SEA: Simple Enrichment Analysis of motifs
Abstract
Motif enrichment algorithms can identify known sequence motifs that are present to a statistically significant degree in DNA, RNA and protein sequences. Databases of such known motifs exist for DNA- and RNA-binding proteins, as well as for many functional protein motifs. The SEA ("Simple Enrichment Analysis") algorithm presented here uses a simple, consistent approach for detecting the enrichment of motifs in DNA, RNA or protein sequences, as well as in sequences using user-defined alphabets. SEA can identify known motifs that are enriched in a single set of input sequences, and can also perform differential motif enrichment analysis when presented with an additional set of control sequences. Using in vivo DNA (ChIP-seq) data as input to SEA, and validating motifs with reference motifs derived from in vitro data, we show that SEA is is faster than three widely-used motif enrichment algorithms (AME, CentriMo and Pscan), while delivering comparable accuracy. We also show that, in contrast to other motif enrichment algorithms, SEA reports accurate estimates of statistical significance. SEA is easy to use via its web server at https://meme-suite.org, and is fully integrated with the widely-used MEME Suite of sequence analysis tools, which can be freely downloaded at the same web site for non-commercial use.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 180
- Organism
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · xhigh effort · 17 Oct 2026 · from the full text of version 1 (24 Aug 2021) · CC BY-ND 4.0
Motif enrichment scored with a held-out threshold and a single test per motif yields p-values close to uniform under the null, where AME and CentriMo come out conservative.
Motif enrichment tools are used to guess which transcription factors or binding proteins act on a sequence set, but their reported significance values are not trustworthy: two popular tools under-call enrichment by an order of magnitude and a third over-calls it. SEA keeps the familiar primary-versus-control design while separating threshold selection from significance estimation, so a single test per motif gives a usable p-value.
Key results
ShownReported significance is close to its nominal meaning on null data, unlike the three comparison tools, which matters because miscalibrated p-values silently drop or invent enriched motifs at a fixed threshold.
On 100 random datasets of 1000 sequences each, SEA p-values stayed within a factor of 2 of the theoretical uniform line in Q-Q plots, for DNA, RNA and protein alphabets. AME and CentriMo were consistently too large, often by over an order of magnitude, while Pscan values were consistently too small, apparently not corrected for multiple testing. · 2.3 SEA accurately estimates motif enrichment…; 4 Discussion
ShownRanking accuracy is on par with established tools on real binding data, so the calibration gain does not come at the cost of finding the right motif.
Across 44 ENCODE ChIP-seq datasets from K562 and Gm12878 cells, each queried with all 843 SELEX motifs of the Jolma compendium, the median rank of the correct motif for the pulled-down factor was 2 for SEA and for AME, with CentriMo and Pscan slightly behind and similar overall rank distributions. · 2.2 Comparison of SEA with other MEA algorithms…; 2.2.1 Accuracy
ShownRanking stays comparable as the input is degraded, with one documented exception at very small peak sets.
Shuffling a growing fraction of sequences to reduce purity, or subsampling the 44 datasets, left the four tools' success rates mostly within each other's error bars, Pscan possibly weaker at high noise. Below 100 input sequences CentriMo ranked the reference motif first more often. Relaxing success to a top-3 rank did not change these patterns. · 2.2.2 Sensitivity
ShownSEA was the fastest of the four tools benchmarked, with Pscan far behind.
Runtimes were measured on the same 44 ChIP-seq experiments, single-threaded on a 3.2 GHz Intel Core i7 with 16GB of memory; the actual timings are plotted in a figure that is not part of this text, so only the ordering is available here. · 2.2.3 Speed; 4 Discussion
Methods, data and reproducibility
Each motif, as a position weight matrix, scores primary and control sequences against a Markov background built from the controls; a random 10% hold-out picks the score threshold and the rest give one Fisher exact or Binomial test per motif. Benchmarking used ENCODE ChIP-seq peaks with SELEX reference motifs, plus random sequences for calibration.
- ENCODE transcription factor ChIP-seq peak files (22 in K562, 22 in Gm12878), cut into 100bp and 500bp hg19 sequences around peak centres, with k-mer preserving shuffled controls · human · n = 44 datasets, 772 to 67,417 sequences each
- Jolma compendium of motifs derived from high-throughput SELEX, used both as the query set and as the independent reference for each pulled-down factor · n = 843 motifs
- Randomly generated DNA, RNA and protein sequences with column-shuffled motifs from the Jolma, CISBP2 and ELM2018 databases, for checking p-value calibration · n = 100 datasets of 1000 sequences; 1000 shuffling repeats
- Methods & tools
- SEA, AME, CentriMo, Pscan, Clover, MEME Suite, fasta-fetch-centered, fasta-shuffle-letters, gendb, meme-shuffle-motifs, CISBP2, ELM2018
- Assay conditions & compute
- Benchmarks ran single-threaded on a 3.2 GHz Intel Core i7 with 16GB memory.
- Reproducibility gaps
- The web server address, the MEME Suite download link, the UCSC location of the ENCODE peak files and the pointer to the motif file are all truncated in this text, so no URL can be given here.
Limitations (5)
- CentriMo ranked the reference motif first more often when the input held fewer than 100 sequences, the one input regime where the authors report another tool being more sensitive than SEA. · stated by the authors
- Clover was left out of the comparison because it was far too slow on large sequence sets, so one of the established tools goes untested here. · stated by the authors
- The authors note the comparison does not cover the settings the other tools were built for, such as AME's use of a per-sequence numeric covariate or Pscan's promoter sets for supported model organisms. · stated by the authors
- The evaluation was designed by the tool's own authors and rests on one motif compendium, one validation assay type and two cell lines, with the correct answer defined by name matching between the pulled-down factor and a SELEX motif. · noted in this summary
- Speed, accuracy and calibration results are all reported through figures that are missing from this text, so the magnitudes behind the comparisons cannot be checked. · noted in this summary
Who should read:Read this if you run motif enrichment on peak sets or promoters and need significance values you can threshold, or if you are choosing between enrichment tools for a pipeline.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- XSTREME: Comprehensive motif analysis of biological sequence datasets
Charles E. Grant, Timothy L. Bailey·bioRxiv·Sep 2, 2021
- STREME: Accurate and versatile sequence motif discovery
Timothy L. Bailey·BioinformaticsPeer-reviewed·Mar 23, 2021
- Factorbook Motif Pipeline: A de novo motif discovery and filtering web server for ChIP-seq peaks
Bong-Hyun Kim, Jiali Zhuang, … Zhiping Weng·bioRxiv·Dec 4, 2015
- Regmex, Motif analysis in ranked lists of sequences
Morten Muhlig Nielsen, Paula Tataru, … Jakob Skou Pedersen·Algorithms for Molecular BiologyPeer-reviewed·Dec 8, 2018
- AmalgaMo: flexible DNA motif merging
Orsolya Lapohos, Gregory J. Fonseca·Bioinformatics AdvancesPeer-reviewed·Jan 1, 2026
- A new sequence logo plot to highlight enrichment and depletion
Kushal K. Dey, Dongyue Xie, Matthew Stephens·BMC BioinformaticsPeer-reviewed·Dec 1, 2018