MTAG-derived polygenic risk score shows a worse prediction performance

[WQ-CHENG]

I'm testing whether MTAG can help with polygenic risk score prediction for a pair of traits with poor genetic correlation. My primary phenotype is bipolar disorder (BIP), and I performed MTAG on BIP versus three other variables: education level, cortical thickness, and cortical surface area. I then identified BIP cases and healthy controls using the polygenic risk score (PRS) for BIP from GWAS and from MTAG. All input GWAS were obtained from PGC, UKB, and SSGAC, and are based on a large cohort.

For BIP vs. surface area and BIP vs. cortical thickness, the MTAG-based PRS of BIP performed very identically to the GWAS-based PRS of BIP. This makes sense given BIP has non-significant rg with those cortical measures. The MTAG-based PRS of BIP for BIP vs education attainment, however, performs even worse than the GWAS-based PRS. This seems a bit strange to me. I expected that the MTAG-based PRS would perform a little bit better or equally to the GWAS-based PRS because BIP has a low but significant genetic correlation with educational achievement.

I checked the log files from MTAG. The number of SNPs used in MTAG looks fine but the “GWAS equiv. (max) N” is much smaller in the MTAG analysis for BIP vs education attainment, comparing with the other two. The barplot and log files are provided below. I’m wondering is this the reason that the MTAG-based PRS performs worse than GWAS-based PRS?Thank you!

BIP_vs_EDU.log BIP_vs_SA.log Pcutoff.pdf

[paturley]

Hmm. This is a bit of a puzzle to me to. The main thing I find strange is that the mean chi2 for trait1 falls after MTAG. Because MTAG SEs are strictly smaller than the corresponding GWAS SEs, this can only happen if the the MTAG on average shrinks the important effect size estimates. Or if MTAG drops those SNPs because they don't overlap across the two datasets.

What happens if you make a PGS only based on the set of SNPs that are overlapping in the two datasets? Also what happens if you plot the MTAG effect size against the GWAS effect size for the SNPs included in the 10e-5 SNPs for the GWAS PGS?

On Thu, Mar 9, 2023 at 3:03 PM W.Q.CHENG @.***> wrote:

I'm testing whether MTAG can help with polygenic risk score prediction for a pair of traits with poor genetic correlation. My primary phenotype is bipolar disorder (BIP), and I performed MTAG on BIP versus three other variables: education level, cortical thickness, and cortical surface area. I then identified BIP cases and healthy controls using the polygenic risk score (PRS) for BIP from GWAS and from MTAG. All input GWAS were obtained from PGC, UKB, and SSGAC, and are based on a large cohort.

For BIP vs. surface area and BIP vs. cortical thickness, the MTAG-based PRS of BIP performed very identically to the GWAS-based PRS of BIP. This makes sense given BIP has non-significant rg with those cortical measures. The MTAG-based PRS of BIP for BIP vs education attainment, however, performs even worse than the GWAS-based PRS. This seems a bit strange to me. I expected that the MTAG-based PRS would perform a little bit better or equally to the GWAS-based PRS because BIP has a low but significant genetic correlation with educational achievement.

I checked the log files from MTAG. The number of SNPs used in MTAG looks fine but the “GWAS equiv. (max) N” is much smaller in the MTAG analysis for BIP vs education attainment, comparing with the other two. The barplot and log files are provided below. I’m wondering is this the reason that the MTAG-based PRS performs worse than GWAS-based PRS?Thank you!

BIP_vs_EDU.log https://github.com/JonJala/mtag/files/10935590/BIP_vs_EDU.log BIP_vs_SA.log https://github.com/JonJala/mtag/files/10935591/BIP_vs_SA.log Pcutoff.pdf https://github.com/JonJala/mtag/files/10935592/Pcutoff.pdf

— Reply to this email directly, view it on GitHub https://github.com/JonJala/mtag/issues/175, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFBUB5JJV3AMVU4VWRE54N3W3IZSXANCNFSM6AAAAAAVVR62MA . You are receiving this because you are subscribed to this thread.Message ID: @.***>

[WQ-CHENG]

Hi Patrick, Thank you for your quick response. I tested the performance of a new BIP-PRS based on the GWAS estimates of the overlapped SNPs between BIP and EDU GWAS. This PRS showed a very similar performance as the original GWAS-based PRS and even slightly better (please see the fig1). The number of SNPs used to construct PRS based on GWAS, overlapped_GWAS, and MTAG are similar (285667, 283862, and 283822).

For independent SNPs used to construct a PGS at a P threshold of 10e-05, I examined the MTAG and GWAS beta values. Fig.2 displays the distribution of beta values for those SNPs. Then  I compared the absolute values of each SNP's MTAG beta and GWAS beta to see how they varied from one another. I also provided the results for BIP_vs_TH as a comparator. Although it appears that the MTAG beta are both lower than the GWAS beta, the BIP_vs_EDU returns lower beta than BIP_vs_TH. If I understand it correctly, the smaller estimates may lead to smaller difference of PRS between cases and controls, resulting to a smaller PRS.R2. However, the MTAG results for BIP_vs_TH showed a very similar performance with GWAS, while MTAG results for BIP_vs_EDU is worse. May I ask what other variables might have an impact on MTAG estimates, besides their genetic correlation and GWAS power?  Thank you!

Fig1.pdf Fig2.pdf

Here is a bit of details if needed:

1. I have removed MHC region from all analysis, including the MTAG and PRS analysis.
2. All PRS analysis was performed using PRSice and the same target samples. All input base file (i.e., GWAS sumstats and MTAG sumstats) has been QCed before PRS analysis following the guidance for PRSice.