HELLO

HELLO calls small genomic variants using customized deep neural network architectures to improve variant-calling accuracy across Illumina, PacBio, and hybrid Illumina-PacBio datasets.

Key Features:

• Customized Deep Neural Networks: HELLO employs deep neural network (DNN) architectures optimized to directly interpret sequencing data rather than reformulating variant calling as an image-recognition problem.
• Comparison to DeepVariant and Inception-v3: Unlike DeepVariant's image-recognition approach that uses Inception-v3, HELLO uses DNNs specifically tailored for sequencing-data characteristics.
• Efficiency and Compactness: HELLO achieves superior accuracy with significantly smaller model sizes, reducing indel call errors by up to 18% for Illumina, 55% for PacBio, and 65% for hybrid Illumina-PacBio datasets.
• Versatility Across Platforms: The method is applicable to Next Generation Sequencing (NGS), Third Generation Sequencing, Illumina, PacBio, and hybrid Illumina-PacBio settings.
• Tailored Variant Inference Functions: The approach incorporates variant inference functions that are explicitly tailored to the nature of sequencing data.

Scientific Applications:

• Genomic research: Improves the reliability of small-variant calls in general genomic analyses.
• Disease association studies: Reduces variant-calling errors to enhance detection of genotype–phenotype associations in disease research.
• Evolutionary biology: Provides more accurate small-variant calls for population- and evolutionary-genetics analyses.
• Personalized medicine: Increases accuracy of variant identification relevant to clinical and precision-medicine applications.

Methodology:

HELLO uses customized DNN architectures that incorporate variant inference functions tailored to sequencing data, with smaller model sizes to reduce computational resource requirements.