A random forest classifier for detecting rare variants in NGS data from viral populations

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Next Generation Sequencing (NGS) generates vast amounts of data, crucial for understanding viral populations.
• Distinguishing rare variants from sequencing errors is a significant challenge in viral genomics.
• Accurate variant detection is essential for viral diversity estimation and *de novo* assembly.

Purpose of the Study:

• To develop a novel method for accurate detection of rare variants in viral NGS data.
• To differentiate sequencing errors from true low-frequency genetic variations.
• To improve the accuracy of viral population diversity analysis and *de novo* assembly.

Main Methods:

• A Random Forest classifier, named MultiRes, was developed.
• The classifier utilizes k-mer counts from viral population reads.
• The algorithm incorporates signal processing concepts, specifically frame-based representations of k-mers.

Main Results:

• MultiRes demonstrated 4 to 500 times fewer false positive k-mer predictions compared to existing methods.
• The method achieved high recall for true k-mers, comparable to established error correction tools.
• MultiRes showed >95% recall for single nucleotide polymorphisms (SNPs) with fewer false positives, identifying more rare variants.

Conclusions:

• MultiRes offers a significant improvement in distinguishing rare variants from sequencing errors in viral NGS data.
• The method enhances the accuracy of viral population diversity estimation and *de novo* assembly.
• MultiRes provides a valuable tool for genomic analysis of viral populations, with software freely available.