The MEME Suite

psp-gen is used to allow MEME to perform discriminative motif discovery—to find motifs overrepresented in one set of sequences compared to in another set. It takes two files as input—the sequence file to be input to MEME, (the "primary" file) and a "control" sequence file of sequences believed not to contain the same motifs as in the "primary" file. psp-gen creates a file for use by MEME that encapsulates information about probable discriminative motifs. psp-gen records its chosen motif width in the file, and MEME is able to adjust the data when it tries different motif widths.

Primary Sequences File

Control Sequences File

A FASTA-like PSP format is written to standard output, it contains a prior for every position of every sequence in the primary set.

Option	Parameter	Description	Default Behaviour
General Options
-h		Print a short usage message and exit.	Run as normal.
-minw	min w	The minimum width to use with selecting the "best" width for PSPs.	The minimum width is set to 4.
-maxw	max w	The maximum width to use with selecting the "best" width for PSPs.	The maximum width is set to 20.
-dna		Use the DNA alphabet.	The DNA alphabet is assumed.
-protein		Use the protein alphabet.	The DNA alphabet is assumed.
-rna		Use the RNA alphabet.	The DNA alphabet is assumed.
-alph	file	Use the alphabet defined in the file.	The DNA alphabet is assumed.
-triples		Use spaced triples instead of whole-word matches. Recommended for protein.	Whole-word matches are used.
-fixedstart		When using the -triples option, only consider triples starting at the start of the site.	Allow triples to start anywhere in a width w site.
-equiv	Letter Groups	Any sequence of letter that appears together should be matched as equal. Separate letter groups using "-", e.g. -equiv "IVL-HKR" means treat all occurrences of I, V or L as the same, and all occurrences of H, K or R as the same. Note that in alphabets where '-' is an allowed symbol separate groups may be specified by repeating the -equiv option for each group.	No letter groups are used.
-revcomp		Both strands are considered when calculating PSPs for complementable alphabets.	Only the given strand is considered.
-scalemin	min score	The lowest score value after scaling.	The lowest score is set to 0.1, unless the -scalemax option is set in which case the lowest score is max score - 1.
-scalemax	max score	The highest score value after scaling.	The highest score is set to 0.9, unless the -scalemin option is set in which case the highest score is min score + 1.
-maxrange		Choose the width with the biggest difference between minimum and maximum scores (before scaling).	Choose the width with the biggest maximum score (before scaling).
-raw		Output scores instead of priors.	The program output PSPs.
-reportscores		Send to standard error primary and control file names, min and max scores and min and max widths, tab-separated.	Do not report scores to standard error.
-verbose		Send detailed stats to standard error, reporting frequency of each score.	No extra information is reported.

In each example, the sequence file is the same as the postive set used for psp-gen. While there is no check that you have done this because MEME runs as a separate program (other than that the sequence names need to match, as do the number of sites), using a primary set different to the sequence set MEME searches is unlikely to be useful.

Generate a discriminative prior for a DNA sequence set (pos-dna.fasta) likely to contain TF binding sites, given another DNA sequence set unlikely to contain TF binding sites (neg-dna.fasta):
psp-gen -pos pos-dna.fasta -neg neg-dna.fasta > dna.psp

Use these position-specific priors in MEME, searching both strands:
meme pos-dna.fasta -psp dna.psp -oc meme_out_dna -revcomp

Generate a discriminative prior for a protein sequence set (pos-prot.fasta) associated with a specific function, given another protein sequence set unlikely to relate to that function (neg-prot.fasta):
psp-gen -pos pos-prot.fasta -neg neg-prot.fasta -protein -triples -maxrange > prot.psp

Use these position-specific priors in MEME:
meme pos-prot.fasta -psp prot.psp -oc meme_out_prot

View usage information:
psp-gen -h

The psp-gen program generates a position-specific prior (PSP) in a data file, which can be used as an additional input to MEME to bias the search to sites more likely to contain a motif (or motifs) of interest. Without using a PSP, MEME assigns the same prior probability to all sites. A PSP is generated by psp-gen using two sequence sets, each in FASTA format:

• the primary set (or X) contains the same sequences as you intend to search for a motif using MEME
• the control set (or Y) is a contrasting set of sequences that is unlikely to contain the motif or motifs of interest, but is otherwise similar to the primary set.

The basic principle of a discriminative prior is to start from the question:

The intuition is that any w-mer that is common in the primary set but not the control set could form part of motif of interest, based on the way the two sets are chosen.

PSP calculation starts from the following equation based on the D-prior described in the Narlikar et al. RECOMB 2007 paper "Nucleosome Occupancy Information Improves de novo Motif Discovery". We count the number of instances X_i,j(w) of the subsequence w wide starting at position j in sequence X_i of the primary set, and similarly count Y_i,j(w) in the control set, then calculate for each site $\begin{array}{c}{D}_{i,j}=\genfrac{}{}{0.1ex}{}{X_{i,j}\left(w\right)}{X_{i,j}\left(w\right)+Y_{i,j}\left(w\right)}.\hfill \end{array}$ As pseudocounts, we add 1 to the enumerator, and 1 + |X|/|Y| to the denominator. The purpose of the pseudocounts is to avoid giving a high score to infrequent subsequences that are not (or rarely) found in the control set. Once we have scored all sites for a given value of w, we scale scores to a range that can be set on the command line, with the default range [0.1,0.9].

We repeat scoring for the whole range of motif widths under consideration, and choose a "best" width using one of two methods: either we choose the value of w that maximizes the score over all sequences, or the value of w that maximizes the difference between the maximum and minimum score over all sequences. We also allow an option of using spaced triples instead of whole words to match subsequences. See Variations below for more details on triples and choosing the "best" width.

