Velvet: choice of hash length k

[wclee47]

Hi,
In choosing the hash length k, I found some tip from the Velvet manual.

It says the relation between k-mer coverage C(k) and standard (nucleotide-wise) coverage C: C(k) = C*(L-k+1)/L where k is hash length, and L is read length. It recommends that C(k) should be above 10 to start getting decent results.

For test, I wanted to use 1 lane SOLEXA result whose read length is 101 and total read count is 69,084,522 which is corresponding to standard (nucleotide-wise) coverage C, 2.46.

With the C(2.46), L(101), I wanted to find the hash length k when setting C(k) to 10. However, calculated k is about -308 which is strange negative number.

I've also seen that if I increase 'C' value, I can obtain the reasonable 'k' value. e.g) when 'C' is 14.23, 'k' can be about 31.
Does it mean that I should increase 'C' value to get the reasonable hash length k?
To do so, I think I should pool multiple lanes to increase 'C' value.
However, it may cause memory problem.

Is my approach correct? or is there anybody who has a different idea?
Please let me know. Thanks in advance.

Won-Chul.

[Zigster]

C(k) simply refers to "kmer-coverage" or coverage depth measured in read-kmers per contig kmer.

If you coverage is 2x, your read length is 101, and your kmer is 31 then your kmer-coverage is 1.4
C(k) = C*(L-k+1)/L

You should choose kmer based on specificity/sensitivity judgements.

[wclee47]

Thanks to your answer.

As you said, kmer-coverage is 1.4 with the conditions I have.
However, Velvet manual says that kmer-coverage should be above 10 to start getting decent results.
If so, I think increasing 'C' is the only way to get kmer-coverage above 10 if I do not want to change other values such as 'L' and 'k'.
Is that right?

Won-Chul

[Zigster]

Ye, of course the only way to actually get deeper coverage is to generate more sequence. Assemblies improve rapidly as you go from 10x to 20x to 30x and slowly from there on.

[cavefish]

Hi,
In choosing the hash length k, I found some tip from the Velvet manual.
Won-Chul.

BTW, you want de novo assemble genome? Your standard (nucleotide-wise) coverage C was so low, 2.46. Big genome?