OK, so there are over 100 different read mappers out there. Why another one? First, its faster than all the rest on long noisy reads while being equally sensitive to the best, second, because it uses a chain model of alignments that accounts for and reveals the frequent drop outs in quality within a Pacbio read, and third, because it actually estimates and reports those regions of reads that involve repeat elements and their approximate copy number in the genome to which they are being mapped. Finally, I was just curious as to whether the daligner‘s sort and merge k-mer strategy would be competitive with the prevailing paradigm of using FM-indices or Suffix tree/arrays. The answer is that it is, albeit it depends on M, the amount of memory available, and G, the size of the reference genome. The larger G, the slower damapper becomes relative to index-based mappers. Morever, while damapper will run in any reasonable (e.g. 16Gb) amount of memory, it runs faster the more memory there is available. For example, with 48Gb, damapper is 1.8 times faster than BWA on the human genome, 15 times faster on the fly genome, and 36 times faster on E.Coli. Performance will be reported on more extensively in a subsequent post.
damapper REF DB.1 DB.2” will map the reads in blocks 1 and 2 of DB to a reference genome REF. It really doesn’t care if REF is a .dam or a .db, and it will handle whatever size it happens to be in the amount of memory available or specified with the -M option. On the other hand the DB blocks must contain sequences that are not too long, i.e reads only, and is expected to be of an appropriate block size (e.g. 250Mbp). The difference between damapper and daligner is that damapper is looking for the best or near best map of the read to a location in REF, whereas daligner will find every possible local alignment between the reads in DB.1 and DB.2 including all repeat induced LA’s. Thus damapper has much less work to do and so can be and is much faster.
Most of the parameters to damapper are the same as for the daligner, i.e. -v, -b, -k, -t, -M, -e, -s, and -m, but with a much larger default of -k (20) and a higher correlation for -e (.85), and the -w, -h, and -l parameters are not needed. Like daligner, damapper finds local alignments consistent with the error rate -e, and does not extend these alignments into regions where the sequences are no longer reasonably correlating at the specified rate. This means that if a Pacbio read has, say, two very low quality regions internal to the read, a frequent occurrence in such data, then damapper will find and report a chain of 3 local alignments and not a single alignment that is basically nonsense in the two drop out regions. This is distinctly different from other mappers that force an alignment through these regions. This seems to my mind a more principled approach and identifies the fact that parts of a read are very low quality. I’m not sure why seeing 2,000 garbage bases in a forced alignment to the reference sequence is valuable :-).
So a damapper mapping of a read is a chain of one or more local alignments to a corresponding contig of the reference. Occasionally reads are chimers, in which case one wants to see a mapping of each of the two distinct parts of the read. damapper accomplishes this by finding the best chain mapping for the read, then finding the best mapping to any portion of the read not spanned by the first chain, and so on, until all significant portions of the read are mapped (if possible). Moreover, damapper will report other good matches to a segment if requested with the -n option. For example, with -n.95, it reports all matches whose score is within 95% of the best score covering a given portion of a read.
Apart from the -n option, the other options unique to damapper are the -C, -N, and -p flags. By default, damapper returns .las files where the A-reads are the mapped reads, and the B-reads are the contigs of the reference. That is, “
damapper REF DB.1 DB.2” will output DB.1.REF.M1.las, DB.1.REF.M2.las, … DB.2.REF.Mn.las, where T is the number of threads it is run with (specified by the -T option, default 4). The T files for each block can be combined into a single .las file with for example “
LAcat DB.1.REF.M# >DB.REF.1.las“. If you would like to see the mapping from the point of view of the reference contigs, specifying the -C option, will further produce files of the form REF.DB.1.C1.las, … REF.DB.2.Cn.las, where the A-reads are the contigs of the reference, and the B-reads are the mapped reads. Specifying -N suppresses the production of the .M#.las files in case you only want to see the coverage of the contigs by reads.
As is customary, there is an HPC script generator, HPC.damapper, that will generate all the calls necessary to produce the mapping of reads to contigs for blocks of a data set (and vice versa if -C is set). For example, “
HPC.damapper -C -n.95 REF DB” will produce DB.#.REF.las and REF.DB.#.las where DB.# is a block of DB. For example, if DB has 4 blocks then HPC.damapper generates a script equivalent to the following:
# Damapper jobs (1)
damapper -v -C -n0.95 REF DB.1 DB.2 DB.3 DB.4
# Catenate jobs (8)
LAsort -a DB.1.REF.M*.las && LAcat DB.1.REF.M#.S >DB.1.REF.las
LAsort -a DB.2.REF.M*.las && LAcat DB.2.REF.M#.S >DB.2.REF.las
LAsort -a DB.3.REF.M*.las && LAcat DB.3.REF.M#.S >DB.3.REF.las
LAsort -a DB.4.REF.M*.las && LAcat DB.4.REF.M#.S >DB.1.REF.las
LAsort -a REF.DB.1.C*.las && LAmerge -a REF.DB.1 REF.DB.1.C*.S.las
LAsort -a REF.DB.2.C*.las && LAmerge -a REF.DB.2 REF.DB.2.C*.S.las
LAsort -a REF.DB.3.C*.las && LAmerge -a REF.DB.3 REF.DB.3.C*.S.las
LAsort -a REF.DB.4.C*.las && LAmerge -a REF.DB.4 REF.DB.4.C*.S.las
# Check all .las files (optional but recommended)
LAcheck -vS DB REF DB.1.REF DB.2.REF DB.3.REF DB.4.REF
LAcheck -vS REF DB REF.DB.1 REF.DB.2 REF.DB.3 REF.DB.4
# Cleanup all intermediate .las files
rm DB.*.REF.*.las REF.DB.*.*.las
Its important to note that all the commands that deal with .las files, i.e. LAsort, LAmerge, LAcheck, LAshow, and LAdump have been upgraded to properly handle damapper-produced .las files that encode chains of LAs. Internally, Dazzler software can distinguish between .las files that encode chains and those that don’t (e.g. as produced by daligner or datander). For sorting, chains are treated as a unit, and sorted on the basis of the first LA. Display commands will indicate chains when present, and LAcheck checks chains and their internal encoding for validity. Another thing to note very carefully, is that chains are best sorted with the -a option and not the default. This is particularly important for files produced in response to the -C option. For example, to produce a mapping of all reads to REF in say REF.DB.las, one should merge the relevant files above with “
LAmerge -va REF DB REF.DB.*.las” One can then see how the reference sequence is covered by the database by then opening REF.DB.las in DaViewer.
In developing the chaining algorithm for damapper, it became obvious that at little additional overhead one could estimate the repetitiveness of the sequence in a region of a read simply by counting how many (sub-optimal) chains covered the region. The -p option requests that damapper produce a repeat profile track for each read based on these counts which roughly estimate the repetitiveness of a read segment within the underlying genome. That is, for each trace-point sized interval (established by the -s parameter) the number of times, c, this interval is involved in a distinct alignment to the reference is estimated. 0 is recorded for segments that don’t match anything in the reference, 1 for segments that match uniquely, and otherwise floor(log10c/10) up to a cap of 40 corresponding to 10,000 copies. Observe carefully that the track has the same form as intrinsic quality values: a vector of small byte-sized integers for each trace point interval. These vectors can be output by DBdump with the -p option and DaViewer is able to graphically display them. These profiles should make it obvious when a read does not have a unique location in a reference sequence due to its being entirely or almost entirely repetitive. I think having this information is critical in any subsequent analysis.