MEME performance: running time and data set size

[cegrant]

 MEME's running time depends on the type of model you have chosen: oops < zoops < anr. Running time also depends on the objective function you have chosen: ce,cd < de, se < classic « nc, for small datasets; the order reverses (except for nc) for very large datasets. Running time also is affected by the options -csites, -searchsize, -hsfrac, or -cefrac. Running time is shorter with smaller alphabets (e.g., DNA) compared to with larger alphabets (e.g., Protein). Running time increases with the number of motif widths that MEME considers (see the options -w, -maxw, -minw, and -allw). The table below gives more details on the relationship between sequence dataset size and MEME's running time with large datasets.

objective function	running time per motif (for large datasets)
classic, nc	approximately nlog(n) for motifs with up to csites sites, where n is the size of the sequence file; thereafter, the running time is essentially constant
de, se	approximately nlog(n) for sequence files with up to searchsize divided by hsfrac letters, where n is the size of the (primary) sequence file; thereafter, the running time is essentially constant
ce, cd	approximately nlog(n) for sequence files with up to searchsize divided by cefrac letters, where n is the size of the (primary) sequence file; thereafter, the running time is essentially constant

The following table shows actual running times for MEME on DNA datasets containing different numbers of 100bp sequences on a 4 GHz Intel Core i7 processor with 16GB of memory. MEME was run using either one or 6 processors (option -p), different models and objective functions (options -mod and -objfun), and the option -revcomp. (The parallel version of MEME scales up to about 128 processors. Please see https://academic.oup.com/bioinformatics/article/12/4/303/183600 for a discussion of the parallel version of MEME.)

processors	OOPS model	ZOOPS model	ANR model
1
6