Dear Community,
I would like to reconstruct the phylogeny of 16,000 E. coli strains based on their genome sequences. What do you think about the following concept?
1. ORF prediction in all genomes. (Prodigal)
2. Identification of the core genes based on reciprocal BLAST of the proteins of the reference E. coli strain to all proteomes. (Diamond)
3. Align the core genes. (MAFFT)
4. Keep SNV sites only and concatenate them to a "super" alignment.
5. Calculate a maximum likelihood tree based on the "super" alignment. (PhyML)
My questions are the following:
A) Should we use only those core genes that we found in all genomes? This is impossible since the overlap contains 0 genes. (We didn't find a single common gene.) Then should I exclude some strains to gain common core genes or can I use genes that are not common in all strains but represented at least in 15,000 strains?
B) Is it a good idea to concatenate SNV sites and infer a tree based on them?
C) How should I choose nucleotide substitution model for such a long and multitaxa alignment?
Cheers,
Eszter
Thanks for the suggestion but this pipeline you suggested must take much longer than ours.
How about maximising the pairwise comparison information like this:
- Extract all potential gene / protein sequences from all strains and make a library of them
- when comparing two genomes find the shared genes from your library and define your genetic distance based on the comparison of those
- you will end up with a 16k x 16k pairwise distance matrix that you can use for any tree (ML, UPGMA, NJ) construction
So, instead of creating multiple alignments, distances will be defined by maximised (in terms of information) pairwise comparisons/distances. Number of shared genes should be somehow taken into account in distance calculation as well.
Thanks for the suggestion, Lajos. It is a good idea but seems to be slower than the one I wrote. Since the Nr. of comparisons in our case is the gene Nr. of the reference genome (approx. 4500) x 16,000. In your case it is much more.
Can we calculate a ML tree based on a distance matrix? How?
OK, UPGMA or NJ but not ML :)
Pairwise alignments are relatively fast, especially compared to a multiple alignment of 16k sequences.
I am not sure, that the approach to concatenate all SNVs from all genes will give you the answer you need. How about doing ML tree for each individual genes and then form clusters of trees that are in agreement and make a consensus tree out of them. I don't know how to do it of course (I mean the second step) but I think it is doable. This way you use all information you have but not restricted by gene set differences. This way you may also find interesting genes / gene clusters that clearly show different evolutionary pattern compared to the others.
Finally I found MMseqs2, an ultra fast and sensitive protein search and clustering suite (https://github.com/soedinglab/MMseqs2), which can predict ortholog groups very-very fast for thousands of taxa.
- Badges
- Science topic
- Similar topics
- Computer Communications (Networks)
- Simulators