STREME is a tool for discovering ungapped motifs that are enriched in the primary dataset that you provide. (If you have fewer than 50 sequences, MEME may work better than STREME.) It can work with sets of sequences in the DNA, RNA, Protein alphabets, or in a custom alphabet designed by you. You can also use STREME to discover motifs that are enriched in the primary dataset relative to a second (control) dataset that you provide. By default, STREME creates a control dataset for you by shuffling each of the sequences in the primary dataset.
Note: We recommend using STREME when you wish to find motifs in sets of 50 or more sequences, and we recommend using MEME for smaller datasets (1 to 49 sequences).
STREME searches for motifs by iterating the following five steps to until the stopping criterion you selected is met. The stopping criterion can be either the statistical significance (p-value) of the motif, or the total number of motifs found.
STREME builds a single suffix tree that includes both the primary and control sequences (but not the hold-out set sequences, see below under "Motif Statistical Signficance" for an explanation of the hold-out set). By default, STREME looks for motifs that optimize a "Differential Enrichment" objective function, but it can also find motifs that tend to occur in the centers of sequences using a "Central Distance" objective function.
STREME uses the tree to efficiently evaluate all seed words of length up the maximum motif width, computing the p-value of each such word's relative enrichment in the primary sequences using the chosen objective function. (Note: With the Differential Enrichment objective function, STREME will use the Binomial test instead of Fisher's exact test if the primary and control sequences have different average lengths. With the Central Distance objective function, STREME computes the cumulative Bates distribution of the average distances of the seed word from the centers of the sequences.)
STREME converts each of the best seed words into a motif, and iterative refines each motif, selecting the motif that best discriminates the primary sequences from the control sequences. At each iteration of refinement, the current motif and the k-order background are used with the suffix tree to efficiently find the best site in each sequence. The primary and control sequences are then sorted by the log-likelihood score of their best site, and the score threshold that optimizes the p-value of the statistical test (which depends on the chosen objective function) is found. The iteration ends by estimating a new version of the motif from the single best site in each primary sequence whose score is above the optimal threshold. This new motif is used in the next refinement iteration. Refinement stops when the p-value fails to improve or the maximum number allowed iterations have been performed.
STREME computes the unbiased statistical significance of the of the motif by using the motif and the optimal discriminative score threshold (based on the primary and control sequences) to classify the hold-out set sequences, and then applying the statistical test (Fisher's exact test, Binomial test, or the cumulative Bates distribution) to the classification. Classification is based on the best match to the motif in each sequence (on either strand when the alphabet is complementable).
STREME "erases" each of the sites of the best motif from both the primary and control sequences by converting the sites to the "N" (DNA and RNA) or "X" (Protein) character.
STREME works best with lots of relatively short (≤ 1000 character) sequences. If you have a just (a few) long sequences, then you probably should split them into many smaller sequences. With ChIP-seq data we recommend using 100bp regions centered on the peaks. With CLIP-seq data, we recommend using the actual CLIP-seq peaks (without centering or trimming). If you have fewer than 50 sequences, you might want to consider using MEME instead of STREME.
STREME always uses a control sequence set, but you don't have to supply it as STREME will create one by shuffling the input sequences. If you wish to use your own sequence set then there are a few guidelines you should follow.
The sequence lengths of the control sequences should be roughly the same as the sequences to search for motifs. This is because STREME uses a null model that assumes that the probability of finding a match in a sequence in either sequence set will be roughly the same for an uninteresting motif. If the average length of your control sequences is longer than that of the primary sequences, STREME trims the control sequences so that both sets have the same average length.
STREME can compute very accurate statistical significance estimates (p-values) for the motifs it discovers, as long as there are at least 50 sequences in the primary set of sequences. The way STREME does this is by holding out some sequences (10% by default) to use only for computing motif significance. To make the significance estimate accurate, STREME selects the "hold-out set" at random before searching for motifs. If your primary sequence set is too small (< 50 sequences), STREME will not report motif p-values, and the SCORE that it reports instead should not be used as an indication of motif significance.
When searching for RNA motifs, you can provide the sequences in either the RNA or DNA alphabets. STREME will automatically convert the sequences to the RNA alphabet, and report RNA motifs. When searching for RNA motifs, STREME treats the sequences as single-stranded, even if they use the DNA alphabet.
STREME has a parameter, order, that controls how it compensates for the innate non-randomness of biological sequences. It uses this parameter to control the order of the Markov model it uses for modeling sequences, and for determining how to shuffle the primary sequences when creating control sequences (unless you provide them). Setting order to higher values can help STREME to avoid finding uninteresting motifs that are just due to the non-randomness of biological sequences (e.g., CpG islands).
By default, STREME sets order to 2 for DNA and RNA sequences, and to 0 for protein sequences and sequences in user-specified alphabets. STREME always constructs and uses a Markov model of the control sequences (or the primary sequences if you don't supply control sequences). When shuffling sequences, STREME does so in such a way that the frequencies of words of length order+1 in each sequence remains the same before and after shuffling. You can change the value of order via the web-server or on the command line.
STREME essentially ignores ambiguous characters by changing them all to the sequence alphabet's wildcard character (e.g. "N" for DNA and RNA, or "X" for protein). STREME will not include any portion of a sequence that overlaps a wildcard in its search for motifs.