A multi-task convolutional deep neural network for variant calling in single molecule sequencing.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Accurate DNA sequence variant identification is crucial but challenging, particularly with single-molecule sequencing's high error rates (~5-15%).
• Existing methods struggle to reliably detect variants in noisy sequencing data.

Purpose of the Study:

• To develop a robust deep learning model for accurate DNA variant calling from diverse sequencing data.
• To address the limitations of current variant identification techniques, especially for single-molecule sequencing.

Main Methods:

• Developed Clairvoyante, a five-layer convolutional neural network (CNN) model.
• Employed a multi-task learning approach to predict variant type (SNP/indel), zygosity, alternative allele, and indel length.
• Trained and validated the model on human sample data (NA12878) using Illumina, PacBio, and Oxford Nanopore sequencing reads.

Main Results:

• Achieved high F1-scores for variant calling on NA12878: 99.67% (1KP common variants) and 98.65% (whole-genome) with Illumina data.
• Demonstrated strong performance with PacBio (95.78%/92.57%) and Oxford Nanopore (90.53%/87.26%) data.
• Showcased sample agnosticism and rapid variant detection (<2 hours on a standard server).
• Identified 3,135 variants missed by Illumina but supported by both PacBio and Oxford Nanopore reads.

Conclusions:

• Clairvoyante offers a significant advancement in DNA variant identification, particularly for error-prone single-molecule sequencing data.
• The model's sample-agnostic nature and efficiency make it a valuable tool for large-scale genomic studies.
• Open-source availability facilitates broader adoption and further development in the genomics community.