MEME: Multiple Expectation Maximization for Motif Elicitation

1
MEME Multiple Expectation Maximization for Motif
Elicitation

• Quinn S. Lewis
• Bioinformatics (COSC6371)
• 28 November 2006

2
Problem
A motif is a pattern common to a set of nucleic
or amino acid subsequences which share some
biological property of interest (such as being
DNA binding sites for a regulatory protein).

• In biology terms
• Identify and characterize shared motifs in a set
  of unaligned genetic or protein sequences
• Only consider contiguous motifs (i.e. insertions
  or deletions are not allowed but appearances of a
  motif may differ in point mutations)
• In computer science terms
• Given a set of strings, find a set of
  non-overlapping approximately matching substrings
• Approximately matching substrings must all have
  the same length

3
A Solution

• Discover (conserved) motifs in a group of
  unaligned and related sequences (DNA or protein)
• Automatically choose the following (with little
  or no prior knowledge)
• Best width of motifs
• Number of occurrences in each sequence
• Composition of each motif

4
Uses

• Find similar biological function and structure in
  sequences
• Sequence variation can be significant
• Motifs sometimes small (6-8 base pairs)
• e.g. Sequence-specific binding sites for proteins
• Reduce unnecessary wetlab experimentation

5
Context

• Local multiple sequence alignment (MSA)
• As opposed to global MSA (e.g. CLUSTAL, T-COFFEE)
• Statistical method
• As opposed to profile or block analysis which
  depends on first producing a global MSA
• Unsupervised learning algorithm
• As opposed to supervised learning which requires
  human intervention

The word meme was coined in 1976 by Richard
Dawkins, author of The Selfish Gene, and refers
to a unit of cultural information transferable
from one mind to another. A meme propagates
itself as a unit of cultural evolution and
diffusion analogous in many ways to the
behavior of the gene. --Wikipedia
6
Algorithm Components

• Expectation maximization (EM)
• EM-based heuristic for choosing the starting
  point for EM
• Maximum likelihood ratio-based (LRT-based)
  heuristic for determining the best number of
  model free parameters
• Multi-start for searching over possible motif
  widths
• Greedy search for finding multiple motifs

Another explanation for the name MEME it is a
greedy algorithma me! me! algorithm.
7
In a perfect world

• If start locations were known, alignment would be
  easy
• Could leverage the fact that no insertions or
  deletions exist in the sequences

8
Types of Possible Motif Models

• OOPS
• One occurrence per sequence of the motif in the
  dataset
• ZOOPS
• Zero or one motif occurrences per dataset
  sequence
• TCM
• Motif to appear any number of times in a sequence
  (two-component mixture)

9
Expectation Maximization

• Expectation step initial guess about the
  location of a (variable) sequence pattern in a
  set of sequences
• Maximization step improve/update pattern as set
  of sequences is iteratively scanned

10
Expectation Maximization Idea
11
Expectation Maximization Algorithm

• dataset - unaligned set of sequences (training
  data) S1, S2, , Si, , Sn each of length L
• W - width of motif
• p - matrix of probabilities that the motif starts
  in position j in Si
• Z - matrix representing the probability of
  character c in column k (the character c will be
  A, C, G, or T for DNA sequences or one of the 20
  protein characters)
• e - epsilon value

12
MEME Algorithm
13
Contributions

• Subsequence derived starting points for EM
• May be useful with other methods
• Saves time (only need to run EM for one iteration
  from each starting point and greedily selecting
  the best starting point based on the likelihood
  of the learned model)
• Little or no prior knowledge requirement
  (unsupervised learning)
• Drops the assumption that each sequence contains
  exactly one appearance of a motif and fit the
  n-per model to a dataset can discover motifs in
  datasets which contain many sequences which do
  not contain the motif
• Erase appearances of the motif found after each
  pass (using the probabilistic weighting scheme)
  finds multiple different motifs and motifs with
  multiple parts

14
Areas for Improvement

• Allowances for gaps and substitutions in the
  conserved regions are not included
• Ability to test significance of the pattern is
  often not included in the analysis
• Erases input data each time a new motif is
  discovered using the assumption that this motif
  is correct
• Limits the model exclusively to the two-component
  case
• Time complexity
• Overly pessimistic about an alignment
• could conceivably lead to missed signals

15
Other Tools

• MAST - http//meme.sdsc.edu
• Uses output of MEME
• Searches biological sequence databases for
  sequences that contain one or more of a group of
  known motifs
• ParaMEME - http//meme.sdsc.edu
• Parallel version of MEME
• Can download run
• Can run from website (http//meme.sdsc.edu)
• MetaMEME - http//metameme.sdsc.edu
• Toolkit for building and using motif-based hidden
  Markov models of DNA and proteins
• Taverna - http//taverna.sourceforge.net
• Uses MEME (among other bioinformatics application
  and data) to create and run workflows