Note: the database used must be the same as the one used in the original Kraken run
What it does
At present, we have not yet developed a confidence score with a solid probabilistic interpretation for Kraken. However, we have developed a simple scoring scheme that has yielded good results for us, and we've made that available in the kraken-filter script. The approach we use allows a user to specify a threshold score in the [0,1] interval; the kraken-filter script then will adjust labels up the tree until the label's score (described below) meets or exceeds that threshold. If a label at the root of the taxonomic tree would not have a score exceeding the threshold, the sequence is called unclassified by kraken-filter.
A sequence label's score is a fraction C/Q, where C is the number of k-mers mapped to LCA values in the clade rooted at the label, and Q is the number of k-mers in the sequence that lack an ambiguous nucleotide (i.e., they were queried against the database). Consider the example of the LCA mappings in Kraken's output:
562:13 561:4 A:31 0:1 562:3
would indicate that:
the first 13 k-mers mapped to taxonomy ID #562 the next 4 k-mers mapped to taxonomy ID #561 the next 31 k-mers contained an ambiguous nucleotide the next k-mer was not in the database the last 3 k-mers mapped to taxonomy ID #562
In this case, ID #561 is the parent node of #562. Here, a label of #562 for this sequence would have a score of C/Q = (13+3)/(13+4+1+3) = 16/21. A label of #561 would have a score of C/Q = (13+4+3)/(13+4+1+3) = 20/21. If a user specified a threshold over 16/21, kraken-filter would adjust the original label from #562 to #561; if the threshold was greater than 20/21, the sequence would become unclassified.