Statistical modeling for sensitive detection of low-frequency single nucleotide variants.

Area of Science:

• Genomics
• Bioinformatics
• Molecular Biology

Background:

• Low-frequency single nucleotide variants (SNVs) are crucial in cancer genetics and population studies.
• Challenges include tumor heterogeneity, low circulating tumor DNA in liquid biopsies, and signal dilution in pooled sequencing.
• Existing methods lack sensitivity and fail to account for sequencing artifacts, with detection limits around 2-5%.

Purpose of the Study:

• To develop a method for sensitive detection of low-frequency SNVs, pushing the detection limit close to sequencing error rates.
• To model erroneous read counts based on genomic sequence contexts.
• To evaluate different statistical distributions for count data modeling in generalized linear models.

Main Methods:

• Modeled observed erroneous read counts using genomic sequence contexts.
• Characterized four count data distributions (generalized linear models) for goodness-of-fit and performance on real sequencing data.
• Tested the method on two sequencing technologies (Ion Proton and Illumina MiSeq) with different chemistries to identify systematic errors.

Main Results:

• The zero-inflated negative binomial distribution generalized linear model demonstrated superior performance, especially for variants between 0.5% and 1%.
• Achieved high recall and precision across different sequencing platforms: 95.3% recall and 79.9% precision for Ion Proton, and 95.6% recall and 97.0% precision for Illumina MiSeq at >=1% frequency.
• The method's detection limit was found to be around 0.5% SNVs and is generalizable to various sequencing technologies.

Conclusions:

• The developed method enables sensitive detection of low-frequency SNVs across diverse sequencing platforms.
• Facilitates advancements in pooled sequencing, early cancer detection, prognostic assessment, and monitoring of metastasis, relapse, or acquired resistance.