Comparing feature selection and machine learning approaches for predicting CYP2D6 methylation from genetic variation.

Area of Science:

• Genomics
• Pharmacogenetics
• Epigenetics

Background:

• Pharmacogenetics currently relies on limited genetic variants for clinical decisions, impacting pediatric prescribing accuracy.
• Integrating genomic information like methylation can enhance pharmacogenetic models and improve prescribing.
• Cytochrome P450 2D6 (CYP2D6) is a key gene in drug metabolism, making it a target for precision dosing.

Purpose of the Study:

• To predict epigenetic loci from single nucleotide polymorphisms (SNPs) related to CYP2D6 in children.
• To develop and validate SNP-based prediction models for CYP2D6 CpG methylation.
• To explore the potential of integrating genomic and demographic data for improved pharmacogenetic models.

Main Methods:

• DNA methylation was quantified using Illumina Infinium Methylation EPIC beadchip.
• Linear Regression, Elastic Net, and XGBoost models were employed to predict CpG sites associated with CYP2D6.
• Feature selection involved SNPs from GWAS, GTEx, and proximity to the CYP2D6 gene, with and without demographic data.

Main Results:

• Elastic Net models using genetic features showed marginally better performance than heritability estimates and significantly outperformed Linear Regression and XGBoost.
• The inclusion of non-genetic features improved model performance for certain probes and feature sets.
• The optimal machine learning approach and feature set varied across different CpG sites, with several top variables identified for each model.

Conclusions:

• SNP-based prediction models for CYP2D6 CpG methylation were developed in Singaporean children of diverse ethnicities.
• These models have potential clinical applications in precision medicine and pharmacogenetics-based dosing.
• Further validation is needed to integrate these models into clinical practice for improved drug prescribing.