Detecting and Phasing Small Variants with Highly Accurate Long Reads

We summarize the challenges around small variant detection for highly accurate (>=99%) long reads and present workflow solutions using existing tools (GATK) and new tools (DeepVariant with trained CCS model). Presented at PacBio SMRT Informatics Workshop in San Diego.

Transcript

1. For Research Use Only. Not for use in diagnostic procedures. © Copyright 2019 by Pacific Biosciences of California, Inc. All rights reserved.
   Detecting and Phasing Small Variants with
   Highly Accurate Long Reads
   William Rowell, Senior Scientist, Bioinformatics Applications, PacBio
   SMRT Informatics Developers Conference, January 16, 2019 @nothingclever
   

   View Slide

2. AGENDA
   -Differences between highly accurate long
   reads and short reads
   -Calling variants with existing tools
   -Training new tools on long reads
   -Making use of phase information to
   improve variant calls
   

   View Slide

3. AGENDA
   -Differences between highly accurate long
   reads and short reads
   -Calling variants with existing tools
   -Training new tools on long reads
   -Making use of phase information to
   improve variant calls
   

   View Slide

4. CCS READS HAVE A DIFFERENT ERROR PROFILE FROM
   SHORT READS
   

   View Slide

5. CCS READS HAVE A DIFFERENT ERROR PROFILE FROM
   SHORT READS
   

   View Slide

6. SMALL VARIANT DETECTION WITH HIGH FIDELITY READS
   STARTS WITH A FAMILIAR WORKFLOW
   CCS Reads
   Align with minimap2
   Detect variants with
   GATK HaplotypeCaller
   Hard filter variants with
   GATK VariantFiltration
   Diploid variant calls
   

   View Slide

7. SMALL VARIANT DETECTION WITH HIGH FIDELITY READS
   FOLLOWS A FAMILIAR WORKFLOW
   CCS Reads
   Align with minimap2
   (pbmm2)
   Detect variants with
   GATK HaplotypeCaller
   Hard filter variants with
   GATK VariantFiltration
   Diploid variant calls
   pbmm2 --preset CCS
   

   View Slide

8. SMALL VARIANT DETECTION WITH HIGH FIDELITY READS
   FOLLOWS A FAMILIAR WORKFLOW
   CCS Reads
   Align with minimap2
   (pbmm2)
   Detect variants with
   GATK HaplotypeCaller
   Hard filter variants with
   GATK VariantFiltration
   Diploid variant calls
   pbmm2 --preset CCS
   --pcr-indel-model AGGRESSIVE
   

   View Slide

9. SMALL VARIANT DETECTION WITH HIGH FIDELITY READS
   FOLLOWS A FAMILIAR WORKFLOW
   CCS Reads
   Align with minimap2
   (pbmm2)
   Detect variants with
   GATK HaplotypeCaller
   Hard filter variants with
   GATK VariantFiltration
   Diploid variant calls
   Strand Bias tests ❌
   Mapping Quality tests ❌
   Read position tests ❌
   Variant Quality tests ✅
   

   View Slide

10. SMALL VARIANT DETECTION WITH HIGH FIDELITY READS
    FOLLOWS A FAMILIAR WORKFLOW
    CCS Reads
    Align with minimap2
    (pbmm2)
    Detect variants with
    GATK HaplotypeCaller
    Hard filter variants with
    GATK VariantFiltration
    Diploid variant calls
    

    View Slide

11. SMALL VARIANT DETECTION WITH HIGH FIDELITY READS
    FOLLOWS A FAMILIAR WORKFLOW
    CCS Reads
    Align with minimap2
    (pbmm2)
    Detect variants with
    GATK HaplotypeCaller
    Hard filter variants with
    GATK VariantFiltration
    Diploid variant calls
    

    View Slide

12. SMALL VARIANT DETECTION WITH HIGH FIDELITY READS
    FOLLOWS A FAMILIAR WORKFLOW
    CCS Reads
    Align with minimap2
    (pbmm2)
    Detect variants with
    GATK HaplotypeCaller
    Hard filter variants with
    GATK VariantFiltration
    Diploid variant calls
    pbmm2 --preset CCS
    --pcr-indel-model AGGRESSIVE
    SNV → QD >= 2.0
    1 bp Indels → QD >= 5.0
    >1 bp Indels → QD >= 2.0
    Precision Recall F1
    SNVs 99.468% 99.559% 99.513%
    Indels 78.977% 81.248% 80.097%
    

    View Slide

13. GOOGLE’S DEEPVARIANT CAN BE TRAINED ON NEW DATA
    TYPES
    CCS Reads (chr. 1-19)
    + GIAB Truth Set
    DeepVariant CNN
    training
    CCS model
    Nature Biotechnology volume 36, pages 983–987 (2018)
    

    View Slide

14. GOOGLE’S DEEPVARIANT CAN BE TRAINED ON NEW DATA
    TYPES
    CCS Reads (chr. 1-19)
    + GIAB Truth Set
    DeepVariant CNN
    training
    CCS model
    Nature Biotechnology volume 36, pages 983–987 (2018)
    CCS Reads
    DeepVariant
    + CCS model
    Diploid
    variant calls
    

