• info@icbcongress.com
• English
• Persian

• Login

Create Free Account

INTERNATIONAL CONGRESS OF

PharmacoGenetics

• AI-Assisted Ancestry-Specific Pharmacogenomics: Bridging the Representation Gap in Precision Medicine

• Arman Taran,^1,* Mina Shirmohammadpour,² Bahman Mirzaei,³
  1. Zanjan University of Medical Sciences
  2. Zanjan University of Medical Sciences
  3. Zanjan University of Medical Sciences
  
  
  
• Introduction: The field of pharmacogenomics has long struggled with a critical limitation: the overwhelming bias toward European-ancestry populations in genomic datasets, which compromises medication safety and efficacy for underrepresented groups. Current pharmacogenomic guidelines, derived primarily from Caucasian populations, demonstrate reduced predictive accuracy when applied to African, Asian, and Indigenous populations due to differences in allele frequencies, haplotype structures, and gene-environment interactions. Artificial intelligence offers transformative potential to address this disparity by integrating multi-ancestry genomic data with clinical outcomes to develop population-specific dosing algorithms. This review examines how machine learning approaches are enabling ancestry-aware pharmacogenomic predictions and their impact on reducing adverse drug reactions in globally diverse populations.
• Methods: Our analysis incorporated data from 23 pharmacogenomic studies (2018-2023) encompassing 1.2 million individuals across 7 ancestry groups. We evaluated three AI approaches: 1) Admixture-aware neural networks trained on the TOPMed cohort (n=450,000) to predict warfarin dosing requirements, 2) Federated learning models analyzing CYP2D6 metabolizer status across 12 healthcare systems while preserving data privacy, and 3) Graph convolutional networks mapping population-specific haplotype structures in the PharmGKB database. Model performance was assessed using ancestry-stratified metrics including sensitivity for actionable variants (MAF >1%), dosage prediction accuracy (mean absolute error), and clinical outcome improvement (hospitalization reduction).
• Results: The AI models demonstrated significant improvements over conventional methods. Admixture-informed neural networks reduced warfarin dosing errors by 38% in African Americans (MAE=0.7 mg/day vs 1.13 mg/day for standard algorithms; p<0.001) by incorporating ancestry-specific CYP2C9*8 and VKORC1 variants. Federated learning achieved 92% concordance in CYP2D6 phenotyping across Asian subpopulations, addressing historical misclassification of "intermediate metabolizers" due to uncharacterized variants. Most notably, graph-based approaches identified 17 novel population-specific drug-gene interactions, including a clinically significant association between NAT2 haplotypes and isoniazid toxicity in Aboriginal Australians (OR=3.2, 95% CI 1.7-5.9).
• Conclusion: These findings highlight AI's capacity to mitigate healthcare disparities through ancestry-conscious pharmacogenomics. The success of admixture-aware models underscores the importance of explicitly accounting for genetic ancestry rather than treating it as a covariate. However, challenges persist in clinical implementation, including: 1) Limited representation of certain populations (e.g., Pacific Islanders) even in large datasets, 2) Ethical considerations in commercial genetic testing, and 3) The need for real-world validation in diverse healthcare settings. Future directions should prioritize: 1) Development of dynamic models that adapt to newly sequenced populations, 2) Integration of socioenvironmental factors with genomic data, and 3) Creation of regulatory frameworks for equitable AI deployment in global precision medicine initiatives.
• Keywords: Pharmacogenomics, Artificial intelligence, Drug metabolism, Population genetics, Machine learning