question about susie_rss

[myz1118]

Thanks for your tool on fine-mapping with GWAS summary data.
I have some questions about it. I used the susie_rss function with the following code: fitted_rss3 <- susie_rss(c(susie_z$zscore), ld_matrix1, n=susie_z$N, estimate_residual_variance = FALSE)
However, I encountered an error: WARNING: XtX is not symmetric; forcing XtX to be symmetric by replacing XtX with (XtX + t(XtX))/2
Error in init_finalize(s) : Input residual variance sigma2 must be a scalar
How can I resolve this? Also, I want to know that the sample sizes N for the loci in my data are inconsistent. Can I use different N values for each locus, or do I have to fix a single N value for all loci? Thanks very much!

[pcarbo]

@myz1118 The susie model assumes that the sample size is the same for all SNPs; it is unclear how accurate the fine-mapping will be if you have different sample sizes for the different SNPs. Ideally, you could make the sample sizes the same, e.g., by genotype imputation. See this paper for some related discussion.

[pcarbo]

Also note that you can try not providing the n input to susie_rss(); it is not required.

[myz1118]

thanks! and what about ERROR" However, I encountered an error: WARNING: XtX is not symmetric; forcing XtX to be symmetric by replacing XtX with (XtX + t(XtX))/2
Error in init_finalize(s) : Input residual variance sigma2 must be a scalar"

[pcarbo]

@myz1118 My guess is that this error should go away if you do not provide "n". Have you checked that?

[myz1118]

Sorry for the late reply. I saw the response just now. Thank you for your suggestion; it was very helpful. However, after removing the "n", I encountered other issues：“The estimated prior variance is unreasonably large.
This is usually caused by mismatch between the summary statistics and the LD matrix.
Please check the input.”.
Additionally, I verified that my LD matrix is symmetric, but the warning keeps indicating that it is not. It says “WARNING: XtX is not symmetric; forcing XtX to be symmetric by replacing XtX with (XtX + t(XtX))/2
Error in susie_suff_stat(XtX = R, Xty = z, n = 2, yty = 1, scaled_prior_variance = prior_variance, : ” . Do you have any solutions for these two problems? Thank you very much!

[pcarbo]

@myz1118 Typically the warning can be ignored; it may only be slightly numerically non-symmetric, although you might want to check manually to make sure.

Regarding the prior variance, we have found that sometimes the EM algorithm (estimate_prior_method = "EM") produces more stable estimate of the prior variance, so you could try this. Alternatively, you could estimate_prior_variance = FALSE to see if this at least yields reasonable results, noting that it is generally preferrable to estimate the prior variance.

[myz1118]

Sorry to bother you again, but after trying your suggestion, I’m still getting the ERROR as follow：
fitted_rss3 <- susie_rss(z=susie_z$zscore, R=ld_matrix1, estimate_prior_method = NULL)
WARNING: Providing the sample size (n), or even a rough estimate of n, is highly recommended. Without n, the implicit assumption is n is large (Inf) and the effect sizes are small (close to zero).
HINT: For large R or large XtX, consider installing the Rfast package for better performance.
WARNING: XtX is not symmetric; forcing XtX to be symmetric by replacing XtX with (XtX + t(XtX))/2
Error in susie_suff_stat(XtX = R, Xty = z, n = 2, yty = 1, scaled_prior_variance = prior_variance, :
The estimated prior variance is unreasonably large.
This is usually caused by mismatch between the summary statistics and the LD matrix.
Please check the input.

LD matrix is generated using the following command, and the order is consistent with the order of significant loci Z-values.
Plink v1.90b6.21 64-bit (19 Oct 2020)
Options in effect:
--bfile significant_snps
--matrix
--out susie_ld
--r2

I would like to know if there are any other solutions, or if the issue might be with the data itself. Thank you again for your reply.

[pcarbo]

@myz1118 estimate_prior_method = NULL has no effect; you need to do estimate_prior_variance = FALSE if you don't want to estimate the prior (I'm not sure if that's what you are trying to do).

In PLINK, you have to be careful about making sure that the association testing and LD computations are consistent in what alleles are used as the "counted" allele; see Issue #240 for some discussion on this.

(Also, the default settings are different in PLINK and PLINK2, so one has to be careful about the version.)

[fmorgante]

@pcarbo @myz1118 it looks like you are computing r^2, while susie_rss requires r.

[pcarbo]

@fmorgante Good catch!

[myz1118]

@fmorgante @pcarbo Thank you very much for your valuable suggestions. I have modified my LD matrix from R² to R, as you recommended, but I am still encountering issues. Specifically, when running the susie_rss function: fitted_rss3 <- susie_rss(z = susie_z$zscore, R = ld_matrix_r, estimate_prior_variance = FALSE) WARNING: Providing the sample size (n), or even a rough estimate of n, is highly recommended. Without n, the implicit assumption is n is large (Inf) and the effect sizes are small (close to zero).
HINT: For large R or large XtX, consider installing the Rfast package for better performance.
WARNING: XtX is not symmetric; forcing XtX to be symmetric by replacing XtX with (XtX + t(XtX))/2
Warning message:
In susie_suff_stat(XtX = R, Xty = z, n = 2, yty = 1, scaled_prior_variance = prior_variance, :
IBSS algorithm did not converge in 100 iterations!
Please check consistency between summary statistics and LD matrix.
See https://stephenslab.github.io/susieR/articles/susierss_diagnostic.html

Afterward, I referred to issue #240 and realized that the user encountered a similar problem. I tried using （estimate_s_rss） to check the consistence between my Z-scores and the LD matrix: lambda <- estimate_s_rss(susie_z$zscore, ld_matrix_r)
The result was 0.87. Does this indicate that there might be some trouble with the data? GWAS summary statistics was generated using FIZI software, and the LD matrix was generated using PLINK 1.9.

Thank you very much for your help ！

[pcarbo]

@myz1118 I also suggest using the "kriging_rss" function to assess consistency of the LD and z-scores. I think that will be helpful.