-
|
A common question for people starting to use |
Beta Was this translation helpful? Give feedback.
Replies: 3 comments 45 replies
-
|
Usually, the best way to infer ancestral alleles is to consider the allele present in one or (preferably) more outgroup species. This, of course, requires that the sequences in the outgroup(s) can be aligned with your data set. There are several published methods for doing this (see "Software", below). Note that the trivial approach of taking the REF allele as the ancestral state is likely to prove inaccurate in many cases, as is the shortcut of taking the most frequent allelic state (rough simulations show that in typical demographic simulations the most frequent allele is ancestral roughly ~80-90% of the time, for sample sizes of hundreds or thousands) Sometimes a VCF will already have an ancestral allele defined for each site, in the However, in our experience, the AA field is not always trustworthy: if the VCF is an old one, the ancestral allele it may have been inferred using rather simplistic algorithms. In some cases, other researchers may provide ancestral alleles in a separate file, for example the Ensembl project uses a simple multi-species algorithm to create ancestral allele FASTA files for a limited set of species, including humans. In many cases, however, you may need to infer the ancestral alleles yourself. Even if the ancestral allele states already exist for your species, running the analysis yourself gives you more control over the inference process and you can chose to use more recent approaches. Software for inferring ancestral allelesPeter Keightley's |
Beta Was this translation helpful? Give feedback.
-
|
Another tool http://compgen.cshl.edu/phast/help-pages/prequel.txt (Probabilistic REconstruction of ancestral seQUEnces, Largely) |
Beta Was this translation helpful? Give feedback.
-
|
The est-sfs method and software of Keightley & Jackson (2018) yield estimates of the probabilities of the ancestral states. Does tsinfer accept these probabilities as input? |
Beta Was this translation helpful? Give feedback.
-
Sadly not. It's all-or-nothing at the moment. It would be very nice to improve the ancestral-state handling elements of tsinfer, and we think it's doable, but is a small research project in itself (if you know anyone interested). |
Beta Was this translation helpful? Give feedback.
-
|
I am curious as to how this might be done. But my plate is full just understand what has been accomplished already. But I keep this in mind as a potential project. The potential for error correction is closer to my biological research interests. |
Beta Was this translation helpful? Give feedback.
-
|
We've been looking at different methods to do ancestral allele inference and they all start from an alignment with outer species. What would be the best approach to deal with non-aligned regions that can’t be inferred: leaving them as missing, using major allele as proxy, ... other? |
Beta Was this translation helpful? Give feedback.
-
|
@hyanwong above you suggested that SNP array data likely won't be a good starting point to infer tree sequences. Why? Thanks! |
Beta Was this translation helpful? Give feedback.
-
|
HI! As you sugested @hyanwong, we have been exploring the inference by chromosome arm vs. whole chromosome and this is the difference between the results for inferring the whole chromosome at once (Chr1) - the sum of both arms (Chr1_1 + Chr1_2):
There are some changes but they are maybe minor, (for all biallelic sites + rice recombination rate there are 3,000 nodes + mutations less when running by chromosome arm in ~800,000, about 0.3%). Do you agree that the is not much improvement in doing it separately? Here you can see the SNP data coverage and the centromeres (red dots) for a 1Mb window (left) and for 1kbp window (right). Also, the centromere information is of a particular approximate point and in not very precise numbers (e.g. 6.2Mbp, 12Mbp)
We have information of two different parts of the breeding programs that basically use different subspecies, indica and japonica, although there is some mix between them. Our plan is to test genomic prediction with a relationship matrix inferred from the trees using a genotypic matrix generated with both subspecies together ("All" genotypic matrix) vs. prediction with a relationship matrix inferred from the trees using a genotypic matrix generated with only one of the subspecies (Ind / Jap). The All genotypic matrix has more variable sites than Ind/Jap genotypic matrices because there are many sites that do not segregate within each subspecies or do so in such a frequency that wasn't enough to call that SNP in the calling process. Up to now, inferences were based on the All genotypic matrix. In order to see if the pattern of sites + mutations was the same with the other ones (Ind and Jap genotypic matrices), we inferred trees for each subspecies separately. These are the results for inferring ts from Chr1 for each ssp in their separate genotypic matrix (Ind[Ind] and Jap[Jap]) and also Ind and Jap from the All genotypic matrix (All[Ind] and All[Jap]) and All[All] for comparison as well. I could confirm that the number of excluded sites grows when inferring for Ind/Jap from the All matrix because fixed alleles in the subset apart from triallelic ones are automatically excluded.
Do you think there is something wrong with the data/way we are processing it or is it "normal" to see ~20 mutations per site and we should continue? Thank you!!! |
Beta Was this translation helpful? Give feedback.
-
@gregorgorjanc: mainly because it is sparse. You'll get the most efficient inference when there is enough information to resolve each edge - i.e. if there is at least one variant per edge. With SNP data you are likely to be missing lots of rare variants, I suspect, so the trees won't necessarily be well resolved. |
Beta Was this translation helpful? Give feedback.
-
It looks like that, yes. These are quite small centromeres. They are unlikely to bias your results much. It's simply that it might not be good practice to quite statistics that include large regions with no variant data (i.e. centromeres).
It does seem very high for that small number of samples, yes. The density plots are useful. It's reassuring that at the inference sites the number of mutations is (relatively) low, but the fact that it's so high at the non-inference sites (assuming these are mostly ones that don't have known ancestral states) implies to me that there's lots of the topology that you aren't correctly capturing. The sites with (say) >50 mutations would, I think, mostly be errors. It might be useful to have a look at how many mutations on a site with multiple mutations lie immediately above a sample node: these are almost certainly sequencing errors. I wonder if there are alignment problems, or extensive paralogous segments in these genomes? |
Beta Was this translation helpful? Give feedback.
-
|
@hyanwong, @minesrebollo has genomic data that pertains to individuals from a breeding programme spanning ~25 years or so. So, the samples are not from one generation, but span ~25 generations and are related through a pedigree. Breeding pedigrees are quite involved with lots of loops due to inbreeding. Could this have an impact on large number of mutations or are we likely facing "genome" issues? Thanks! |
Beta Was this translation helpful? Give feedback.
-
|
Ensembl has ancestral sequences inferred by Ortheus for some species https://m.ensembl.org/info/genome/compara/ancestral_sequences.html |
Beta Was this translation helpful? Give feedback.
-
|
Is this the direct link? https://ftp.ensembl.org/pub/release-100/fasta/ancestral_alleles/ |
Beta Was this translation helpful? Give feedback.
-
|
There are much more species with their ancestor inferred than the ones on that link. In Ensembl's compara API there is this script to get them https://github.com/Ensembl/ensembl-compara/blob/release/104/scripts/ancestral_sequences/get_ancestral_sequence.pl but couldn't make it work for rice... I downloaded the EPO alignment from the FTP site, converted it to .maf with https://github.com/Ensembl/ensembl-compara/raw/master/scripts/dumps/emf2maf.pl and accessed the sites with WGAbed. |
Beta Was this translation helpful? Give feedback.
Usually, the best way to infer ancestral alleles is to consider the allele present in one or (preferably) more outgroup species. This, of course, requires that the sequences in the outgroup(s) can be aligned with your data set. There are several published methods for doing this (see "Software", below). Note that the trivial approach of taking the REF allele as the ancestral state is likely to prove inaccurate in many cases, as is the shortcut of taking the most frequent allelic state (rough simulations show that in typical demographic simulations the most frequent allele is ancestral roughly ~80-90% of the time, for sample sizes of hundreds or thousands)
Sometimes a VCF will already have …