Does the DRAGEN pipeline support VQSR (Variant Quality Score Recalibration)?

Illumina recommends running DRAGEN as default, without adding the VQSR step. VQSR is a time-consuming, post-VCF filtering step, which builds a re-calibration model to ultimately filter out variants in the output VCF.

In older DRAGEN versions (and GATK), because of the variant caller’s inability to distinguish between correlated errors and true variants, it was necessary to apply hard-filtering rules post-VCF to filter out the excess number of FP calls. Several VCF annotations could be used (eg, QD, MQ, FS, MQRankSum) to flag calls as FP. Alternatively, those annotations could be fed to a machine learning algorithm and trained against a truth set, and false positives could then be filtered out based on the training (eg, VQSR).

In newer DRAGEN versions, the algorithms were improved at the core of the variant caller by incorporating the properties of base calling and/or mapping errors directly into the probability calculations. Therefore, the dependency on post VCF filtering was substantially reduced. DRAGEN V3 hard-filtering simply uses QUAL with a threshold corresponding to the best Fmeas (best trade-off between sensitivity and specificity). Therefore, to get the best accuracy and run time results, Illumina recommends using the DRAGEN output VCF as is, without adding the VQSR step.

With the release of DRAGEN 3.5, the VQSR module was deprecated and no longer available in DRAGEN 3.5.7 or later versions.

For any feedback or questions regarding this article (Illumina Knowledge Article #6139), contact Illumina Technical Support [e-mail].