Causal signal with high PIP while not reaches significance.

[Xi-Cao]

Hi~thanks for your helpful tool on fine-mapping with GWAS summary data.

I recently performed fine-mapping analysis using SuSiE with GWAS summary statistics. Variants with MAF < 0.01, multi-allelic sites, and those located in the major histocompatibility complex region were prefiltered. The analysis ran smoothly without any errors or warnings.

In one complex region, SuSiE identified six credible sets (Figure 1). One of them (L6) contains a single variant, rs11204088, with a high PIP of 0.98. However, this variant did not even reaches nominal significance in the GWAS summary statistics (Z = -1.18, P = 0.23) (Figure 2). I checked the LD mismatch using the kriging_rss function, which appeared to be well-matched (Figure 3). The logLR of this variant is also negative (Figure 4). In addition, I searched for the related variants in high LD with rs11204088 (rs4401880, r = 0.84), which also showed no significance in GWAS summary statistics (P = 0.47).

I'm a bit puzzled by this result. Could you provide any guidance on why this might have occurred, and how I should interpret or further validate my results?

Best regards,
xi cao

[gaow]

@Xi-Cao this might not be an uncommon observation when you compare marginal association analysis vs joint analysis (eg fine-mapping). Some variants, when conditional on others, may appear significant even though marginally they were not. This happens especially in regions with multiple causal variants (is this expected for your particular GWAS)?

In this case it also seems your marginally most significant variant has a low PIP at the end ... Could you try to run a simple conditional regression by setting max_iter to 1 in susie_rss and see if that still keeps the strongest marginal association significant, but already promoting some marginal non-significant variants to having high PIP?

[Xi-Cao]

Hi~Sincerely thank you for your response.

According to LD clumping, this region contains four independent significant variant clusters in my GWAS. I reran susie_rss with max_iter = 1 as you suggested. This time, SuSiE identified four credible sets, and excluded rs11204088 (PIP = 0.14). In contrast, the most marginally significant variant, rs115849089, showed a higher PIP of 0.21 than before and was included in one of the credible set s(L1).

So, in this case, would it be reasonable to interpret that fine-mapping excluded potential false positives from the marginal association analysis and helped uncover likely causal variants that were previously masked due to LD?

Additionally, as mentioned above, I pre-filtered variants with MAF < 0.01, as I found that the EAF and LD estimates for ultra-rare variants are highly dependent on sampling and prone to LD mismatch. However, I’m wondering that if there were some marginally significant variants among those filtered out, could this compromise the reliability of the fine-mapping results? In such cases, what would be the best strategy to ensure the robustness of the conclusions?

Best regards,
xi cao

[pcarbo]

@Xi-Cao You could also check the MAF of rs11204088. Judging by the statistics you report, I would be suspicious of rs11204088 being causal for your phenotype.

[Xi-Cao]

Hi~Sincerely thank you for your advice.

I checked the MAF of rs11204088 in both the GWAS and LD panel. There is a slight discrepancy between them (MAF = 0.099 in GWAS vs. 0.094 in the LD panel). However, no LD mismatch was detected using the kriging_rss function. Given this, would the final fine-mapping results still be considered reliable?
Additionally, I’m really confused about the second question I mentioned earlier regarding the filtering of marginally significant rare variants before analysis. Could you kindly offer some suggestions on this?

Best regards,
xi cao

[pcarbo]

@Xi-Cao There is no single "correct" answer to your second question because there is an inherent tradeoff. It sounds like the rarer SNPs are less reliable. On the other hand, when reporting your results, it is possible that a reported causal SNP is actually tagging a rarer SNP that is the truly causal SNP. This sort of tradeoff is not unique to your analysis; the most important thing, in my opinion, is to be transparent in reporting your results.

[Xi-Cao]

I see. Thank you very much for your patient explanation！