ekg
commented
6 years ago This seems to me like a good plan, and it is similar to what I am exploring in various contexts, although with a different assembly input. Typically I am using assembly graphs from pacbio or ONT data.
I would add several possible alternative approaches that avoid the potential problems with BAM export, which is provided for use in cases where the graph is generally partially ordered, and is likely to require further development to work in a reasonable way in a generic graph such as one made by cactus.
First, I would do single-sample calling using vg call, then add the results back to the reference graph and iterate the process. The goal would be to generate a pangenome object that encodes all of the variation and sequence of interest in the entire set. At the end, another round with vg call would generate a per-sample graph for each sample, and we would have minimized bias with respect to the alleles in the pangenome. It is possible (although not well explored) to then convert this into a vector representation, and of course a matrix can be built up from all the samples (vg pack). At this point we would have something very similar to a multisample VCF, but over a graph. This matrix could then be used directly for various analyses, such as PCA or GWAS-like experiments. Retaining coverages in some normalized way would allow the exploration of copy number variation alongside other kinds of variation. This is admittedly extremely exploratory, but I think it's interesting and I'll be working through it in other contexts.
You can also avoid the BAM step entirely, and project your results out to VCF. That's pretty well supported and avoids the issues with the BAM step.
WimSpee
mdkeehan
jelber2
glennhickey
ac2278
Dear Variant Graph (VG) developers,
Thank you for the software and the recent preprint which I read with great interest. https://www.biorxiv.org/content/early/2017/12/15/234856
One important usecase for graph references is plant and animal breeding. In this context whole genomic segments are often introgressed from wild to cultivated species. Reference bias is thus a big issue.
In for example tomato breeding many genomic segments are intogressed from wild species like Solanum pennellii to the cultivated Solanum lycopersicum tomato species.
See for example for some background information:
The genome of the stress-tolerant wild tomato species Solanum pennellii https://www.nature.com/articles/ng.3046
The tomato genome is diploid and has c.a. 1/3 the size of human genome, c.a. 1 GigaBase.
How would you recommend that VG can be used in the case where one has at least 1 wild and 1 cultivated reference genome, and hundreds to many thousands of whole genome re-sequenced samples?
Creating the base reference genome graph from 1 linear reference genome and a multi-sample VCF file seems strange to me. This would just introduce the reference bias to the graph that we are trying to avoid.
Do I understand it correct that: 1) I could use Cactus to create a pangenome reference graph of all my cultivated and wild reference genomes? (https://github.com/glennhickey/progressiveCactus ? ) 2) I could then align all re-sequenced individuals against the pan genome graph using VG? 3) I could then export per sample BAM and 1 multi-sample VCF file for each input reference genome using VG? ( i.e. a projection of the reads and variants against each input linear reference)
Thank you.
Wim Spee