What is the best cut-off for e-value in BLAST? Some say 0.01 ,others 0.0001? Even BLAST tutorials don't give a very clear idea about which cutoff suits which purpose? Could some experts give me a clear idea on this. Why are we using 0.0001 and not 0.00011?
$\begingroup$
$\endgroup$
3
asked Jun 26, 2014 at 11:04
-
2$\begingroup$ There's virtually no difference between 0.0001 and 0.00011; cutoffs work in orders of magnitude. Depending on what I'm doing, sometimes I'll use 0.1, sometimes I'll use 1E-30 (0.000000000000000000000000000001). $\endgroup$– prooffreaderJun 26, 2014 at 15:20
-
$\begingroup$ Honestly, e value means very little and is hard to judge. Its often case dependent $\endgroup$– A. Radek MartinezOct 3, 2016 at 19:17
-
$\begingroup$ For nucleotide alignment use 1e-05 --> pathblast.org/docs/e_value.html $\endgroup$– Jaqueline AraujoOct 3, 2017 at 13:47
Add a comment
|
5 Answers
$\begingroup$
$\endgroup$
1
There is no perfect cut-off. It always depends on what you're doing. The e-value is basically a measure of how many such alignments you would expect to find in a database this size by chance. Therefore, e-values greater than 1 mean that you'd expect at least one alignment similar to what you've found by chance alone.
As others have stated, the e-value is not a rigorous measure of significance. It depends on
- The length of your alignment.
- The alignment's bit-score score (this takes into account amino acid substitution scores based on the matrix being used, BLOSUM65 by default).
- The size of the database you are searching in.
I have at different times and for different projects used e-values ranging from 10 to 1e-100 or less. You always need to take into account what you are looking for. For example, to run a tBLASTn looking for distant homologs, you'd use a relatively high e-value, say 1e-10 or even 1 or greater depending on the details. For example, the default e-value for ncbi-tBLASTn is 10.
To find extremely similar sequences in closely related species, you'd use a much smaller one. Say 1e-20 or 1e-50. Normally, you tweak these settings after seeing your results.
After running a few hundred blasts, you get the hang of it and you can have a pretty good estimate of the e-value cutoff you want to use. Until you get that experience, you will have to use trial and error. Just remember that the conclusions you can draw from your data will be heavily influenced by that e-value. Don't try to publish something claiming you've found a homolog if the supporting BLAST hit has an e-value of 45 for example.
Personally, when looking at BLAST out files, I check the e-value, the alignment's length (as a percentage of the query), the alignment's %id score and the bit score itself. Of these, the most reliable is the bit-score but, again, that too changes depending on what question you are attempting to answer.
-
1$\begingroup$ @OP I would also add this point that the cutoffs, scoring schemes and inferences are dependent on what you really want to find. It is to be first reasoned if BLAST is the right tool for what you are seeking. $\endgroup$– WYSIWYGJul 1, 2014 at 8:02
$\begingroup$
$\endgroup$
6
The e-value is supposed to be a metric for the chance that an alignment could occur at random, but it is a crude estimate. As pointed out in other answers, this significantly does not include the length of the query sequence.
It also does not include the conservation of the gene or the frequency of amino acids (in protein blasts). It does take into account the database of sequences queried, which might be something you know very little about if you didn't make it yourself.
That being said, 1e-5 is a pretty poor score. Its really awful for nucleotides - maybe a perfect alignment for 10 or 12 bases. So I'm going to assume that you are looking at peptides.
You should also look at the coverage statistic. That is the length of the alignment divided by the query or the hit sequence. It should be 70%+ depending on what you want. If the coverage is good even answers pretty far down the score rankings might be relevant.
Similarity is usually performed with a BLOSUM matrix. Regardless, the number of exact matches compared to the number of similar matches of amino acids in an alignment poorly approximates what is going on in evolution.
Lastly protein alignments tend to have gaps - lots of gaps. The penalty to the scores for allowing gaps might low for long evolutionary distances or completely unreasonable for more recently diverged results. So you have to look at those qualities of the alignment yourself.
