GC bias and coverage holes

A promoter can look deleted because the polymerase disliked its base composition. GC bias is a property of how the library was made and how the instrument was fed. A coverage hole is an observation: this stretch of the reference has little depth. The observation becomes a copy-number call only after you have asked whether the chemistry would have refused that stretch anyway. Library construction as a whole is described in next-generation sequencing from library to reads. This page is the explainer for the holes that survive an otherwise healthy run.

Polymerases and fragmentation notice base composition

Double-stranded DNA is not copied with equal enthusiasm at every composition. Stretches that are extremely rich in G and C, and stretches that are extremely rich in A and T, are both awkward, for different physical reasons. Very GC-rich duplexes separate reluctantly. Very AT-rich duplexes can be fragile, and some polymerases fall off them. Amplification during library preparation, which many short-read kits still use, magnifies a small preference into a large difference in depth.

Fragmentation adds its own signature. Enzymatic fragmentation and mechanical shearing do not cut every sequence at the same rate. A motif that is rarely cut produces fragments outside the size window you then select, so those fragments vanish before sequencing. The hole in the data can sit on a gene that is very much present in the cell.

The practical principle is modest. Depth is the joint product of biology and chemistry. A copy-number tool that assumes chemistry is flat will report the chemistry.

What you can change, and what you follow

The reagent classes that move GC bias are the library kit as a whole: polymerase, fragmentation method, and whether a PCR step is present. PCR-free library classes exist because omitting amplification reduces one major source of bias. They demand more DNA and they do not repeal every sequence preference. Capture and amplicon kits add another layer: a probe or a primer that dislikes a locus will look exactly like a deletion of that locus.

Follow the manufacturer's protocol for the kit in your hand. Temperatures, cycle counts and input ranges are part of how that kit's bias was characterised. A local "optimisation" that adds cycles to rescue a weak library often deepens GC bias. If you need a different bias profile, change kit class deliberately and document it. Method notes that laboratories publish on protocols.io can show you what others recorded. They are not permission to ignore the insert that came with your reagents.

Equipment is the sequencer plus a fragment-length check. An instrument overview for the short-read class is published by Illumina. Long-read library classes have different compositional failure modes. A hole that is purely an amplification artefact may shrink when the same DNA is sequenced as long molecules, which is one reason to read long-read sequencing for structural variants before you call a missing exon a deletion.

A GC plot is not a copy-number track

Take windows along the reference, or along your targets. For each window, compute GC fraction and mean depth. In a biased library the cloud of points slopes: depth falls as GC moves toward an extreme. That slope is a property of the whole library. A true deletion is local. Its window drops, and neighbouring windows of similar GC do not. Breakpoints, split reads, or a change in insert size may sit at the edges. A bias dip has no breakpoint. It follows composition wherever that composition occurs, including in places you know are diploid.

Copy-number callers that normalise by GC are trying to remove the slope before they look for local drops. The normalisation can succeed, and it can also erase a real event that happens to sit in a GC-rich gene, or invent an event if the model is fit on too little data. Look at the plot yourself for any locus you intend to publish. Compare with a control library made the same way. If the control has the same hole, you are looking at the method.

The reference you plot against must be named. GC content of a window depends on the assembly. Check the locus on Ensembl so you know whether the reference itself has a gap there. A gap is not a deletion in your sample.

When the control library shows the same hole

Branch there. If an unrelated sample, prepared with the same kit, misses the same GC-rich exon, do not report a shared deletion. Change chemistry or validate that exon another way. If the control is covered and the sample is not, and the GC plot of the sample is otherwise flat, a real loss is a fair hypothesis. Confirm it. A small locus can be amplified and read by Sanger, discussed as a method through the Sanger DNA sequencing enquiry reference, when the primers themselves can be designed. A primer that fails on the same GC-rich template is another chemistry result, not a confirmation.

If identity checks fail, stop. A coverage hole across a sex chromosome is sometimes a swapped label, which sample identity mix-ups and plate maps treats as its own subject.

Failure modes that produce holes without a mutation

Too many PCR cycles. A polymerase matched to a different fragment length. Bead cleanups that lose the shortest or the longest fragments, which are not GC-neutral in every genome. Capture probes designed on a different build, so a real exon has no bait. Reads trimmed so aggressively that the remaining insert cannot map, which looks like missing sequence. Alignment settings that refuse to place reads in repeats, so those repeats become holes and then become false deletions.

A copy-number call that ignores the GC plot will publish those holes. Ask for the plot whenever a deletion sits in a composition extreme. The words for depth and breadth, if the report mixes them, are in a glossary of sequencing terms.

Research claims stop at the evidence

A coverage hole is not a pathogenic variant, a diagnosis, or a statement about a breeding value. Even a well-supported deletion in a research genome is a research result until a validated assay says otherwise. This page does not provide that validation. Biosafety and ethics attach to the sample, not to the shape of the coverage track.

Kits that have been warm

Enzymatic fragmentation and some polymerases change behaviour after a warm excursion, and the change can be a worse GC slope rather than a total failure. In a hot laboratory, record how library reagents were stored and whether a shipment thawed. A power cut during a PCR step is a reason to abandon that library, not to add cycles and hope. Follow the manufacturer on whether a kit tolerates a stated temperature range. Do not guess a rescue programme.

What to ask for when a locus keeps disappearing

Tell the person you are writing to which locus, what its GC content is on your named build, and whether a control library made the same way shows the same hole. Say whether you already used a PCR-free class. Reagents for an alternative library class live in the genomics and sequencing catalogue, as categories rather than as a promise about a given lot. The work sits on the genomics research pathway.

A genome-wide library strategy can be discussed through the whole-genome sequencing enquiry reference. If the hole is the reason you are considering long reads, use the long-read DNA sequencing enquiry reference. A locus you can still amplify can be discussed through the Sanger DNA sequencing enquiry reference. Send the plot and the kit class with the quote request.