LDpred2 - Strong Overfitting in Variance Explained Between GWAS and Non-GWAS Samples

[TheoCavinato]

I am observing significant overfitting when comparing the variance explained (r²) for individuals included in a GWAS versus those outside of it.

Pipeline Overview:

1. We perform GWAS on a subset of the UK Biobank (~150K individuals).
2. We apply LDpred2 on the summary statistics to estimate betas.
3. We compute polygenic scores (PGS) for UK Biobank individuals using these betas.
4. We assess the variance explained (r²) separately for individuals inside and outside the GWAS subset.

Results:

For standing height:

• r²_in = 0.30
• r²_out = 0.16

For BMI:

• r²_in = 0.47
• r²_out = 0.1

The large discrepancy suggests substantial overfitting.

Question:

Are there specific diagnostics or metrics I could use to better understand the source of this issue? Any recommendations for reducing the overfitting?

Thank you for your time!

[privefl]

Of course, you cannot test in individuals (or any of their relatives) that were used in the training (GWAS).

A similar question was opened a few days ago: #537.