Differences between PRS-derived single-trait MTAG GWAS and GWAMA summary statistics for SBP & DBP

[brendaudosen]

Hello MTAG Developer,

I’m using MTAG to generate single-trait GWAS summary statistics for systolic blood pressure (SBP) and diastolic blood pressure (DBP). As input, I start from GWAMA-derived summary statistics (univariate inverse-variance meta-analyses). After running MTAG in single-trait mode, I derive polygenic risk scores (PRS) from both the MTAG outputs and the original GWAMA summary statistics. However, I’m observing systematic differences in downstream PRS performance and effect‐size distributions between the two.

What I’ve tried so far

Data inputs: GWAMA summary stats for SBP and DBP (QC’d, harmonized across cohorts). MTAG run with only that single trait (no additional phenotypes).

PRS construction
Clumping + thresholding using identical parameters.
Same LD reference panel.

Comparisons
Effect‐size distributions (β, SE, Z) between MTAG vs. GWAMA.
PRS‑R² in an independent validation cohort.

All processing steps were held constant, yet MTAG-derived PRS consistently shows slightly lower effect sizes and marginal differences in R².

Questions to clarify
Weighting & intercept adjustment
Does MTAG apply its LD Score intercept‐based bias correction and noise‐covariance reweighting even in single-trait mode?

Shrinkage/Prior assumptions
Is there any empirical-Bayes shrinkage applied when k = 1 that would systematically pull βs toward zero?

Implications for PRS
Should we treat MTAG single-trait GWAS as fundamentally a different estimation method (and thus expect differences in PRS), or is there a way to "turn off" the multivariate adjustments to match GWAMA exactly?

Recommended best practice
If the goal is downstream PRS construction for a single trait, would you recommend running GWAMA directly over MTAG, or is there a configuration to align MTAG outputs with standard meta-analysis outputs?

Any guidance on why these discrepancies arise, despite identical input GWAMA stats, and how best to interpret or mitigate them for PRS applications would be greatly appreciated. Thank you!

[paturley]

Hello, MTAG summary statistics actually transform the effect sizes to correspond to a standardized phenotype, so I would expect the effect size to be a constant multiple of the input summary statistics. In terms of prediction, single-trait MTAG results adjust the SEs using the LD score intercept. To the degree that this adjustment is large, that could affect the output of the PRS algorithm that you are using. If you are doing a pruning and thresholding approach, I'd expect the R2 of the PRS to be smaller since intercept-corrected SEs generally lead to larger SEs, so fewer SNPs will get included. If you are using a genome-wide method, I'm not sure why MTAG-based PRS would have higher or lower predictive power than GWAMA -based PRS. Best, Patrick

…

On Tue, Jul 22, 2025 at 2:36 PM Brenda Udosen ***@***.***> wrote: *brendaudosen* created an issue (JonJala/mtag#238) <#238> I’m using MTAG to generate single-trait GWAS summary statistics for systolic blood pressure (SBP) and diastolic blood pressure (DBP). As input, I start from GWAMA-derived summary statistics (univariate inverse-variance meta-analyses). After running MTAG in single-trait mode, I derive polygenic risk scores (PRS) from both the MTAG outputs and the original GWAMA summary statistics. However, I’m observing systematic differences in downstream PRS performance and effect‐size distributions between the two. What I’ve tried so far Data inputs: GWAMA summary stats for SBP and DBP (QC’d, harmonized across cohorts). MTAG run with only that single trait (no additional phenotypes). PRS construction Clumping + thresholding using identical parameters. Same LD reference panel. Comparisons Effect‐size distributions (β, SE, Z) between MTAG vs. GWAMA. PRS‑R² in an independent validation cohort. All processing steps were held constant, yet MTAG-derived PRS consistently shows slightly lower effect sizes and marginal differences in R². Questions to clarify Weighting & intercept adjustment Does MTAG apply its LD Score intercept‐based bias correction and noise‐covariance reweighting even in single-trait mode? Shrinkage/Prior assumptions Is there any empirical-Bayes shrinkage applied when k = 1 that would systematically pull βs toward zero? Implications for PRS Should we treat MTAG single-trait GWAS as fundamentally a different estimation method (and thus expect differences in PRS), or is there a way to "turn off" the multivariate adjustments to match GWAMA exactly? Recommended best practice If the goal is downstream PRS construction for a single trait, would you recommend running GWAMA directly over MTAG, or is there a configuration to align MTAG outputs with standard meta-analysis outputs? Any guidance on why these discrepancies arise, despite identical input GWAMA stats, and how best to interpret or mitigate them for PRS applications would be greatly appreciated. Thank you! — Reply to this email directly, view it on GitHub <#238>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFBUB5L6DSTWXC6U6DNEQNT3J2AERAVCNFSM6AAAAACCDWEBN2VHI2DSMVQWIX3LMV43ASLTON2WKOZTGI2TGNRZGQ3TKNA> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[brendaudosen]

Thanks again, Patrick, for clarifying how MTAG rescales effect sizes and adjusts standard errors via the LD‐score intercept. I wanted to share the follow‑up from my PRS comparisons:

Single‐trait MTAG vs. GWAMA single‑trait
For SBP, the pruning‐and‐thresholding PRS built on MTAG single‐trait summary statistics explained 0.44 % of phenotypic variance in AWI‑Gen, versus 0.19 % for the GWAMA PRS. For DBP, the MTAG single‑trait was 0.59 % vs. 0.22 % for GWAMA.

This improvement aligns with MTAG’s use of the LD‐informed meta‐analysis weights: by leveraging cross‐cohort LD patterns and standardizing the trait, MTAG produces more powerful SNP effect estimates even after accounting for the intercept correction. In contrast, GWAMA’s inverse‐variance meta doesn’t re‐scale to unit variance or incorporate the LD‐score intercept in the same way, resulting in less optimal weights for PRS.

Regards,
Brenda