A question on MEME results

Hello everyone,
Recently I'm searching for a sequence pattern from some fasta sequences using MEME, I have 821897 sequences in total fed into MEME for de novo motif searching using meme default parameters meme -nmotifs 3 file.fa -searchsize 1520000 -oc file_meme -seed 0620 -dna -revcomp and found a significantly strong motif like this (here I say the motif is strong because of 821411/821897, this may be argued):

Use all 821897 sequences:

all-seqs

I naturally think that, given so strong motif, the motif will remain largely similar when I randomly choose some sequences, however, things became weird when I sampled 3 times of 500000 sequeces like below:

Use sampled 500000 sequences for three times: shuf1 shuf2 shuf3

It seems these three motifs are all strong still, but vary a lot. I am not sure what I did wrong, and your advice would be much appreciated.

P.S. I add the result of the motif generated by Weblogo3 for comparison.
Use all 821897 sequences: weblogo
Use sampled 500000 sequences for three times: shuf1 shuf2 shuf3

The sampled 500000 sequence for MEME and Weblogo is exactly same, My questions are:
1. Why the motif generated by MEME using almost all sequences is different to Weblogo's, which also used all sequences. I know that MEME will use some algorithm to refine motif, and weblogo simply stack all base nucleotides, but will this differ so much?
2. Why three sampled results of Weblogo are similar, but differ a lot in MEME's?
Thank you for your time!

meme motif

• 318 views • link

written 7.1 years ago by Hughie •

1 answer

You seem to have a palindromic motif here:

ACAGTxTAxACTGT

However, the distribution of motifs in your 3 subsets makes me think that you didn't choose a palindrome option in your search. If not, I suggest that you do so. I don't think your motifs will vary that much and the motif itself will be stronger. Also suggest that you use the anr switch (any number of repeats) as that may find additional sequences if there are multiple copies of your motif. Lastly, I have no clue how you managed to find >800K sequences that have the same motif, but it could be easier to work with fewer sequences if you remove the redundancy at 90% identity.

• 0 views • link

written 7.1 years ago by Mensur Dlakic

Thanks for your reply, Mensur.
Yes, you are insightful for finding this is a palindromic motif with such low bits. However, I didn't designedly filter out palindrome or something for sequences.

I don't think your motifs will vary that much and the motif itself will be stronger.

Would you please give more cues why you think the motif is not strong?
Also, duplicated sequences will be removed in MEME by default, so, there is no a copy of any sequence.
I just used mostly default parameters of MEME like this:

meme -nmotifs 3 file.fa -searchsize 1520000 -oc file_meme -seed 0620 -dna -revcomp

Sorry for that I'm a beginner in analysis and I don't understand the algorithm very well, Thanks again for your kind reply!
BTW, I have updated my post for more details.

• 0 views • link

written 7.1 years ago by Hughie •

However, I didn't designedly filter out palindrome or something for sequences.

Using a palindrome switch (-pal) asks meme to specifically look out for palindromes. I think you should use it given that your motif clearly is a pseudo-palindrome (with TxA in the middle rather than TA I surmised originally). It will also improve the signal and hopefully push some of your letter stacks above 0.5 bits. Since you are using -revcomp (correctly so), adding -pal will (almost) always find matches on both strands and it will (almost) double the number identified motifs.

I think that meme by default uses -mod zoops to look for zero or at most one copy of a motif per sequence. That is usually a sensible choice. I suggest that you try -mod anr which will look for Any Number of Repeats. That includes zero or at most one copy of a motif per sequence as specified by the zoops switch, but it will also look for multiple copies of the same motif in one sequence (if any). Beware that this will increase the run time 5-10x, but it may find more than one copy per sequence. Then again, it is unlikely that you have >800K natural sequences that contain the same motif, so maybe you know already that there aren't multiple copies of it.

Also, duplicated sequences will be removed in MEME by default, so, there is no a copy of any sequence.

I don't have a recent version installed, but this not the case at least up to V4.0. As far as I know, meme works with sequences as given without removing anything.

• 0 views • link

written 7.1 years ago by Mensur Dlakic

Thank you again Mensur,
I will try your advice and show the result later :)

• 0 views • link

written 7.1 years ago by Hughie •

Hi Mensur, I have followed your suggestions using -pal and got a reasonable result below:

sampled results:

I have also tried different -mod parameters including anr, but results are all weird except for -mod zoops, so, I may use this later.
Thanks for your kind help!

• 0 views • link

written 7.1 years ago by Hughie •