    View Slide

15. GOOGLE’S DEEPVARIANT CAN BE TRAINED ON NEW DATA
    TYPES
    CCS Reads (chr. 1-19)
    + GIAB Truth Set
    DeepVariant CNN
    training
    CCS model
    Nature Biotechnology volume 36, pages 983–987 (2018)
    CCS Reads
    DeepVariant
    + CCS model
    Diploid
    variant calls
    Precision Recall F1
    SNVs 99.914% 99.959% 99.936%
    Indels 96.901% 95.980% 96.438%
    autosomes
    

    View Slide

16. GOOGLE’S DEEPVARIANT CAN BE TRAINED ON NEW DATA
    TYPES
    CCS Reads (chr. 1-19)
    + GIAB Truth Set
    DeepVariant CNN
    training
    CCS model
    Poplin, R. et al. A universal SNP and small-indel variant caller using deep neural networks. Nature Biotechnology 36, 983–987 (2018)
    CCS Reads
    DeepVariant
    + CCS model
    Diploid
    variant calls
    Precision Recall F1
    SNVs 99.914% 99.959% 99.936%
    Indels 96.901% 95.980% 96.438%
    Precision Recall F1
    SNVs 99.807% 99.904% 99.855%
    Indels 95.387% 94.501% 94.942%
    autosomes
    chromosome 20
    

    View Slide

17. BUT WE’RE STILL LEAVING INFORMATION ON THE TABLE…
    

    View Slide

18. BUT WE’RE STILL LEAVING INFORMATION ON THE TABLE…
    

    View Slide

19. PHASE INFORMATION CAN IMPROVE PRECISION AND RECALL
    

    View Slide

20. INCORPORATING PHASE INFORMATION LEADS TO
    IMPROVEMENTS IN VARIANT RECALL AND PRECISION
    Ebler, J. et al. Haplotype-aware genotyping from noisy long reads bioRxiv doi: 10.1101/293944
    Precision Recall F1
    SNVs 99.468% 99.559% 99.513%
    Indels 78.977% 81.248% 80.097%
    Precision Recall F1
    SNVs 99.693% 99.792% 99.742%
    Indels 81.102% 83.818% 82.438%
    GATK HC
    GATK HC + WhatsHap
    Incorporate phase
    information and
    re-genotype
    variant positions
    with WhatsHap
    

    View Slide

21. INCORPORATING PHASE INFORMATION LEADS TO MODEST
    INCREASES IN VARIANT RECALL AND PRECISION
    Precision Recall F1
    SNVs 99.468% 99.559% 99.513%
    Indels 78.977% 81.248% 80.097%
    Precision Recall F1
    SNVs 99.693% 99.792% 99.742%
    Indels 81.102% 83.818% 82.438%
    Precision Recall F1
    SNVs 99.914% 99.959% 99.936%
    Indels 96.901% 95.980% 96.438%
    Precision Recall F1
    SNVs 99.904% 99.963% 99.934%
    Indels 97.835% 97.141% 97.486%
    GATK HC
    GATK HC + WhatsHap
    DeepVariant CCS
    DeepVariant CCS + haplotype sorting
    Incorporate phase
    information and
    re-genotype
    variant positions
    with WhatsHap
    Tag reads with
    phase information
    from trio data and
    sort alignments by
    haplotype
    https://goo.gl/4cnMeC
    Ebler, J. et al. Haplotype-aware
    genotyping from noisy long reads bioRxiv
    doi: 10.1101/293944
    

    View Slide

22. CONCLUSIONS
    -GATK HaplotypeCaller (optimized for short reads) can be used to detect
    SNVs with high recall and precision, but has trouble discriminating between
    biological indels and sequencing errors.
    -DeepVariant, when trained on long reads, can be used to detect both SNVs
    and indels with high recall and precision.
    -Both workflows can be improved by providing long-distance phasing
    information, but there’s still work to be done in this area.
    

    View Slide

23. ACKNOWLEDGEMENTS
    Google AI Genomics
    Alexey Kolesnikov
    Pi-Chuan Chang
    Andrew Carroll
    Mark DePristo
    Saarland University/Max Planck
    Institute for Informatics
    Jana Ebler
    Tobias Marschall
    PacBio
    Yufeng Qian
    Richard Hall
    Aaron Wenger
    Paul Peluso
    David Rank
    Mike Hunkapiller
    DNANexus
    Jason Chin
    NIST
    Nathan Olson
    Justin Zook
    

    View Slide

24. For Research Use Only. Not for use in diagnostic procedures. © Copyright 2019 by Pacific Biosciences of California, Inc. All rights reserved. Pacific Biosciences, the Pacific Biosciences logo, PacBio,
    SMRT, SMRTbell, Iso-Seq, and Sequel are trademarks of Pacific Biosciences. BluePippin and SageELF are trademarks of Sage Science. NGS-go and NGSengine are trademarks of GenDx. FEMTO
    Pulse and Fragment Analyzer are trademarks of Advanced Analytical Technologies.
    All other trademarks are the sole property of their respective owners.
    www.pacb.com
    

    View Slide