Related Experiment Video
Updated: Aug 8, 2026
07:59
Training Dogs for Awake, Unrestrained Functional Magnetic Resonance Imaging
Published on: October 13, 2019
7.5K
Overcoming Limitations to Deep Learning in Domesticated Animals with TrioTrain
Jenna Kalleberg1, Jacob Rissman1, Robert D Schnabel1,2
1University of Missouri, Division of Animal Sciences, Columbia, MO, 65201 USA.
Biorxiv : the Preprint Server for Biology
|April 25, 2024
Summary
Developing a multi-species DeepVariant (DV) model improves variant calling accuracy in non-human genomes. This new approach, TrioTrain, reduces Mendelian Inheritance Error (MIE) rates, overcoming limitations of human-genome-trained models for broader genomic applications.
Area of Science:
- Genomics and Bioinformatics
- Comparative Genomics
- Machine Learning in Biology
Background:
- Variant calling tools often assume human genome parameters, limiting their effectiveness in diverse species.
- DeepVariant (DV) shows promise for variant calling but its performance on 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 limitations of human-genome-trained DeepVariant models when applied to non-human genomes, specifically bovine.
- To introduce TrioTrain, a novel multi-species DeepVariant model for improved variant calling in diverse species.
- To automate the extension of DV for species lacking Genome In A Bottle (GIAB) resources and mitigate cluster-based barriers.
Main Methods:
- Developed TrioTrain, a multi-species DeepVariant training approach using trio data from cattle, yak, and bison.
- Implemented region shuffling for SLURM-based clusters and removed Mendelian discordant variants to offset imperfect truth labels.
- Trained 30 model iterations across five phases, evaluating performance on human trios (HG002) and bovine F1-hybrid genomes.
Main Results:
- The multi-species DV model achieved a mean SNP F1 score >0.990 on GIAB human trios.
- The phase 4 bovine model identified more variants with a lower Mendelian Inheritance Error (MIE) rate than DeepTrio in HG002.
- Substantial reduction in inheritance errors for bovine F1-hybrid genomes, with a mean MIE rate of 0.03 percent.
Conclusions:
- Multi-species, trio-based training produces a robust variant calling model, demonstrating the limitations of exclusively human-trained models.
- TrioTrain effectively extends DeepVariant for species without GIAB resources, enhancing its applicability in comparative genomics.
- The developed model significantly improves variant calling accuracy and reduces inheritance errors in non-human genomes.
