Expanding biobank pharmacogenomics through machine learning calls of structural variation

Area of Science:

• Genomics and Bioinformatics
• Pharmacogenomics
• Clinical Research

Background:

• Biobanks integrating genetic and clinical data are valuable for pharmacogenomic (PGx) research.
• Existing array-based genotypes can identify many PGx alleles, but challenges remain for complex variations.

Purpose of the Study:

• To create a comprehensive pharmacogenomic (PGx) allele and phenotype callset for over 80,000 participants in the Michigan Genomics Initiative (MGI) biobank.
• To develop and validate a novel computational method for identifying PGx alleles dependent on structural variation, specifically the CYP2D6*5 deletion.

Main Methods:

• Utilized PyPGx software on TOPMed imputed genotypes to generate a central PGx callset.
• Developed a support vector machine model trained on genotype array SNV probe intensities to detect CYP2D6*5 structural variants.
• Validated array-based PGx calls against PCR-validated clinical data.

Main Results:

• Achieved >92% concordance between array-based PGx calls and PCR-validated alleles.
• The support vector machine demonstrated >99% accuracy in identifying CYP2D6*5 carriers.
• Reclassified approximately 7% of African American and 4% of White participants to lower activity metabolizer phenotypes, indicating potential for increased adverse drug reactions.

Conclusions:

• Central PGx callsets derived from existing biobank data are feasible and valuable for research.
• Customized computational methods can effectively identify challenging PGx alleles, including structural variants, enhancing biobank utility.
• Augmenting standard PGx callsets with advanced methods broadens research potential and clinical applicability in pharmacogenomics.