Equitable machine learning counteracts ancestral bias in precision medicine

Area of Science:

• Genomics
• Bioinformatics
• Artificial Intelligence in Medicine

Background:

• Genomic datasets disproportionately represent European ancestries, limiting understanding of human disease and equitable medical advancements.
• Ancestral bias in genomic data hinders the discovery of population-specific disease insights and therapeutic targets.

Purpose of the Study:

• To introduce PhyloFrame, a machine learning (ML) method designed to correct ancestral bias in genomic data for equitable precision medicine.
• To improve the predictive power and gene identification capabilities of genomic models across diverse ancestries.

Main Methods:

• PhyloFrame integrates functional interaction networks and population genomics with transcriptomic data to adjust for ancestral bias.
• The method was applied to breast, thyroid, and uterine cancer datasets and validated across fourteen diverse genomic datasets.

Main Results:

• PhyloFrame demonstrated improved predictive power and reduced model overfitting across all ancestries, particularly for underrepresented groups.
• The method showed a higher likelihood of identifying known cancer-related genes and effectively mitigated performance impacts due to phylogenetic distance from training data.

Conclusions:

• Equitable artificial intelligence (AI) approaches like PhyloFrame can mitigate ancestral bias in genomic training data.
• These advancements are crucial for achieving equitable representation in medical research and improving healthcare outcomes for all populations.