I am running Mutect2 (GATK v4.x) on ctDNA sequencing data generated from Illumina NovaSeq X Plus.

NovaSeq X Plus produces discrete base quality bins (Q9, Q24, Q40). With the default Mutect2 setting (--min-base-quality-score 10), only Q24 and Q40 bases are used. When I lower this threshold to include Q9 bases, some low-VAF variant calls show noticeable changes in detection or TLOD scores.

This behavior is not observed in NovaSeq 6000 data, where base qualities are typically Q11/Q25/Q37.

My specific questions are:

Is the observed change in variant calls mainly due to how Mutect2 models base quality likelihoods, or are there known NovaSeq X Plus–specific error patterns associated with Q9 bases?

In practice, are Q9 bases generally excluded in somatic variant calling to avoid inflating false positives?

Are there recommended additional filters when including low base quality reads (e.g., strand bias, read position, orientation bias)?

Any insights into Mutect2’s handling of low base quality data on NovaSeq X Plus would be appreciated.