Corrects the overlaps in DNA sequence graphs by realigning sequence ends.
Optionally outputs FASTA and PAF representing the GFAv1-format graph.
Almost no overlap-based assembly methods produce correct GFAv1 output. Invariably, some overlaps are incorrectly defined. (The major exception to this are De Bruijn assemblers, which have fixed length overlaps that are correct by definition.) In some methods (like shasta), the overlaps are systematically slightly wrong, due to their derivation from run length encoded sequences. It can help to correct these, because it lets us "bluntify" the graph. This produces a graph in which each base in the graph exists on only one node, which is a desirable property for variation graph and other pangenomic reference models.
To build gimbricate, use git to download its source and cmake to build.
You'll need a recent gcc >= v7.4 as well as cmake >= v3.1
git clone --recursive https://github.com/ekg/gimbricate.git
cd gimbricate
cmake -H. -Bbuild && cmake --build build -- -j 4
By default, it builds into bin/ in the directory root.
gimbricate reads a GFA and uses the CIGAR strings attached to L records to guide realignment of the ends of the sequences and the subsequent correction of the CIGAR strings on the overlaps. It takes the sum of CIGAR operation lengths in the cigar to determine how much sequence to realign. Then it realigns the appropriate ends of the Sequence records referred to by the Link record with GSSW.
For instance, these link records might be approximately accurate:
L 256 + 1072 + 11M L 1072 + 1074 + 20M L 110 + 1072 - 20M L 1072 - 1074 - 20M
But realignment suggests that only 11bp of the last one match:
L 256 + 1072 + 11M L 1072 + 1074 + 20M L 110 + 1072 - 20M L 1072 - 1074 - 9D11M9I
gimbricate can also convert its GFA output to a combination of FASTA and PAF.
These may then be fed into seqwish, which induces a variation graph including the original contigs as paths.
By adding additional alignments (in PAF format) to its input, seqwish can further compress redundant regions of the graph, but without loss of the original graph structure (which is given by the FASTA and PAF outputs of gimbricate).
By example (using test/h.gfa), this would bluntify the graph in h.gfa:
gimbricate -g h.gfa -n -p h.paf -f h.fasta >h.gimbry.gfa
seqwish -s h.fasta -p h.paf -g h.seqwish.gfa
Optional flag -P can be used when trying to bluntify a graph with already correct overlaps (such as those outputted by De Bruijn assemblers), skipping the time- and memory-consuming process of realignment.
This would additionally compress the graph by adding alignments between the nodes:
gimbricate -g h.gfa -n -p h.paf -f h.fasta >h.gimbry.gfa
minimap2 -cx asm20 -k 11 -w 1 -X h.fasta h.fasta >h.self.paf
cat h.paf h.self.paf >h.self+.paf
seqwish -s h.fasta -p h.self+.paf -g h.seqwish+.gfa
Note that this techinque can be extended to provide graph to graph alignment and pangenome assembly.
All that we need to do is to concatenate together the gimbricate PAFs, FASTAs, and minimap2 PAFs for the full collection of graphs, and then feed these to seqwish.
By default, gimbricate renames all input segments to numerical ids from 1 to N, where N = |Sequences|, but this can be disabled with -n.
This can simplify preprocessing when importing the graphs into vg, odgi, or other tools that expect numerical GFA names.
When building pangenomes from many inputs, it's often helpful to name input sequences uniquely.
Add a prefix to the names in the file with -x or --name-prefix.
This will propagate to the PAF, FASTA, and GFA outputs.