Getting more SNPs with PIP > 0.9 than expected

[DivyanshiSrivastava]

Hi,

I'm trying to use CARMA to finemap. However, I run into an issue where I'm getting multiple variants fine-mapped to high PIPs despite these variants being in strong LD with each other.

As an example, here I'm finemapping a region of chromosome 1: Here are all variants with CS != 0:

SNP	CHR	BP	A1	A2	MAF	Z	PIP	CS
rs17400053.	1	161362821	C	G	0.138954	-2.554073	0.01236889	5
rs34550741	1	161373788	G	A	0.136364	-2.787327	0.01639989	5
rs12758795	1	161377161	C	T	0.139193	-2.627630	0.03521533	5
rs35825219	1	161377619	G	A	0.139197	-2.625070	0.02333522	5
1:161378027_GC_G	1	161378027	GC	G	0.139471	-2.738607	0.34884646	5
rs34370781	1	161381532	G	C	0.139121	-2.685047	0.25500650	5
rs66579967	1	161382179	G	T	0.139386	-2.712943	0.30011416	5
rs11578929	1	161388081	C	T	0.107278	4.371659	1.00000000	4
rs148566739	1	161388081	C	CATATATAT	0.076377	-1.301230	1.00000000	1
rs6680630	1	161388100	T	A	0.108031	4.368306	1.00000000	2
rs78358576	1	161388104	T	A	0.103696	4.590882	1.00000000	3

However, when I look at the corresponding (matching) indices in the LD matrix (UKB derived LD matrix provided by PolyFun), the last 3 SNPs ( all assigned PIP=1) are in high-LD, with LD values are close to 1.

I'm passing a dense LD matrix, with sumstats in the format described above. I'm using no annotations, but get the same result with annotations, and the outlier detection is False. This is happening across several loci for me, any idea on what could be going on here, and how I can troubleshoot? I'm using R 4.2.1.

Thanks! Divyanshi

[ZikunY]

Hi Divyanshi,

Thank you for taking interest in CARMA! Yes, the results you showed here are problematic and the typical results biased by the outliers, which is rs148566739 in this incidence. Please take on the outlier detection as "outlier.switch=T", which should help the situation. Please let me know it could solve the problem.

Thanks.

Best, Zikun

[davsmlee]

Hi,

Just wanted to follow-up on this - are variants with CS = 0 not part of a credible set? Or should these be considered as part of a credible set as well? Thanks!

[igarcia17]

Hi @davsmlee As far as I know, variants that do not belong to any credible set (CS=0) are not causal, and shouldn't be considered part of any credible set. However I struggle to link the concept of PIP to credible set. When I run CARMA with no annotations, in some cases I get that the variant with the highest PIP (approx 0.45) is not part of any credible set (CS=0). I understand that the set of variants present in a credible set have to sum up to 1 in order to consider them as jointly contributing to the trait. Maybe I am in a missconception, but how can it be that a 0.45 PIP variant is not considered as causal? I apologize in advance if it is a naive question, or if I am confused grasping the concepts. Kind regards Inés

[ZikunY]

Hi @davsmlee and @igarcia17 ,

If the variants do not belong to any credible set (CS=0), then it can only be interpreted as the variants do not belong to any credible sets...... It should not be interpreted as the variants are not causal.

The credible set is composited with a group of variants with high LD (characterized by a lower bound of LD) such that the sum of PIPs of this group is over 0.95 (or 0.99).

If in a loci, the highest PIP is around 0.45 but not included in any credible set which means that this variant is not highly correlatd to any credible set, or on the other hand, the groups of variants that highly correlated to this variant didn't have a sum of PIP over 0.95. But this definitely ignores the situation of outliers, which would surely complicate the situation.

In the end, the PIP characterized the probability of variants being causal, so 0.45 is a decent chance that the variants being causal, maybe worth of looking for annotations or ANNOVAR, depending on the actual situation.

Best, Zikun