Bridging Ancestry Gaps in Genomic Risk Prediction with Tabular Foundation Models

Area of Science:

• Genomics
• Machine Learning
• Population Genetics

Background:

• Genomic prediction models show performance disparities across diverse populations due to sample size imbalances and varying genotype-phenotype effect sizes.
• Existing models struggle with ancestry-driven heterogeneity in genetic effect sizes, limiting their clinical utility.

Purpose of the Study:

• To evaluate the effectiveness of in-context learning (ICL)-capable tabular foundation models for genotype-to-phenotype prediction across diverse ancestries.
• To develop and assess an instruction-tuning framework to improve model robustness to ancestry-driven effect size heterogeneity.

Main Methods:

• Utilized large, ancestrally diverse biobank data to assess ICL foundation models.
• Developed an instruction-tuning framework treating genetic ancestry as a continuous variable, incorporating synthetic tasks with ancestry-dependent non-stationary effects.
• Evaluated model performance across the genetic ancestry continuum.

Main Results:

• ICL foundation models demonstrated reduced performance degradation in under-sampled ancestry groups compared to traditional supervised methods.
• Prevailing models failed when allele effect sizes varied across ancestry space.
• Instruction-tuned models achieved improved and more stable predictive performance across the genetic ancestry continuum, even for individuals distant from training exemplars.

Conclusions:

• Instruction-tuned tabular foundation models offer a promising approach to bridge ancestry gaps in genomic risk prediction.
• This method enhances predictive accuracy and stability, improving the clinical utility of genomic prediction models for diverse populations.