Diagnosing migraine from genome-wide genotype data: a machine learning analysis.

Area of Science:

• Genetics
• Computational Biology
• Neurology

Background:

• Migraine has a polygenic basis, but genome-wide association studies (GWAS) explain only part of its heritability.
• A significant portion of migraine heritability remains unexplained, suggesting the involvement of non-additive and interactive genetic effects.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for migraine prediction, aiming to capture non-additive and interactive genetic effects.
• To address the 'missing heritability' in migraine genetics by employing advanced computational approaches.
• To compare the performance of ML models against traditional polygenic risk scoring (PRS).

Main Methods:

• A population-based, cross-sectional study using data from the Trøndelag Health Study (43,197 participants).
• Genome-wide genotyping and phenotyping based on International Classification of Headache Disorders criteria.
• Development and optimization of various ML and deep learning models, including gradient boosting and Naïve Bayes, using PLINK and LDPred2 for PRS.

Main Results:

• Machine learning models achieved superior classification performance (Area Under Curve [AUC] 0.62-0.63) compared to PRS (AUC 0.52-0.59) across datasets with varying genetic variant numbers (p<0.001 to p=0.02).
• The best performing ML models included a gradient boosting machine for smaller datasets and a multinomial Naïve Bayes model for the largest dataset.
• ML identified known migraine-associated genes and pathways, alongside novel pathways related to signal transduction, neurological function, botulinum toxins, and the calcitonin gene-related peptide receptor.

Conclusions:

• Migraine genetics may involve non-additive and interactive causal structures, better captured by complex ML models than additive PRS.
• The effectiveness of ML models highlights their potential to uncover genetic architectures masked by large data dimensionality and limited sample sizes.
• Future research with larger sample sizes using ML can enhance migraine precision medicine by elucidating complex genetic interactions.