-
$\begingroup$ "It does take into account the database of sequences queried, which might be something you know very little about if you didn't make it yourself." This is incorrect. Database size directly affects E-values. $\endgroup$– 5heikkiJun 27, 2014 at 10:45
-
2$\begingroup$ @5heikki.. what are you refuting here? Both the statements mean the same thing... $\endgroup$– WYSIWYGJun 27, 2014 at 12:49
-
$\begingroup$ Sorry but this is really not correct. As explained in the page you link to, the e-value is given by
E=Kmne^λSwhereSis the alignment's score. Since this is a measure of (among other things) the conservation of the two sequences, the e-value includes both "conservation of the gene" and "the frequency of amino acids". Also, to my knowledge, all popular BLAST implementations (and certainly ncbi and wu-blast) use BLOSUM65 matrices, not PAM by default. In fact, I don't know of any program that uses PAM matrices by default and I can't imagine why you would want to. $\endgroup$– terdonJun 27, 2014 at 18:51 -
$\begingroup$ en.wikipedia.org/wiki/Point_accepted_mutation#Use_in_BLAST $\endgroup$– shigetaJun 27, 2014 at 22:35
-
$\begingroup$ CLUSTALW uses PAM and BLOSUM both.. It is preferable to use BLOSUM for local alignments and PAM for global alignments.. Makes intuitive sense, thats it; no hardcore reason.. $\endgroup$– WYSIWYGJun 28, 2014 at 5:03
$\begingroup$
$\endgroup$
E-value refers to the expected number of random hits for a given alignment score. Smaller it is more reliable is your match. There is no hard and fast rule for e-value cutoff. You can keep whatever you want depending on the level of stringency that you require. But you should note that for smaller sequences (< 30nt) there is always a higher likelihood of random matches. In such cases it is practical to relax the e-value cutoff.
$\begingroup$
$\endgroup$
2
The E-value of 0.0 indicate the number of alignments with scores equivalent to or greater than that are expected to occur in a database by chance therefore the lower the E-value the more significant the score hence a better quality of the alignment blast search.
-
1$\begingroup$ can you share a reference that others can read for more information on your answer? it would definitely strengthen your answer - thanks for your contribution! $\endgroup$ Oct 3, 2016 at 0:35
-
$\begingroup$ A score of 0.0 zero isn't actually zero it's just lower then 10^-100 $\endgroup$ Oct 3, 2016 at 18:42
$\begingroup$
$\endgroup$
3
The e-value is required to have confidence in the hit, along with hit. Generally, an e-value of e-4 is preferred because, this cutoff is found to be sufficient enough by blast to confirm a hit as homolog. But, for a hit, an e-value greater than e-4 does not mean it is a homolog neither we can say it is not a homolog. We cannot surely infer anything from these hits. Therefore, researchers choose a value
-
1$\begingroup$ There is no magical threshold that defines homology! It depends completely on the specifics of your situation. There are cases where you can find true homologs and an evalue >10. Everything depends on the length of your sequences, the size of your database, the conservation of your homologs etc. $\endgroup$– terdonJul 14, 2014 at 14:34
-
$\begingroup$ Sorry if my answer seems confusing. But, is it not that an evalue much closer to zero is considered homolog? Because, i see in some papers and tutorials using e-4 or lesser value to assess homology as a rule of thumb. Those hits, where the evalue > 10 may or may not be a homolog, but i think we cannot definately say they are homolog, without further insights. Would like to hear your comments. Thanks. $\endgroup$ Jul 15, 2014 at 5:02
-
1$\begingroup$ No. We can never decide on homology based on the e-value alone and there is no single magical threshold. The threshold you choose will always depend on the particular question you are asking (how diverged are the species? How long is your sequence? How large is the database? Is this nucleotide or protein blast? etc). What is important for deciding homology is the particular regions of a sequence that are conserved (e.g. domains). For example, if you search for a homolog of a very short sequence with repetitive regions, you will never have a small e-value. $\endgroup$– terdonJul 15, 2014 at 9:58