Machine learning improves SNP microarray performance in challenged samples

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning Applications

Background:

• SNP microarrays are a cost-effective genotyping tool across disciplines.
• Current microarray analysis methods often require high-quality DNA, limiting their use with challenged samples.
• Uncertainty in genotype calls from low-quantity DNA needs robust handling.

Purpose of the Study:

• To evaluate machine learning algorithms for genotype and genotype likelihood estimation from SNP microarray data.
• To address uncertainty in genotype calling for low-quantity DNA samples.
• To develop a more direct estimate of genotype quality for microarray data.

Main Methods:

• Application of several machine learning algorithms, including neural networks and XGBoost.
• Estimation of genotypes and genotype likelihoods using Illumina Omni5-4 microarray data.
• Comparison of algorithm performance and generalization capabilities.

Main Results:

• XGBoost demonstrated strong performance and better generalization across different sample types on Omni5-4 chips compared to neural networks.
• Machine learning approaches can represent genotype uncertainty probabilistically, improving data utility.
• XGBoost provides a direct estimate of genotype quality, a valuable feature for microarray analysis.

Conclusions:

• XGBoost is a promising machine learning method for improving genotype accuracy and quality assessment in SNP microarray analysis, especially with challenged samples.
• Probabilistic genotype calling enhances data compatibility with downstream analyses.
• The development of direct genotype quality estimates addresses a key limitation in current microarray analysis.