glennhickey
commented
1 year ago Cactus is aligning multiple tandem repeats together as paralogs. This touches on two issues:
1) If multiple copies are in every genome, then Cactus should leave them apart. It tries to do this already using outgroup information, but it is not perfect. This is something we are aware of and are working on (but it's tricky in the general case).
2) Even if tandem repeats are collapsed, MAF export should have a way to sensibly chain them if possible. The present logic is lacking in this department. I fixed a very similar issue here for the HAL browser. Something similar in hal2maf or taffy would be useful.
Things you can try now, as I unfortunately can't give a timeline for fixing the above issues:
-
use
--dupeMode singleor at the very least--filterGapCausingDupeswithhal2maf. The latter will prune some problem paralogs, leading to much cleaner alignments. The former will guarantee you at most one copy per line. -
try using
--rootto specify a subtree forcactus-hal2mafwith the hope that an outgroup will help keep paralogs apart. -
try
halSyntenyand work with the alignment chains (probably far from ideal). -
increasing the chain size for cactus. For example setting
deannealingRounds="2 32 512 2048"in the config.xml. This is finicky though and you may get substantial tradeoffs in coverage for more contiguous alignments.
sivico26
Hi again,
As promised, continuing with the discussion started in #1201,
cactus(ortaffy?) sometimes breaks some sequences within the same alignment block for no reason, making it look like the affected taxa have several paralogs/duplicates. A closer inspection reveals that it is the same locus but fragmented. Take for instance the following alignment:Here,
ccarandcdanhave multiple duplicates, but if you look closely, some of these sequences are contiguous. Take, for instance,ccar.contig_11: on the third row, if you sum its start position and its size, it gives the start position of the previous row (49710 + 297 = 50007). The same applies to that row: 50007 + 88 = 50095. They are also in the same strand. And finally, if you look at the alignment block, they go one after the other without overlapping. In short, they are contiguous.So, correcting these "broken" sequences, is what I referred to intra-merging in the previous issue. After running intra-merging, this alignment block looks like this:
After doing this, we went from what seemed to be 4 paralogs for
ccarand 3 forcdanto the real 2 for each on this alignment block. Also, the alignment block is now more coherent and homogeneous. This is what I meant when I said thatcactus/taffyshould not be doing this. Now, what is going wrong exactly, I am not sure. I can only imagine some heuristics are failing.Finally, this issue is quite pervasive in my data. To give you an idea, from the same loci I discussed in the previous issue. My raw alignments go from:
Before intra merging, to:
So I reduce the alignments with duplicates from more than 1/3 to ~ 1/5, just by doing intra-merging. To be fair, I am running a slightly more sophisticated version, where I also introduce gaps if the sequence positions do not match exactly but they are close enough (say 10 bp), so these numbers are rather a ceiling. Still, I can assure you that in most cases I do not have to introduce gaps because the matches are exact, so the numbers are not far off either.
Anyway, I hope this gives a better idea of what I meant. Cheers
Originally posted by @sivico26 in https://github.com/ComparativeGenomicsToolkit/cactus/issues/1201#issuecomment-1779156228