A genotype-phenotype transformer to assess and explain polygenic risk

Area of Science:

• Genetics
• Bioinformatics
• Machine Learning

Background:

• Genome-wide association studies (GWAS) identify genetic variants linked to traits but lack mechanistic insight.
• Epistatic interactions (gene-gene effects) complicate understanding of genetic architecture.
• Transformer models offer a novel approach to analyze complex biological data.

Purpose of the Study:

• Introduce the Genotype-to-Phenotype Transformer (G2PT) for modeling hierarchical genetic information flow.
• Apply G2PT to predict the triglyceride to high-density lipoprotein cholesterol (TG/HDL) ratio, a key metabolic health indicator.
• Demonstrate G2PT's capability in uncovering complex genetic interactions and improving predictive accuracy.

Main Methods:

• Developed G2PT, a hierarchical graph transformer framework.
• Utilized G2PT to model the genetic basis of the TG/HDL ratio.
• Analyzed attention patterns to identify key genetic variants and systems involved.

Main Results:

• G2PT achieved state-of-the-art accuracy in predicting TG/HDL ratio.
• Identified 1,395 genetic variants across at least 20 biological systems.
• Discovered 40 epistatic interactions, including a novel interaction between *APOA4* and *CETP* in phospholipid transfer.

Conclusions:

• Hierarchical graph transformers represent a powerful next-generation tool for polygenic risk analysis.
• G2PT provides mechanistic insights into complex genotype-phenotype relationships.
• This approach enhances understanding of metabolic health genetics and potential therapeutic targets.