is why although height is >0.6 heritable (meaning genetics explain more than 60% of the trait variance in a given population), only 20% of the variance can be attributed to GWAS finds" I don't see how you can assign numbers like this. As a simple example, say the amount of protein absorbed while growing up affects adult height. Then you can have a certain protein catabolyzing gene that is highly correlated with height when protein in the diet is low, but has little relationship when dietary protein is high (since the person gets enough absorbed protein either way). So the variance explained (either by a given genetic locus or in total) is not going to be any kind of constant value, and can in fact vary wildly.
- alextheparrot 9y ago
  He is probably measuring something called "narrow-sense heritability", which just includes just additive genetic effects (Gene A + Gene B = Sum(A, B)) and excludes other things such as gene interaction. Most of the statistical techniques used to estimate the narrow-sense heritability assume environmental and other, non-additive, genetic effects follow a Gaussian. Your example clearly exposes a situation where these studies may not be powerful. The inability to explain heritability as more than a "sum of the parts" is a problem which I'm not familiar with a solution for at the genome scale.
  - nonbel 9y ago
    Do you know of an example of someone doing one of these heritability calculations, saving the model, then plugging in data from a new dataset into the same model? (ie the new data did not exist at the time the model was created)

nonbel 9y ago

>"This is why although height is >0.6 heritable (meaning genetics explain more than 60% of the trait variance in a given population), only 20% of the variance can be attributed to GWAS finds" I don't see how you can assign numbers like this. As a simple example, say the amount of protein absorbed while growing up affects adult height. Then you can have a certain protein catabolyzing gene that is highly correlated with height when protein in the diet is low, but has little relationship when dietary protein is high (since the person gets enough absorbed protein either way). So the variance explained (either by a given genetic locus or in total) is not going to be any kind of constant value, and can in fact vary wildly.

alextheparrot 9y ago
He is probably measuring something called "narrow-sense heritability", which just includes just additive genetic effects (Gene A + Gene B = Sum(A, B)) and excludes other things such as gene interaction. Most of the statistical techniques used to estimate the narrow-sense heritability assume environmental and other, non-additive, genetic effects follow a Gaussian. Your example clearly exposes a situation where these studies may not be powerful. The inability to explain heritability as more than a "sum of the parts" is a problem which I'm not familiar with a solution for at the genome scale.
- nonbel 9y ago
  Do you know of an example of someone doing one of these heritability calculations, saving the model, then plugging in data from a new dataset into the same model? (ie the new data did not exist at the time the model was created)