Machine Learning for Variant Calling

DRAGEN secondary analysis employs machine learning based variant recalibration (DRAGEN-ML) for germline SNV VC. Variant calling accuracy is improved using powerful yet efficient machine learning techniques that augment the variant caller, by exploiting more of the available read and context information that does not easily integrate into the Bayesian processing used by the haplotype variant caller. A supervised machine learning method was developed using truth from the PrecisionFDA v4.2.1 sets to build a model that processes read and other contextual evidence to remove false positives, recover false negatives and reduce zygosity errors, for both SNVs and INDELs.

Setup

No additional setup is required. ML model files for the hg38 and hg19 human references are packaged with the DRAGEN installer. After installation, the files are present at <INSTALL_PATH>/resources/ml_model/<ref> DRAGEN-ML is enabled by default as needed, when running the germline SNV VC. DRAGEN will automatically detect the reference used for analysis, and use the correct model files. It either hg38 or hg19 reference type is not detected, ML recalibration will automatically be disabled and SNV VC falls back to legacy operation.

Inputs

DRAGEN-ML requires a run with BAM or FASTQ input, since the machine learning model extracts information from the read pile-up. DRAGEN-ML runs concurrently with DRAGEN SNV VC. DRAGEN-ML can be applied to WGS or WES samples. Re-calibration of existing VCF files is not supported.

Outputs

DRAGEN-ML recalibrates all quality scores, changing the values of the QUAL and GQ fields in the output VCF/GVCF.

• DRAGEN-ML also updates PL and GP in the output VCF/GVCF.

• The genotypes (GT field) of some variants may be changed by ML e.g., 0/1 to 1/1 or vice versa.

• DRAGEN-ML PHRED scores (e.g. QUAL) are better calibrated than and differ significantly from those with ML disabled and, as a consequence, QUAL scores tend to not exceed 75. By comparison, QUAL scores with ML disabled can exceed 1000. For this reason, the QUAL filtering threshold for both SNP and Indel is set to 3.0103 when DRAGEN-ML is enabled, compared to 10.41 of SNP and 7.83 of Indel for DRAGEN-VC when DRAGEN-ML is disabled.

The following variants types are recalibrated:

• Biallelic and multiallelic variants

• Autosomes and sex chromosomes, including haploid positions

• Force GT calls

• Non primary contigs

Accuracy Improvements

DRAGEN-ML typically removes 30-50% of SNP FPs, with smaller gains on INDELS. FN counts are reduced by 10% or more. The output QUAL/GQ of DRAGEN-ML is empirically more accurately calibrated than DRAGEN SNV VC without ML. There are significant gains in accuracy statistics across the entire genome with ML enabled. Note that a small number of variant calls may have degraded accuracy with ML enabled compared to VC without ML.

Run time

DRAGEN-ML adds about 10% to the run time compared to runs without ML.