Huge thanks to my co-authors and mentors - Douglas Wightman, Christiaan de Leeuw, and @daniposthu.bsky.social — for their guidance and collaboration, and to the REALMENT consortium for supporting this work.

🧵 6/6

Take home:
Additive PGSs remain the most robust default for most complex traits.
ML/DL can help when traits are:
• highly heritable
• low in polygenicity
• driven by strong dominance deviations

Full paper + code: github.com/nybell/non-a...

🧵 5/6

In the UK Biobank (10 traits), ML/DL models outperformed additive PGSs for traits known to show dominance - including lipoprotein(a), alkaline phosphatase, and ApoB - but not for height (no dominance).

🧵 4/6

Across most scenarios, additive PGSs were remarkably robust - even when up to 20% of SNP-h² came from dominance SNPs.

Performance dropped mainly for traits with:
• high SNP-h²
• low polygenicity
• strong dominance deviations

🧵 3/6

Most PGS methods assume additivity - each allele contributes linearly to risk - but real traits can show dominance deviations.

We simulated phenotypes varying in:
• SNP heritability (SNP h²)
• % heritability from dominance
• polygenicity
• dominance deviation strength

🧵 2/6

When do machine learning models actually outperform standard polygenic scores? 🤔

In our new preprint, we benchmark how non-additive genetic effects (i.e, dominance deviations) shape polygenic prediction across simulated and UK Biobank traits.

👉 www.medrxiv.org/content/10.1...

🧵 1/6