Overcoming limitations to customize DeepVariant for domesticated animals with TrioTrain

Area of Science:

• Genomics
• Bioinformatics
• Comparative Genomics

Background:

• Variant calling tools often assume human genome characteristics, limiting their effectiveness in diverse species.
• DeepVariant (DV) is a powerful variant caller, but its performance with non-human genomes is not fully understood.
• The development of universal algorithms necessitates evaluating their impact across different species.

Purpose of the Study:

• To assess the performance of human-trained variant callers, including DeepVariant with allele frequency (DV-AF) and DeepTrio (DT), on bovine genomes.
• To develop and validate a novel approach, TrioTrain, for extending DV to diploid species lacking established reference resources like Genome-in-a-Bottle (GIAB).
• To create the first multispecies-trained DV-AF checkpoint using bovine trios.

Main Methods:

• The TrioTrain approach automates DV extension for diploid species by using a region shuffling strategy to overcome SLURM-based cluster limitations.
• Animal truth labels were curated to exclude Mendelian discordant sites before training DV for accurate offspring genotyping.
• Bovine trios (cattle, yak, bison) were utilized to generate a multispecies-trained DV-AF checkpoint.

Main Results:

• A multispecies-trained DV-AF checkpoint achieved a mean SNV F1 score >0.990 during GIAB benchmarking, despite limitations in bovine truth sets for repetitive regions.
• A bovine-trained DV checkpoint (checkpoint 28) reduced the Mendelian inheritance error (MIE) rate by 50% compared to the default human-trained DV when tested on HG002.
• Checkpoint 28 demonstrated a mean MIE rate of 0.03% across three bovine interspecies cross genomes.

Conclusions:

• A multispecies, trio-based training strategy effectively reduces inheritance errors in single-sample variant calling.
• Exclusively human-trained models hinder the application of deep learning-based variant calling to new species.
• The diverse ancestry within bovids highlights the need for advanced comparative genomics tools tailored for non-human species.