Bridging ancestry gaps in genomic risk prediction with tabular foundation models

Area of Science:

• Genomics
• Artificial Intelligence
• Population Genetics

Background:

• Genomic prediction models show variable performance across different populations, hindering clinical use.
• Key limitations include unequal sample sizes and shifting genotype-phenotype effect sizes across ancestries.
• The efficacy of tabular foundation models with in-context learning (ICL) for genotype-to-phenotype prediction and their resilience to ancestry-related effects are not well understood.

Purpose of the Study:

• To evaluate the effectiveness of ICL-capable tabular foundation models for genomic prediction in diverse populations.
• To address the challenge of ancestry-driven heterogeneity in genotype-phenotype effect sizes.
• To develop and assess an instruction-tuning framework for improving genomic prediction models across the genetic ancestry continuum.

Main Methods:

• Utilized large, ancestrally diverse biobank data to assess model performance.
• Compared conventional supervised methods with ICL-capable tabular foundation models.
• Developed an instruction-tuning framework using synthetic tasks with ancestry-dependent non-stationary effects, treating genetic ancestry as a continuous variable.

Main Results:

• ICL-capable foundation models demonstrated reduced performance degradation in under-sampled ancestry groups compared to traditional approaches.
• Existing models trained on standard synthetic data struggled with varying allele effect sizes across ancestries.
• The instruction-tuned models exhibited enhanced and more consistent predictive performance across the genetic ancestry spectrum, even for individuals genetically distant from training examples.

Conclusions:

• Instruction-tuned tabular foundation models offer a promising approach to bridge ancestry gaps in genomic risk prediction.
• This method improves the equity and robustness of genomic prediction tools.
• The developed framework and models show potential for broader clinical utility by accounting for population-specific genetic effects.