Unique K-mers & coverage depth

[sigma]

Hi friends,

I'm working with some reads' simulations from Illumina paired end data and I try to find the link between the coverage depth and unique k-mer.

My hypothesis : with a HIGH coverage depth we must have LESS unique k-mer. Most of nucleotides are covered.
However I see for a 20 mers the complete opposite result :for a high coverage depth I have got a high uniqueness ratio whereas my hypothesis is validated for a 80 mers.

Actually I look for a paper about it but can not find it.
Have you got any idea about my hypothesis or these results ?

advance thanks,

[jimmybee]

This would be completely dependent on the type of organism you're working on wouldn't it?

[sigma]

I don't think so...

K-mer and coverage depth do not depend on the type of organism.
n.b : I am working on maize

[jimmybee]

Thats not what I meant. Nevermind

[arvid]

How do you define uniqueness ratio?

My hypothesis would be that sequencing errors causing unique k-mers will continue to add to the number of unique k-mers for much longer than your actual underlying sequence will (once you've saturated your sequence). If you continue with extreme sequencing depths, all possible errors will have been seen, so this trend will eventually flatten out.

You should probably have a look at how your read simulator generates sequencing errors, at high depths, and check that these are similar to real datasets - I would grab a big public dataset and do the comparison (not sure if there are good maize sets, however). I'd be interested to see whether the errors are as random in real datasets (at extremely high depths, where the underlying libraries are saturated) as with typical simulators.

[sigma]

Hi arvid,

uniqueness ratio = unique k-mers/ total distinct k-mers
And yes unique k-mers result from sequencing errors.

BUT when i simulated data I simulate reads WITHOUT errors (some parameters help to simulate a perfect sequence) so with a hight depth i must find more identical kmers (it depends on the sequence).

[cedance]

based on what distribution do you sample k-mers? gamma?
Also could you attach a plot of your k-mer frequency distribution/histogram?

[arvid]

Hi arvid,

uniqueness ratio = unique k-mers/ total distinct k-mers
And yes unique k-mers result from sequencing errors.

BUT when i simulated data I simulate reads WITHOUT errors (some parameters help to simulate a perfect sequence) so with a hight depth i must find more identical kmers (it depends on the sequence).

If you are simulating reads, why without errors? If you are developing an algorithm or heuristic that should deal with real reads, I don't see the point... because they will behave quite differently.

[cedance]

@arvid, Unless the hypothesis itself is to test the validity of the claim that "unique k-mers are solely attributed to sequencing errors"?

[arvid]

@arvid, Unless the hypothesis itself is to test the validity of the claim that "unique k-mers are solely attributed to sequencing errors"?

Right, but then why use a read simulator? To me it sounds more feasible to dissect the problems into units that can be more easily solved separately:

1. check k-mer distributions in the genome of interest - by simply counting all the k-mers
2. check other biases introduces by the sequencing pipeline, not just substitution or indel errors that leads to skews in these k-mer distributions

With the current approach, problem 1 gets mixed with problem 2. I'm also not sure to what extent a read simulator is feasible for problem 2. Maybe if sigma can elaborate on the actual question he is trying to solve, a better way could be found...