MSRCall: a multi-scale deep neural network to basecall Oxford Nanopore sequences

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Third-generation sequencing technologies like Oxford Nanopore Technologies' MinION offer real-time, long-read data crucial for genomic analysis.
• The MinION device infers nucleotide sequences by analyzing ionic current signals as DNA/RNA fragments pass through nanopores.
• Accurate basecalling, the process of translating these signals into nucleotide sequences, remains a challenge for MinION compared to traditional next-generation sequencing (NGS) methods.

Purpose of the Study:

• To develop an improved basecalling algorithm for MinION sequencing data.
• To enhance the accuracy of nucleotide sequence determination from nanopore-generated signals.
• To address the limitations in MinION's basecalling accuracy compared to existing NGS basecallers.

Main Methods:

• A novel neural network architecture, the multi-scale recurrent caller (MSRCall), was designed.
• MSRCall incorporates a multi-scale structure and recurrent layers to capture dependencies across various time scales.
• The architecture includes a fusion block and a connectionist temporal classification decoder for efficient basecalling.

Main Results:

• The proposed MSRCall model demonstrates superior performance over existing basecallers.
• MSRCall achieves higher read accuracy, leading to more reliable genomic data.
• The model also shows improved consensus accuracy, crucial for variant detection and genome assembly.

Conclusions:

• MSRCall represents a significant advancement in basecalling accuracy for MinION sequencing.
• The neural network's multi-scale recurrent design effectively captures complex signal patterns.
• This improved basecalling accuracy will enhance the utility of portable nanopore sequencing for diverse genomic applications.