EmbedGEM: a framework to evaluate the utility of embeddings for genetic discovery

Area of Science:

• Computational biology
• Genetics
• Machine learning

Background:

• Machine learning embeddings offer compressed representations of complex biological data.
• Embeddings show potential for capturing disease states and aiding genetic discovery.
• A key limitation is determining if genetic variants associated with embeddings are relevant to the specific disease or trait.

Purpose of the Study:

• Introduce EmbedGEM (Embedding Genetic Evaluation Methods), a framework to systematically evaluate the utility of embeddings in genetic discovery.
• Compare embeddings based on heritability and disease relevance.
• Provide a generalizable approach for multivariate traits.

Main Methods:

• EmbedGEM evaluates embeddings on two axes: heritability and disease relevance.
• Heritability is measured by genome-wide significant associations and mean statistics at significant loci.
• Disease relevance is assessed using polygenic risk scores and their association with trait labels in held-out data.

Main Results:

• EmbedGEM was demonstrated on synthetic and UK Biobank data (metabolic and liver traits).
• The framework successfully ranked traits based on heritability and disease relevance.
• Results indicate that high heritability does not necessarily correlate with high disease relevance for embeddings.

Conclusions:

• EmbedGEM provides a systematic method for evaluating the genetic relevance of embeddings.
• The framework is applicable to multivariate traits and can be extended.
• This work highlights the importance of assessing both heritability and disease relevance for meaningful genetic discovery using embeddings.