Model is not converging evena fter running for 10k iterations

[akhilpampana]

Hello,

I am running susie_rss using meta gwas results and 1000g reference dataset. I made the directionality consistent as best I can but still I am getting the model not converging even after checking for 10k iterations. Although its a warning, I am worried that it may will impact the result so would like to know if there is a way to fix this.

Also is there a way to supply prior variance calculated separately?

Regards Akhil

[pcarbo]

@akhilpampana This isn't necessarily a problem. Could you please share the last, say, 100 lines of output from running susie?

Also is there a way to supply prior variance calculated separately?

Yes, you would set argument "scaled_prior_variance" and estimate_prior_variance = TRUE; see help(susie) for details.

[akhilpampana]

Hey , What do you mean when 100 lines of output from running susie? I only see some values and all so thought of asking?

[pcarbo]

If you set verbose = TRUE, this will output the algorithm's progress (change in the objective function over time). Can you share, say, the last 100 lines of this output?

[akhilpampana]

log_file.txt

Here is the file as requested. I have tested with decrease L and increasing max_iterations and noting resolved this warning.

[pcarbo]

Thanks for sharing — yes, I agree that after 100 iterations, susie is still far from reaching a fixed point. Have you assessed the consistency of the LD matrix and z-scores as we suggest following this vignette?

[akhilpampana]

I have tested the consistency of the ld matrix with z_scores and checked other information in the files. I have observed the lambda and plotted the distribution which looked like this (https://github.com/stephenslab/susieR/issues/223#issuecomment-2067049020). I flipped the alleles and performed all the necessary steps needed but still I am getting the same warning. Its not convergent. I have also observed that there are no lead snps in the credible set if there is a warning of not reaching 100 iterations. I have tested 100,1000,10000 even 100000 but still haven't got convergence. I am not sure how to resolve this issue. The gwas analysis is a meta-analysis and uses all individuals of 1000g as a ld reference.

Regards Akhil

[pcarbo]

Judging by this kriging_rss plot, if it is up-to-date, it suggests that your z-scores are very inconsistent with your LD matrix. I'm not sure why they are so inconsistent, but it suggests that either (1) your z-score and/or LD calculations have some bug, or (2) you need to consider a difference reference panel for computing LD. Even if susie_rss did converge, I would not expect your results to be reliable.

Also, since you are dealing with a meta GWAS, you may want to consider other more recent fine-mapping methods that are better suited for meta analyses.

[akhilpampana]

Thank you for the clarification. Do you have any recommendations of any new method that I can test out?

Regards Akhil

On Fri, May 3, 2024, 6:04 AM Peter Carbonetto @.***> wrote:

Judging by this kriging_rss plot, if it is up-to-date, it suggests that your z-scores are very inconsistent with your LD matrix. I'm not sure why they are so inconsistent, but it suggests that either (1) your z-score and/or LD calculations have some bug, or (2) you need to consider a difference reference panel for computing LD. Even if susie_rss did converge, I would not expect your results to be reliable.

Also, since you are dealing with a meta GWAS, you may want to consider other more recent fine-mapping methods that are better suited for meta analyses.

— Reply to this email directly, view it on GitHub https://github.com/stephenslab/susieR/issues/226#issuecomment-2092783288, or unsubscribe https://github.com/notifications/unsubscribe-auth/AEUJ56F3JVFG2HATWVGY5RLZANVLDAVCNFSM6AAAAABHAZQHXSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDAOJSG44DGMRYHA . You are receiving this because you were mentioned.Message ID: @.***>

[pcarbo]

I don't have much experience with these methods, but this preprint is a good starting point.