cisGreedy Motif Finder for Cistematic

1
cisGreedy Motif Finder for Cistematic

• Sarah Aerni
• Mentors Ali Mortazavi
• Barbara Wold

2
cisGreedy

• De novo motif finder which implements a greedy
  algorithm similar to Consensus motif finder
• Goal To provide an efficient algorithm to be
  included in the Cistematic package that performs
  similarly to Consensus and meme

3
Cistematic

• Integrate visualization, refinement of motifs and
  improve performance of multiple motif finders in
  a single package

Mortazavi, 2006
4
Cistematic

• cisGreedy becomes part of Bottom Tier
• Motif finder would be included in the Cistematic
  package (prevents need for complicated
  installations)

Image Ali Mortazavi
5
What is a Motif?

• cis-Regulatory elements
• Transcription Factor Binding Sites(TFBS)
• Binding by transcription factors may increase or
  decrease transcription of genes

6
What is a Motif?

• GAL4 in Yeast
• Activator of galactose-induced genes (convert
  galactose to glucose)
• Protein structure determines motif
• DNA-protein interactions require certain bases at
  specified locations
• Motif reflects homodimer structure

7
What is a Motif?

• cis-Regulatory elements
• Transcription Factor Binding Sites(TFBS)
• Binding by transcription factors may increase or
  decrease transcription of genes
• Gene Regulation believed to be a major source of
  complexity
• Plants may have more genes or larger genomes than
  humans are they more complex?
• Identification of cis-regulatory elements will
  help us understand gene regulatory networks
  (bigger picture)

8
How do we find motifs?

• Hard to identify
• Relatively short sequences (as small as 6 bases)
• Many positions not well conserved
• Factors improving identification
• Usually localized in certain proximity of a gene
  (search within 3 kb upstream)
• Some positions highly conserved
• Use other data (Microarray?)

9
Motif Finders

• Greedy
• Maximizes similarity of motifs from sequences
  through a greedy approach
• Eliminate background modeling by using Cistematic
  package preprocessing steps
• Improves speed
• Prevents false negatives
• Implements multiple models (zoops, oops, TCM)

10
Consensus Scoring

• Use equation similar to log likelihood called
  Information Content

L columns in the matrix A A,C,G,T
frequency of each letter i at each
position j a priori probability of letter
i
Hertz, Gerald Z., and Gary D. Stormo.
"Identifying DNA and protein patterns with
statistically significant alignments of multiple
sequences." Bioinformatics 8 1999 563-577.
11
Removing Background

• Goal of a background model differentiate noise
  from signal
• Issues with background
• What background should be used?
• Whole genome? Conserved regions?
• Selective pressures maintain conserved regions
• Arguably searching in conserved regions
  guarantees there is little noise (it has been
  maintained)
• Solution
• Search in conserved regions
• Use simple repeat masking
• Sequences which reoccur are likely TFBS

12
cisGreedy scoring

• Scoring focuses on maximizing number of identical
  bases
• Percent identity is dependent on number of
  deviations from the strict consensus
• Background adds complexity that may lead to false
  negatives

13
cisGreedy

• Input sequences are analyzed
• Randomly select 2 sequences to be compared

14
cisGreedy

• The two selected sequences are analyzed
  independently of the remaining sequences

15
cisGreedy

• The two selected sequences are analyzed
  independently of the remaining sequences
• Windows of motif size are scanned starting at the
  beginning of each sequence

16
cisGreedy

• Sequences are scanned in an attempt to locate the
  highest scoring alignment
• Alignments are ungapped
• Score is established as the number of sequences
  containing the most frequently occurring base at
  each position

17
cisGreedy

• Reverse Complements are analyzed (user specified)
• Once start locations are established with a top
  alignment score, these are left unchanged (Greedy)

18
cisGreedy

• Select an additional sequence in which to
  identify the location of the motif
• Windows in the additional sequence are aligned to
  previously established windows (Greedy)

