bogdanlab / h2-gre

Implemention of Accurate estimation of SNP-heritability from biobank-scale data irrespective of genetic architecture. (Nature Genetics 2019)

h2-GRE (Generalized Random Effects model)

Scripts for methods described in Accurate estimation of SNP-heritability from biobank-scale data irrespective of genetic architecture.

Example use

Please see the page SNP heritability estimation for Biobank for an example.

User Support

Please first go through documentation. This repository is in active development and your user experience will be invaluable to make our code widely usable. If you have questions or suggestions, feel free to contact us via email or open a Github issue.

h2-gre's Issues

h2 for imbalanced binary traits

Thanks for sharing this useful tool that I am testing in some large datasets. I have some questions regarding the analysis of (imbalanced) binary traits. I notice that h2 estimates are strongly inflated for binary traits with prevalence <0.01, when effect sizes are estimated using OLS.

Do you have any specific recommendations for (imbalanced) binary traits? Would there be any utility in using, for instance, logistic/Firths log(OR) and SEs instead of BETA and SE from OLS, or would this not be valid?

On that note, using some of the provided example scripts, PLINK1.9 and PLINK2 actually automatically revert to logistic regression for binary traits (rather than using OLS; even when using —linear). Not sure if this is known with you guys and how you managed to force plink to run OLS. (Or perhaps it is only newer versions of PLINK that have this issue.)

Either way, thanks greatly and I await your reply.
Best,
Sean

Reference Panel

Hi,

in the help, you mention that currently, the same individuals must be used to perform the association and the heritability estimation. This is not practical for meta-analyses. Do you plan on adding an option for a reference panel soon?

Thanks,

Arthur

Code suggestions

Hi, the software works great, thanks very much, it was possible to perform the whole analysis in a few hours.
A couple of minor issues I found (using 130k individuals and 600k snps);

I had to install fire (as well as python-fire)

I had to use --chr_i (rather than --chr-i)

I first performed the association testing using PLINK2, but for the final step (estimating snp heritability for each chromosome) got a error whose final line was
AttributeError: 'DataFrame' object has no attribute 'CHR'
This was because in PLINK2, the association results files have a column called #CHROM not CHR. Therefore, I manually renamed this column. I likewise changed ID to SNP and OBS_CT to NMISS
If I instead used PLINK 1.9 for the association analysis (as suggested), I got this error
assert((legend['A1'] == sumstats['A1']).all())
AssertionError
I think this was because in association analysis, PLINK1.9 does not automatically use A1 from bim file as the test allele.

Finally, I guess I can get an overall sd by adding up the variances for each chromosome?

Otherwise, as I say, was very quick and easy to use, thanks