DRAGEN–GATK

GATK and DRAGEN: differences discussed in the training

  • GATK single sample pipeline: the goal is to do the best possible single sample calling on a human genome
  • DRAGEN: similar to GATK best practice but what’s new →
    • Train the PCR model
Train the PCR model
Train the PCR model
Train the PCR model
Train the PCR model

GATK vs. DRAGEN-GATK

  • Aligner → DRAGMAP

  • No BQSR step → the training describes a different error-modeling strategy

  • Variant discovery adds an STR error-modeling step while retaining HaplotypeCaller, with three improvements shown below:

    Variant discovery adds an STR error-modeling step while retaining HaplotypeCaller, with three improvements shown below:
    Variant discovery adds an STR error-modeling step while retaining HaplotypeCaller, with three improvements shown below:

STR (short tandem repeats) errors occur in both aligning and sequencing

  • Hard to align
  • PCR often adds/deletes STR units
  • Sequencers struggle to sequence bases around STR units in the genome

STR model (CallibrateDragstrModel)

  1. Find all of these STR in your reference genome
  2. Assumption: We don’t know if this error is an PCR error or an indel. But let’s assume that most of it are PCR errors.
  3. Try to calculate the baseline error rate as a prior → If this region has high baseline error rate, we will be less confident for certain variant call in this region. → We get much better indels (60% → 90% accuracy).

BQD (Base quality dropout) + FRD (Foreign read detection)

  • It characterize the likelihood of getting a wrong call
  • For example if we get A/C, it gives us a model to say how likelihood A or C are actually errors (how likelihood we should get C/C or A/A) An A/C genotype may initially appear most likely; treating low-confidence alternative reads as errors can instead make A/A more plausible.
An A/C genotype may initially appear most likely; treating low-confidence alternative reads as errors can instead make A/A more plausible.
An A/C genotype may initially appear most likely; treating low-confidence alternative reads as errors can instead make A/A more plausible.

BQD

  • Assumption: Errors are often non-independent →

    • Sequencing quality fails at the end of the read
    • Certain reference regions cause base quality errors
    • Penalizes alleles with lower quality reads

  • For the variant highlighted in green, inspect the quality of the supporting reads rather than treating each apparent call independently. The read-level evidence changes confidence in the variant.

    BQD
    BQD

FRD

  • catch the call from the reads that don’t belong → might be contamination or something

  • In the example with a 20:4 imbalance, the less frequent GG-supporting reads may have poor quality. They may be foreign reads rather than evidence for a true genotype.

    In the example with a 20:4 imbalance, the less frequent GG-supporting reads may have poor quality. They may be foreign reads rather than evidence for a true genotype.
    In the example with a 20:4 imbalance, the less frequent GG-supporting reads may have poor quality. They may be foreign reads rather than evidence for a true genotype.

Long-read sequencing

  • PacBio vs. ONT
  • Long reads can span structural variants that exceed the length covered by individual short reads.
  • Some non-unique repetitive regions are difficult to resolve with short reads.
  • Long-read workflows can also provide DNA-methylation information. Short-read methylation workflows typically require additional library preparation, which introduces its own effects on the data.
← Back to all posts