When can higher sequencing depth fail to improve the reliability of a biological conclusion?

Why

Deeper sequencing reduces random sampling error, so a 50x site has a more precise allele fraction than a 10x site. Many failures are not random. Daley and Smith show that in a low-complexity library, extra reads are mostly copies of molecules already seen.

GIAB's stratifications resource notes that no pipeline performs equally across the genome, with large duplications, repeats and homopolymers among the hardest contexts, and that short reads lack the mapping advantage of long reads in large repeats. In those regions, more short reads add more ambiguous alignments, not more truth. The same applies to errors fixed in the sample before sequencing, such as contamination with another person's DNA, oxidative damage or a PCR error early in amplification, which deeper reads faithfully copy.

Analysis errors behave the same way: the wrong genome build or a missing decoy places reads wrongly at any depth. And for biological claims, depth cannot replace replicates. A thousand reads from one mouse are still one mouse.

How to work it through

1. Identify whether the uncertainty is about sampling or about a systematic source.
2. Check library complexity and project what more reads would add.
3. Check whether the problem region is low mappability, a duplication or a homopolymer using GIAB stratifications.
4. Look for contamination, damage signatures and reference mismatches.
5. If the issue is systematic, choose a different method, such as long reads, a PCR-free library, an orthogonal assay or more biological replicates.
6. Write down why depth was or was not the answer.

Where the answer stops holding

In some cases depth does help, for example for low-fraction somatic variants when molecules are not limiting. Long reads and orthogonal assays have their own error profiles. Stratifications describe the human reference genomes, so other organisms need their own difficult-region maps.