Using TrajGWAS for large-scale datasets: how to improve performance?

[parekhpravesh]

I would like to run TrajGWAS on some large-scale longitudinal phenotypes. Specifically, I have 100,000 observations, 48 covariates (+ intercept), and 100 phenotypes. I would like to get effect size estimates as well (so running a Wald test)

As an example, I ran TrajGWAS for one phenotype. I start Julia with: julia --threads 64 and then do:

trajGWAS(@formula(y ~ 1 + X1 + X2 + ... + X48),
@formula(y ~ 1),
@formula(y ~ 1 + X1 + X2 + ... + X48),
:id,
path_to_csv_file,
path_to_plink_file,
pvalfile = p_output_name,
nullfile = null_output_name,
covrowinds = covrowmask,
genetic rowinds = geneticrowmask,
parallel = :true,
test = :wald)

I am doing this as a slurm job with --cpus-per-task=64 and mem-per-cpu=7G specifications. Julia version: 1.10.0

However, after about 22 hours, only about 700 SNPs have been written to the output file. This is quite a bit slow and I wonder if there are any suggestions on how to make this efficient? Perhaps I am not specifying parallelisation correctly?

[parekhpravesh]

Hello, just following up on this - I tried the same settings and after 3+ days of computation, only about 3000 SNPs were written to the output file. Do you have any tips/suggestions on how the performance can be improved?

[kose-y]

Oh, sorry for the late response. The Wald test, giving the effect sizes, is much slower than the score test, which does not give the effect sizes. Our suggestion is first to screen the SNPs with the score test and take a subset of SNPs with low p-values, then compute the effect sizes using the Wald test only for the selected SNPs.

[parekhpravesh]

Thank you - I tried running a score test and could finish the analysis in ~35 hours - could you confirm if I am specifying the parallelisation option correctly? Or is everything implemented for single threaded computation and it doesn't really matter?