Sequencing and coverage depth
The captured library is read on an Illumina sequencer as tens of millions of read pairs. The number of independent reads covering an exon base is its depth, and depth sets the lowest VAF you can see. The mean depth is only a starting point: duplicates and uneven coverage both discount the depth that is actually useful.
Paired-end reads and coverage depth
Illumina sequences by synthesis: library molecules are amplified into clusters on a flow cell, and in each cycle fluorescently labelled bases are added and imaged, one base per cycle. Paired-end sequencing reads a stretch from each end of the fragment (WES commonly uses 2×100 or 2×150 bp, depending on the platform), and the distance between the two aligned reads gives the insert size. Every base comes with a base quality, which usually drops towards the end of the read.
Lay the reads on the exons and count, base by base, how many cover it, and you get a depth profile. It is not flat: deeper near the centre of a probe, thinner at the edges, and lower overall in exons with extreme GC. Duplicate reads, poorly mapped reads and low-quality bases are dropped during calling, and the overlap of a read pair is counted once, so the effective depth behind each decision is lower than the raw depth.
Depth sets the lowest VAF you can see
At a site with true VAF f and depth n, the number of reads supporting the variant roughly follows a binomial distribution: n × f on average, but with scatter from one sample to the next. Callers usually require a minimum number of alt reads so that sequencing errors are not taken for mutations. The detection probability is then the chance of drawing at least that many.
Try a few combinations in the schematic above:
- 100×, VAF 5%, at least 5 alt reads: 5 on average, and a detection probability of only about 56%.
- Same VAF at 200×: about 97%.
- 100×, VAF 2%: about 5%; reliable detection needs depth on the order of 500×.
The VAF here already includes purity: in a sample of 30% purity, a clonal heterozygous mutation is expected at about 15%, and subclonal ones lower still. Low-purity samples need deeper sequencing.
Common pitfalls
- Looking only at the mean. A mean of 150× can still leave some exons at 20×; check the ≥20× (and ≥50×, ≥100×) fractions and the coverage distribution.
- Estimating sensitivity from raw depth. The effective depth after removing duplicates and low-quality reads is what the caller actually uses.
- Calling a gene wild-type because no mutation was found. First check that the position was covered deeply enough; a poorly covered position is “unknown”, not wild-type.