Once we have scored every available site and chosen a specific value of w, we convert the scores to probabilities, based on the proportionalities $\begin{array}{c}\text{Pr}(Z_i=0)\propto 1\text{ and }\text{Pr}(Z_i=j)\propto \genfrac{}{}{0.1ex}{}{D_{i,j}}{1-D_{i,j}},\hfill \end{array}$ where Z_i = j means that the site for a motif in sequence X_i is found at location j (with j = 0 signifying there is no site in the given sequence), and D_i,j is the score for sequence X_i at the site starting at location j, numbered from 1. We convert these proportionalities to a PSP by normalizing them to add up to 1 over each sequence.

Since MEME runs as a separate program, there is no way to check that the primary set is the same as the sequence set given to MEME. There are however checks that every name in the PSP file matches a name in the MEME sequence set, and that the number of sites in the PSP file for a particular sequence name matches the number of sites in the sequence file seen by MEME. While it does not make sense to use a completely different sequence set for PSP generation and for running MEME (with the resulting PSP file), MEME can use a PSP file that does not provide probabilities for every sequence. Any sequences not in the PSP file are given a uniform prior probability. MEME documentation includes more detail on how PSPs are used.

For each variation, "default" means that you do not need to specify a command-line switch to turn on that setting.

• sequence alphabet: choose -dna (default), -rna or -protein. For DNA, you also have the option to score matches on both strands:
  • -revcomp (the default is to score only the given strand). This setting applies to both the primary and control sequences.
• choosing the "best" width: specifying -maxw and -minw provides limits to psp-gen but the algorithm chooses one width to generate priors after scoring the range of allowed widths, using one of two variations:
• maximum score (default): choose the value of w that maximizes the score over all sequences.
• maximum range of values (-maxrange): choose the value of w that maximizes the difference between the highest and lowest score over all sequences. This variant is useful in data sets where a high number of sites achieve similar high scores (more likely to happen with protein sequences).
• type of word match:
  • whole-word match (default)
  • spaced triples (-triples): for a case where exact-word matches are unlikely to succeed, spaced triples are an alternative. Each w-wide word is broken down into triples containing an initial, middle and final letter, and all other letters are treated as don't cares (i.e., they match anything). In spaced triples mode, psp-gen scores a site on the basis of the spaced triple fitting the w-wide word at that site that has the highest score. For example, the subsequence
    MTFEKI contains the following triples:
    MT...I
M.F..I
M..E.I
M...KI
    where "." matches anything when counting matches.
    • An option that only applies to triples is -fixedstart, which only allows triples for the current site under consideration that start at the beginning of the site. This option reduces execution time at the expense of considering fewer possibilities for scoring a site.
    • default: consider triples of all widths that fit the current site.
• scaling variations (-scalemin and -scalemax): the scoring formula creates values in the range [0,1], but the conversion to PSPs from scores requires values strictly less than 1. Choosing an option other than scaling between the default [0.1,0.9] reduces (if you narrow the range) or increases (if you widen the range) sensitivity to small differences in scores. If you only set one of scalemin or scalemax, the other is calculated as described in Options.
• ambiguous characters: you have two levels of control of ambiguous characters:
  • ambiguous characters in the sequence data: The effect of encountering any of these letters is to score a word or triple containing them as if it only occurs once, to minimize its score.
  • treat specific letters as equivalent: use the -equiv option to treat specific groups of letters as if they were all the same (default: all letters are treated as unique). For more detail, see Options.
• reporting and debugging: you have various options for understanding better how PSPs are calculated.
  • see raw scores: -raw stops the last stage of computation that converts the scores to probabilities; the scores sent to standard out are however scaled, using either the defaults or selected values of scalemin and scalemax (default: convert scores to PSPs).
  • report range of scores: see -reportscores in Options (default: do not report).
  • see detailed statistics: -verbose causes detailed statistics on scores to be sent to standard error (default: no statistics reported).

When you use exact-word (w-mer) search, psp-gen run time scales linearly in each of |X|, |Y| and maxw-minw. Formally, time is O((|X| + |Y|)(w_max - w_min)). For triples, psp-gen scales linearly in each of |X| and |Y|, and quadratically in w_max, so run time is O((|X| + |Y|)(w_max)²).

Extra memory in both cases scales linearly with |X| + |Y|.

Since triples have a higher-order term in time complexity, we predict run time scalability in more detail than for exact word match. Adding in constants and a linear term, time complexity for triples is approximated by (where W = w_max-2, the number of variations in width used by the algorithm):

t

=

(k1|X| + k2|Y|)W2 + k3(|X| + |Y|)W

On a real machine (iMac, 3.06GHz Intel Core 2 Duo, 6MB L2 cache, 4GB RAM) the following is a good predictor of the upper bound on run time for likely values of |X|, |Y| and range of widths. Scale any results according to the speed of the actual machine on which you intend to run psp-gen. The formula reflects runs on randomly generated data; real data will likely result in faster runs because random data defeats some optimizations in psp-gen.

t

=

(4.37x10-6|X|+ 2.46x10-06|Y|)W2 + 6.25x10-06(|X| + |Y|)W s

For example, for |X| = |Y| = 60,000 and w_max = 50 (i.e., W = 48), estimated run time = 980s, the same in this case as measured run time. The following command line reproduces this run:

      time psp-gen -neg pos60k.fasta -pos neg60k.fasta -protein -minw 3 -maxw 50 -triples > priors60kmaxw50.psp
      

• Any letters in sequence represening ambiguous bases or amino acids result in a word (w-mer) containing that word (or triple in the case of protein) being counted whether in the primary or control sets as if it occurred once, to minimize its score.