nf-core/pangenome

A FASTA file containing the sequences to build the pangenome graph from. Each sequence can be a full chromosome, a contig, or a very long read. The FASTA file must be BGZIPPED or WFMASH won't be able to process it. If you have your sequences in FASTA format, you can run:

bgzip <SEQUENCES.fa> @<THREADS> > <SEQUENCES.fa.gz> samtools faidx <SEQUENCES.fa.gz>

In order to ensure the most compatible functionality, please format your sequence identifiers so that they follow the https://github.com/pangenome/PanSN-spec.

pattern: ^\S+.fn?a(sta)?(.gz)?$

The constructed graph is defined by the number of mappings per segment of each genome (--n_haplotypes <N> - 1). Ideally, you should set this to equal the number of haplotypes in the pangenome. Because that's the maximum number of secondary mappings and alignments that we expect. Keep in mind that the total work of alignment is proportional to N*N, and these multimappings can be highly redundant.

Set this parameter to your e-mail address to get a summary e-mail with details of the run sent to you when the workflow exits. If set in your user config file (~/.nextflow/config) then you don't need to specify this on the command line for every run.

Use mash dist or mash triangle to explore the typical level of divergence between the sequences in your input (see https://pggb.readthedocs.io/en/latest/rst/tutorials/divergence_estimation.html#divergence-estimation for more information). Convert this to an approximate percent identity and provide it as --wfmash_map_pct_id <PCT>. A list of examples can be found at https://github.com/pangenome/pggb#example-builds-for-diverse-species.

Crucially, --wfmash_segment_length provides a kind of minimum alignment length filter. The mashmap3 step in wfmash will only consider segments of this size. For small pangenome graphs, or where there are few repeats, --wfmash_segment_length can be set low (for example 500 when building a MHC pangenome graph). However, for larger contexts, with repeats, it can be very important to set this high (for instance 50k in the case of human genomes). A long segment length ensures that we represent long collinear regions of the input sequences in the structure of the graph. In general, this should at least be larger than transposons and other common repeats in your pangenome. A list of examples can be found at https://github.com/pangenome/pggb#example-builds-for-diverse-species.

By default, wfmash only keeps mappings with at least 5 times the size of a segment. This can be adjusted with --wfmash_block_length <BLOCK_LENGTH>.

This Nextflow pipeline version's major advantage is that it can distribute the usually computationally heavy all versus all alignment step across a whole cluster. It is capable of splitting the initial approximate alignments into problems of equal size. The base-level alignments are then distributed across several processes. Assuming you have a cluster with 10 nodes and you are the only one using it, we would recommend to set --wfmash_chunks 10. If you have a cluster with 20 nodes, but you have to share it with others, maybe setting it to --wfmash_chunks 10 could be a good fit, because then you don't have to wait too long for your jobs to finish.

Graph induction with seqwish often works better when we filter very short matches out of the input alignments. In practice, these often occur in regions of low alignment quality, which are typical of areas with large INDELs and structural variations in the wfmash alignments. This underalignment is then resolved in the smoothxg step. Removing short matches can simplify the graph and remove spurious relationships caused by short repeated homologies. A setting of --seqwish_min_match_length 47 is optimal for around 5% divergence, and we suggest lowering it for higher divergence and increasing it for lower divergence. Values up to --seqwish_min_match_length 311 work well for human haplotypes. In effect, setting --seqwish_min_match_length to N means that we can tolerate a local pairwise difference rate of no more than 1/N. Thus, INDELs which may be represented by complex series of edit operations will be opened into bubbles in the induced graph, and alignment regions with very low identity will be ignored. Using affine-gapped alignment (such as with minimap2) may reduce the impact of this step by representing large indels more precisely in the input alignments. However, it remains important due to local inconsistency in alignments in low-complexity sequence.

If you run out of memory during the seqwish step, you can lower this value. It will take longer, but it will use less memory.

The last step in smoothxg refines the graph by running a partial order alignment (POA) across segments, so called blocks. The "chunked" POA process attempts to build an MSA for each collinear region in the sorted graph. The length of these sub-problems greatly affects the total time and memory requirements of the pipeline, and is defined by -- smoothxg_poa_length <LEN1,LEN2,...>. Several passes of refinement can be defined by lengths >LEN1,LEN2,...>, and so on. Ideally, this target can be set above the length of transposon repeats in the pangenome, and base-level graph quality tends to improve as it is set higher. Higher values makes sense for lower-diversity pangenomes, but can require several GB of RAM per thread.

For details of the different asm modes, please take a look at https://pggb.readthedocs.io/en/latest/rst/optional_parameters.html#homogenizing-and-ordering-the-graph.

The paths matching ^REF are used as a reference, while the sample haplotypes are derived from path names, e.g. when DELIM=# and with '-V chm13:#', a path named HG002#1#ctg would be assigned to sample HG002 phase 1. If LEN is specified and greater than 0, the VCFs are decomposed, filtering sites whose max allele length is greater than LEN.

Pangenome graphs can represent all mutual alignments of collections of sequences. However, we can't really expect to pairwise map all sequences together and obtain well separated connected components. It is likely to get a giant connected component, and probably a few smaller ones, due to incorrect mappings or false homologies. This might unnecessarily increase the computational burden, as well as complicate the downstream analyzes. Therefore, it is recommended to split up the input sequences into communities in order to find the latent structure of their mutual relationship. For example, the communities can represent the different chromosomes of the input genomes. <Warning> If you know in advance that your sequences present particular rearrangements (like rare chromosome translocations), you might consider skipping this step or tuning it accordingly to your biological questions.