LDSC -Genetic correlation, some results all have NA valuies

Question

[d.s.zimmerman]

Hi there,

I am performing a genetic correlation analysis and receive NA results for some of the traits I'm testing but works for others. This is the error I get:

*ERROR computing rg for phenotype 2/2, from file /home/expcard/Projects/GWAS_SCA/GWAS_NTR/LDSC/NEWFORMAT/Schunkert_CAD_2011.sumstats.gz.

Traceback (most recent call last):

File "/home/expcard/Projects/GWAS_SCA/GWAS_NTR/LDSC/ldsc/ldscore/sumstats.py", line 410, in estimate_rg rghat = _rg(loop, args, log, M_annot, ref_ld_cnames, w_ld_cname, i)

File "/home/expcard/Projects/GWAS_SCA/GWAS_NTR/LDSC/ldsc/ldscore/sumstats.py", line 539, in _rg intercept_gencov=intercepts[2], n_blocks=n_blocks, twostep=args.two_step)

File "/home/expcard/Projects/GWAS_SCA/GWAS_NTR/LDSC/ldsc/ldscore/regressions.py", line 705, in __init__ np.multiply(hsq1.tot_delete_values, hsq2.tot_delete_values))

FloatingPointError: invalid value encountered in sqrt*

Has anyone come across this issue? Could it be that the heritability of my own trait is too low?

Thanks!

Answer 1

[Sam]

It is likely that your heritability is too low or has a huge SE, causing the h2 estimate to be negative, which error out LDSC (because to calculate the rg, it needs to calculate sqrt (h^2a x h^2b)

You can check the h2 doing a single trait LDSC (--h2).

[d.s.zimmerman]

Hi Sam,

Thanks for your reply, here are the --h2 results, would this heritability be too low? Also, is the ratio also not too high here? What could be causing this?

Total Observed scale h2: 0.013 (0.0532)

Lambda GC: 1.0405

Mean Chi^2: 1.0412

Intercept: 1.0388 (0.0075)

Ratio: 0.9416 (0.1822)

[Sam]

The SE of your h2 is too high, that's why. You can check if the number of SNPs overlapped with the baseline LD score is high enough (require at least 200k SNPs across the genome). If the overlap is high and you still got such a low h2, you can try and recalculate the LD score using the 1000G phase 3 data and the baseline annotation provided by LDSC to generate a new set of baseline and use only the SNPs in your data to calculate a corresponding weight file. This might help to improve the power a bit. If that doesn't work, then it is very likely that your GWAS does not have enough power.