Improving machine learning reproducibility in genetic association studies with proportional instance cross validation...

Area of Science:

• Biomedical data analysis
• Statistical genetics
• Machine learning applications

Background:

• Machine learning is increasingly used for complex biomedical data, including genome-wide association studies (GWAS).
• Reproducibility in GWAS analysis, especially for genetic interactions, is challenged by biological and statistical factors.
• Traditional cross-validation can fail due to imbalanced genotype distributions in GWAS data.

Purpose of the Study:

• To introduce a novel cross-validation method, proportional instance cross-validation (PICV).
• To address the challenge of imbalanced genotype distributions in GWAS data analysis.
• To improve the reproducibility of machine learning analyses in GWAS.

Main Methods:

• Proposing proportional instance cross-validation (PICV) to maintain original variable distributions during data splitting.
• Applying PICV to simulated GWAS data with varying epistatic interaction parameters.
• Comparing PICV performance against traditional random cross-validation.

Main Results:

• PICV significantly improved sensitivity and positive predictive value in simulated GWAS data compared to traditional cross-validation.
• Application to primary open-angle glaucoma GWAS data showed improved consistency between training and testing results.
• A previously reported interaction in glaucoma GWAS did not replicate, but PICV enhanced result consistency.

Conclusions:

• Modifications to machine learning procedures, like PICV, are necessary for biomedical data with imbalanced distributions.
• PICV enhances the reproducibility of genetic interaction findings by accounting for variable imbalance.
• The PICV approach is potentially applicable to other domains facing imbalanced variable distribution challenges.