Documentation for Datamonkey's Analyses
To perform a selection analysis, datamonkey.org needs a multiple alignment of at least three homologous coding nucleotide sequences. Codon based methods for estimating dN and dS can be applied to any sequence alignment, but there are several considerations to keep in mind:
Ideally, the alignment should represent a single gene, or protein product, sampled over multiple taxa (e.g. mammalian interferon genes), or a diverse population sample (e.g. Influenza A viruses infecting different individuals). Because comparative methods estimate relative rates of synonymous and non-synonymous substitution, substantial sequence diversity is needed for reliable inference. For example when, Suzuki and Nei applied a REL-type method to a very low divergence (1 or 2 substitutions per sequence along a star phylogeny) sample of the Human T-lymphotropic virus (HTLV), they found that the method performed poorly. Yang and colleagues have suggested that the total length of the phylogenetic tree should be at least one expected substitution per codon site, but it is impossible to give a generally valid range for desirable sequence divergence. However, sequences that are too divergent could lead to saturation, i.e. our inability to reliably infer branch lengths and substitution parameters. The number of sequences in the alignment is important: too few sequences will contain too little information for meaningful inference, while too many may take too long to run. At the time of this writing, Datamonkey permits up to 150 sequences for SLAC analyses, 100 for FEL/IFEL analyses, 40 for REL and PARRIS and 25 for GA-Branch. As a rule of thumb, at least 10 sequences are needed to detect selection at a single site (SLAC/FEL/IFEL/REL) with any degree of reliability, while as few as 4 may be sufficient for alignment-wide inference (PARRIS/GA-Branch). The median number of sequences in an alignment submitted to Datamonkey is 19. Comparative methods are ill suited to study certain kinds of selection. For example, they should not be applied to the detection of selective sweeps (rapid replacement of one allele with a more fit one, resulting in a homogeneous population), unless sequences sampled prior to and following the selective sweep are included in the sample. A number of publications have dealt with this issue extensively (e.g. Selection using HyPhy), and we refer an interested reader to one of these works for further insight.
It is a good practice to visually inspect your data to make sure that the sequences are alignment correctly. Of course, one can never be sure that an alignment is objectively correct, but gross misalignments (e.g. sequences that are out of frame) are easy to spot with software that provides a graphical visualization of the alignment, such as HyPhy, Se-Al, or BioEdit. Datamonkey uses the HyPhy package as its processing engine, and if an alignment does not open in HyPhy on your machine (using the File:Open:Open Data File command), then it will not be properly read by Datamonkey.
You should verify that the alignment is in frame, i.e. that it does not contain stop codons, including premature stop codons (indicative of a frame shift, e.g. due to misalignment, or a non-functional coding sequence) and the terminal stop codon. Your alignment should exclude any non-coding region of the nucleotide sequence, such as introns or promoter regions, for which existing models of codon substitution would not apply. When coding nucleotide sequences are aligned directly, frameshifting (i.e. not in multiples of 3) gaps may be inserted, since the alignment program often does not take the coding nature of the sequence into account. Therefore it is generally a good idea to align translated protein sequences and then map them back onto constituent nucleotides. Datamonkey will perform a number of checks when it receives coding sequences and report all problems it encounters.
If the alignment contains identical sequences, Datamonkey will discard all but one copy before proceeding. This is done to speed up the analyses, because identical sequences do not contribute any information to the likelihood inference procedure (except via base frequencies), but the computational complexity of phylogenetic analyses grows with the number of sequences.
Finally, Datamonkey may rename some of the sequences to conform to HyPhy naming conventions for technical reasons (all sequence names must be valid identifiers, e.g. they cannot contain spaces). This is done automatically and has no effect on the subsequent analyses.
Sequence types accepted by various Datamonkey components. Available analyses are automatically determined by the server based on the contents of the alignment and user description from the upload page.
Other genetic codes are defined in terms of differences with the Universal code.
Datamonkey automatically recognizes five aligned sequence data formats and also autodetects whether the data is nucleotide (codon) or aminoacid.
The following NEXUS blocks are supported:
ASSUMPTIONS (for data partitioning) and
PHYLIP option characters in the first line are ignored for both sequential and interleaved formats.
#), and complete sequence data follow the name of the taxon.
#_, and blocks of sequence data follow in the same order as the names of the taxa.
Complete model selection procedure details can be found in this MBE paper
We recommend that you run a model selection procedure, which sifts all 203 possible time-reversible models through a hierarchical testing procedure combining nested LRT tests with AIC selection to pick a single "best-fitting" rate matrix. Model selection is processed on a remote cluster, and should take no more than a few minutes to complete.
To allow the most general model of nucleotide substituion, select the General Reversible Model (REV), since it does not add much to the overall processing time. However, if your data set is small, it may not be possible to accurately estimate nucleotide substitution bias rates, and HKY85 might not be a bad choice. You can also try several different models and see if the location of inferred sites changes depending on the nucleotide model (it rarely does, unless the model is very wrong).
For more details see MBE paper
All possible resolutions of an ambiguous character contribute, in a weighted fashion, to the computation of EN, ES, NN and NS (see methods paper). Characters without any information (all gaps or all missing) are NOT counted though, to avoid artifically high dN and dS estimates.
The most likely resolution for the given site is used in the computation of EN, ES, NN and NS. Ties are broken randomly.
Consider the site:
ACA ACG ACG ACR
For the resolved option, only most frequent resolution based on the data in the site only, will be considered. In this case, the resolution is 'ACG'
For the averaged option, all four possible resolutions ('ACA' and 'ACG') will be considered. The weight factor for each resolved is determined by the relative frequency of that codon to all possible resolutions. If f(xyz) denotes the frequency of codon xyz in the entire data file, then the contribution of ACA will be f(ACA)/(f(ACA)+f(ACG)) and of 'ACG' : f(ACG)/(f(ACA)+f(ACG)).
For more details see MBE paper
A (somewhat) arbitrary cut-off value for the Bayes factor for the event of positive selection at a given site. When a Bayes factor sufficiently greater than 1 (default is 50; values between 10-100 provide strong evidence for the hypothesis, while values over 100 are decisive) suggests that the data supports positive (or negative) selection at that site. This value does correspond to statistical significance, but the exact mapping between Bayes factors and p-values is data set dependent. The reciprocal of a Bayes Factor is loosely analogous to a p-value. REL methods are most sensitive, but are also most prone to false positives.
|Sequence Alignment||The FASTA formatted multiple sequence alignment that is to be uploaded. Please see TODO for details|
|Notify When Completed|
The FASTA formatted multiple sequence alignment that is to be uploaded. Please see Data Formats for details.