What whole-genome sequencing can and cannot read

The phrase “whole genome” carries a suggestion that the whole of the genome has been read. It has not.

Almost all of the whole-genome sequencing sold to individuals today uses short reads. The genome is fragmented, a very large number of short pieces — around 150 bases each — are read, and each piece is matched back against a reference sequence to work out where it came from. The limits fall out of that last step, putting the pieces back.

Draw this line before the reading starts. Leave it vague and you will read “not found” as “not present.”

What it does well

Single-base substitutions, and insertions and deletions from a few bases up to a few dozen. Most of the variants people discuss fall in here, so in practice the coverage is broad.

For working through variants involved in metabolic pathways, this range is enough for the great majority of cases.

What it does poorly

Large structural changes

Deletions and duplications on the scale of hundreds to thousands of bases are hard to catch by laying down short fragments. When a fragment is absent, it is difficult to tell whether it simply was not read or whether it is genuinely missing.

Repeat expansions

Where the same sequence repeats dozens or hundreds of times, there is no way to decide which copy within the repeat a fragment came from. When the number of repeats is itself the meaningful quantity, that region has not been read.

Regions closely similar to others elsewhere

The genome carries sequences that closely resemble functional genes — copies that have lost their function, still sitting there. Fragments cannot be assigned confidently to one or the other, so reads in these regions are less reliable.

Places read thinly

How many times a given position was read is called coverage depth. It varies across the genome. Regions skewed towards particular base compositions tend to come out thin, and calls made in thin places are not dependable.

What “not found” means

This is the point most often misread in practice.

When nothing appears in the results, it is one of the following.

• There was no change there
• There was a change, but of a kind this method cannot read
• The region was never read in sufficient depth

The results table cannot tell these three apart. Telling them apart requires checking how much that region was actually read. In procedural terms: look at the coverage distribution before starting the reading, so you know in advance which regions were not read.

Check the position, not the gene

There is an easy mistake here. “This gene is well covered” is not enough. What you want to know is whether that one position was read.

An average depth across the whole gene will not tell you whether your position happens to sit in a trough. The reverse happens too: a gene that reads thinly overall can be perfectly well covered at the one position you care about. Fix the position, then look at that.

And bear in mind that variants turn up roughly once in a thousand bases. “Nothing was called near that position” therefore does not translate to “the position was not read.” Nothing nearby and nothing read are two different statements.

Some forms need a different instrument

Among the things this method cannot read, a few have a very definite character.

• Whole-gene deletion polymorphisms. Some people carry the gene and some do not. You cannot establish that someone lacks it from the mere absence of aligned fragments (GSTM1 and GSTT1 are the familiar examples).
• Regions where the number of repeats is itself the polymorphism. A short sequence in a regulatory region repeats a variable number of times. Since the count is what matters, a method that lines up short fragments cannot count it (the repeat in the MAOA promoter is one example).

These were not missed. This method cannot read them in principle. Settling them takes a different approach — assays that count copies, or that measure repeat length.

What matters is to write down that something is unreadable when it is. A blank in a table is uncomfortable, and filling it with something turns it into a result for whoever reads the table next. Leave the blank, and note beside it why it is blank and what it would take to settle. That is the honest form, and the one that is still usable later.

For exercises, use data with known answers

To find out whether your procedure is right, you need something to check it against.

That is what the standard datasets published for research are for. Genome in a Bottle, run by NIST in the United States, publishes “answers” validated to a high degree of confidence — the same specimens analysed repeatedly by several methods, with the range of agreement established.

Your own data has no answer key. Which is exactly why the practice should happen on data that does, with your own data coming after the procedure is solid. Keeping that order turns out to be faster in the end.