Estimating error models for whole genome sequencing using mixtures of Dirichlet-multinomial distributions

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Accurate genotype identification is fundamental for genomic data analysis, including polymorphism identification, disease association studies, and mutation rate determination.
• Genotyping accuracy can be compromised by various biological and technical factors such as copy-number variation, paralogous sequences, library preparation, sequencing errors, and reference-mapping biases.

Purpose of the Study:

• To develop an improved statistical model for read depth analysis to enhance genotype accuracy.
• To identify and mitigate sources of error in genomic data that lead to inaccurate genotype calls.

Main Methods:

• Modeled read depth data using a mixture of Dirichlet-multinomial distributions.
• Developed a model comprising two main distributions: a major component (low error, low bias) and a minor component (overdispersed, high error, high bias).
• Identified sequence sites fitting the minor component as enriched for copy-number variants and low complexity regions.

Main Results:

• The Dirichlet-multinomial mixture model significantly improved upon existing models for read depth analysis.
• The minor component distribution was found to be overdispersed, exhibiting higher error and reference bias.
• Sequence sites fitting the minor component were enriched for copy-number variants and low complexity regions, which are known sources of erroneous genotype calls.
• Removing sites that did not fit the major component led to a demonstrable improvement in genotype call accuracy.

Conclusions:

• The proposed Dirichlet-multinomial mixture model effectively distinguishes between reliable and error-prone sequencing sites.
• By filtering out sites associated with the minor component, genotype accuracy can be substantially enhanced, particularly in the presence of copy-number variations and other biases.
• This approach offers a robust method for improving the reliability of genotype calls in diverse genomic datasets.