EmbedGEM

EmbedGEM systematically evaluates the utility of embeddings in genetic discovery, focusing on heritability and disease relevance.

Key Features:

• Heritability assessment: Measures the number of genome-wide significant associations and the mean chi-squared statistic at significant loci.
• Disease relevance evaluation: Computes polygenic risk scores for each embedding principal component and evaluates their association with high-confidence disease or trait labels in a held-out evaluation patient set.
• General applicability to multivariate traits: The approach extends beyond embeddings to accommodate any multivariate traits and additional metrics along the evaluation axes.
• Dual-dataset validation: Evaluates embeddings and multivariate traits using both synthetic simulated datasets and real-world data from the UK Biobank.

Scientific Applications:

• Genetic discovery with machine learning embeddings: Determines whether genetic variants associated with embeddings are relevant to the disease or trait of interest.
• Metabolic trait analysis: Evaluates embeddings and multivariate traits for metabolic traits using UK Biobank data.
• Liver-related trait analysis: Evaluates embeddings and multivariate traits for liver-related traits using UK Biobank data.
• Ranking trait utility: Ranks traits based on their heritability and disease relevance in synthetic datasets with known ground truth.

Methodology:

EmbedGEM compares embeddings along two axes: heritability and disease relevance. For heritability, it considers the number of genome-wide significant associations and the mean chi-squared statistic at significant loci. For disease relevance, it computes polygenic risk scores for each embedding principal component, then evaluates their association with high-confidence disease or trait labels in a held-out evaluation patient set.