According to NCBI Blast tutorial, e-value indicates the number of hits one can "expect" to see by chance when searching a database of a particular size. However, this definition is somewhat too abstract for me to visualize.
For my research, I am continuously using blast tools, i.e. megablast and blastn. E-value is a complex parameter that takes alignment, size of the database and length of the query sequence into consideration. I am curious for potential explanations that would enable me to better understand the meaning of e-value.
Well, I can't agree. "Statistical significance" of e-value is far more higher than % output. Actually, e-value is the only "real stastitical value" you get from BLAST. But I didn't want it to replace the other values. Of course you should consider all the values for your result, that's why they're here. Me, I also always look for query coverage % and max indent %. But...
Low e-value shows that your result is statistically good, but in fact it doesn't show anything about how close is the relation between sequences. % doesn't show anything about phylogeny relation, neither about how good the result statistically is.
You probably mixed statistical and biological intepretation from my point of view.
Hi Bilgenur, Yes your explanation about the e-value parameter is correct. Generally, e-value tells about the accuracy of the hits. The more the e-value, more is the possibility of it to be a false positive, while least e-value indicates hits to be correct. So an e-value of 0 indicates it to be the best hit. Generally, 1e-05 or 1e-10 is used as a preliminary filter for e-value.. Thanks!
Hi Bilgenur,
try to visualize it on an example:
Let's have database you BLAST against - GenBank database and database of random sequences - RANDOM database. Both of these databases are the same size.
Now you BLAST your query sequence against GenBank database. You get HSP (high scoring pair), a similar sequence in database, with e-value of 5. That means you would get 5 HSPs with the same or better similarity in RANDOM database. There isn't such a random database in real, so you only expect it, that's why expectancy value ;-)
Now, you probably feel that your match isn't good, it's more likely to be an "accident". The lower e-value, the better result (statistically). E-value is better way to rate your HSP than for example % similarity, because it counts also the length of your query sequence as well as the size of the whole database.
Karel
Dear @Karel Sedlar, good explanation. However, e-value alone would not be sufficient to interpret the output.
For instance, a DNA sequence which shows a high degree of similarity over a short region may give a very low E-value for this region even though these two sequences are not closely related. Hence, one would need to look at both e-value and % similarity in order to conclude whether the result is statistically good or not.
Thus, I disagree with you on your comparison of e-value and % outputs with regards to statistical significance. What do you think?
Well, I can't agree. "Statistical significance" of e-value is far more higher than % output. Actually, e-value is the only "real stastitical value" you get from BLAST. But I didn't want it to replace the other values. Of course you should consider all the values for your result, that's why they're here. Me, I also always look for query coverage % and max indent %. But...
Low e-value shows that your result is statistically good, but in fact it doesn't show anything about how close is the relation between sequences. % doesn't show anything about phylogeny relation, neither about how good the result statistically is.
You probably mixed statistical and biological intepretation from my point of view.
When you do a blast search you have a query sequence, a database and BLAST parameters. The BLAST results provide you with a "hit". The hit is somewhat similar to your query. Is this similarity due to random chance or is it because your query and the hit are actually related to each other? To answer this question we can calculate the average number of hits that would match your query in a “random database” of the same size and have the same or better score than the hit you have found in the actual database. This number of hits is called E-value. E-value = 1 means that you expect a hit that is as good in the given database due to chance alone. E-value = 10 means that you expect 10 such hits. Thus if your hit has the E-value of 1 or more you can’t exclude the “null hypothesis” that your query and the hit are unrelated. However if the E-value is small, say 0.0001 you can say that it is unlikely that a match so good would be found in the database by chance alone. Thus you can exclude the null hypothesis and conclude that your query sequence is in fact somewhat related to the hit you found. I hope this helps.
I would also note that while you can compare the e-values of several hits in the same BLAST session to say which hit is closer, you can not compare the e-values of two hits if they are from different BLAST searches against different BLAST databases. This is because e-value depends on the database size. The value of bit score is what you want to use in this case as it measures the degree of similarity between two sequences and does not depend on anything but the alignment and the weight matrix. The bit score of 80 for protein alignments is usually indicative of homology in my personal oppinion.
From the debate going on, I understand that the E-value is most representative compared to the percentage of similarity. What would be the better parameter to consider between the coverage and the E-value? in some cases, one gets low E-value with low coverage.
In that case (low e-value, low query coverage), you have to consider both of the values. It shows that your sequence is long enough to be representative (to get statistically significant results), but there are no matches with high percentage similarity in the database you search in. If I got such a result, I would consider to change the database.
e-value is a probability that this result was obtained randomly. e-value = 1 means completely random, e=0.01 means your result has a chance of 1% to be a fault. In general, E-value is the most important parameter, however other parameters are also important, eg. "coverage" in domains analysis; identity in function predictions etc.
Even combined, the E-value and the % coverage are not enough for, say, molecular species identification. Manual examination is the key, if you ask me. Some examples are given here:
https://www.researchgate.net/publication/235931750_Five_simple_guidelines_for_establishing_basic_authenticity_and_reliability_of_newly_generated_fungal_ITS_sequences
Article Five simple guidelines for establishing basic authenticity a...
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Bioinformatics