19
cisGreedy

• Additional sequence scanned as before, reverse
  complement (user specified)
• Alignment score established as before

20
cisGreedy

• Final motif locations are used in order to build
  position specific frequency matrices
• Reverse complement sequence used in building PSFM
  if used

21
Testing cis-Greedy

• AIY
• 16bp cis-regulatory motif drives expression
• Experimentally verified
• Gene battery consists of a set of genes bound by
  AIY
• Orthologous genes contain highly specified
  binding sites
• Individual binding sites of battery genes within
  a single species can vary considerably

(Wenick and Hobert 759)
22
Cistematic Results for AIY
regions of conservation
orthologous genes
hen-1
hen-1
23
Results for AIY

• AIY Identified AAATTGGCTTCCTCAAA
• cisGreedy TTTGAGGAAGCCAATTT
• (reverse comp) AAATTGGCTTCCTCAAA
• meme AAATTGGCTTCCTCAAA

AIY- Battery Consensus
24
Cistematic Results for AIY
hen-1
25
Results for AIY
hen-1
26
Tompa Bakeoff

• 3 benchmark datasets
• Real
• Markov Chain
• Generic
• 4 organisms
• Human
• Mouse
• Fruitfly
• Yeast
• Each dataset contains 0-1 motifs.
• Each sequence can have 0 or multiple motifs
• Report 0-1 motif per dataset and locations of
  motifs
• Use statistical tools provided by bakeoff to
  analyze runs

27
Bakeoff example (hm03)

• Identify most reasonable motif based in each
  dataset independently

28
Real
Real
Interesting pattern appears between 3 of 10
sequences
29
Markov
30
Generic
31
Bakeoff example (hm03)

• Identify most reasonable motif based in each
  dataset independently
• Determine which motif appears most reasonable
  across 3 benchmarks and map motif in sequences
  using Cistematic
• Compare results to actual locations (provided in
  bakeoff package)

32
Solution
33
Real
Real
34
Solution
35
Markov
36
Bakeoff results

• Correlation Coefficient
• nCC (nTP nTN - nFN nFP) /
  v((nTPnFN)(nTNnFP)(nTPnFP)(nTNnFN))
• Sensitivity (fraction of known sites that are
  predicted)
• sSn sTP / (sTP sFN)
• Positive Predictive Value (fraction of predicted
  sites that are known.)
• sPPV sTP / (sTP sFP)

37
Bakeoff results

• cisGreedy overall 7th best performer (excluding
  those with no data)
• Overall top performer in fly
• Worst performer in yeast
• 3rd worst performer in mouse
• 4th best performer in human

38
Bakeoff results
Adapted from Tompa, 2005
When running programs in parallel, correlation of
motif finder results to true binding sites
improves
39
Future goals

• Complete analysis of results for cisGreedy using
  benchmarks established by Tompa paper (Nature
  Biotech, 2005)
• Document results and algorithm development
• Continue improving cisGreedy

40
References

• Bioalgorithms.info
• Jones, Neil C., and Pavel A. Pevzner. An
  Introduction to Bioinformatics Algorithms . MIT
  Press , 2004.
• Hertz, Gerald Z., and Gary D. Stormo.
  "Identifying DNA and protein patterns with
  statistically significant alignments of multiple
  sequences." Bioinformatics 8 1999 563-577.
• Tompa, Martin et al. Assessing computational
  tools for the discovery of transcription factor
  binding sites." Nature Biotechnology January
  2005 137-144.
• Wenick, Adam S., and Oliver Hobert. "Genomic
  cis-Regulatory Architecture and trans-Acting
  Regulators of a Single Interneuron-Specific Gene
  Battery in C. elegans." Developmental Cell
  6(2005) 757-770.
• http//cistematic.caltech.edu

41
Acknowledgements

• Ali Mortazavi
• Barbara Wold
• Wold Lab funding provided by DOE NASA
• Additional funding by NSF NIH
• SoCalBSI faculty, staff and fellow students