GECCO Report

1
GECCO Report Some Issues about TFBSs

• Tak-Ming Chan
• August 2 2007

2
GECCO Report

• Genetic and Evolutionary Computation Conference
  (GECCO) 2007
• Submissions totaled 577, Total accepted full
  papers 266Accepted as posters (abstracts) 210

3
GECCO07

• Tutorials Workshops (2 days)
• Introductory GA, GP, EDA,
• Advanced Representations, Fitness landscapes,
  Problem hardness,
• Specialized Open-source software,
  Bioinformatics, Complex Networks,
• Presentations (3 days)
• 14 tracks GA(109/43), GP(54/27), ES/EP(21/11),
  Real-World Applications (99/48), Biological
  Applications (25/10 ),
• All sessions go simultaneously

4
GECCO07

• Keynote Event
• Public Debate on Complexity and Evolution
• with Richard Dawkins (The Selfish Gene), Lewis
  Wolpert and Steve Jones (Almost Like a Whale)
• Different view from the biologists about
  evolution, constraints, complexity
• Audio download

http//www.cs.ucl.ac.uk/staff/p.bentley/evodebate.
html
5
Digest of some Presentations

• Real-World Applications
• A robust GP solution for hedge fund stock
  selection 1
• Robustness for volatile and extreme scenarios
• Fitness mean profit over the volatility (std.)
  of the scenario
• Genetic Algorithms
• Simple diversity mechanisms analysis 2
• Different diversities (in two ( 1) EAs)
  suitable for different problems (two plateau
  functions)
• Genotype diversity better in a multimodal
  problem
• Phenotype diversity better in a needle in a
  haystack problem

6
Digest of some Presentations

• Biological Applications
• Prof. Congdon, the chair of my session had a
  conversation with me
• GAMI (consensus-led) addressed the TFBS
  identification in a multimodal way
  (unintentionally)
• In further work, they found Information Content
  was not ideal as expected for their datasets
• Check it out for more in
• 1 W. Yan, C. D. Clack, Evolving Robust GP
  Solutions for Hedge Fund Stock Selection in
  Emerging Markets, Proceedings of GECCO 07,
  pp2234-2241
• 2 T. Friedrich, N. Hebbinghaus, F. Neumann,
  Rigorous Analyses of Simple Diversity Mechanisms,
  Proceedings of GECCO 07, pp1219-1225

7
A little bit more about London

• Tired of the Big Ben, Tower Bridge, Buckingham
  Palace?
• Try the Wellcome Collection near UCL
• Especially good for seeing some interesting
  things about Medicine and Genetics

8
Some Issues about TFBSs

• Information Content (IC)
• Similarity measures between PWMs

Regulates Gene Expression
Transcription
Transcription Factor
TFBS
Gene
9
Evaluation of IC as a Metric for TFBS
Identification 3

• Incorporated IC in GAMI which originally employed
  Match Count (MC) (by Congdon et al)
• Expected IC should be more accurate than MC
  but it turned out to be not
• IC missed some 100 conservation regions while MC
  did not
• Several possible problems of IC addressed
• Background frequencies
• Different IC scores to a motif and the reverse
  complement
• Synonyms problem

3 An Evaluation of Information Content as a
Metric for the Inference of Putative Conserved
Noncoding Regions in DNA Sequences Using a
Genetic Algorithms Approach, in 2006 IEEE
TRANSACTIONS ON COMPUTATIONAL BIOLOGY AND
BIOINFORMATICS
10
Scores for MC and IC

• Rough correlation different peaks

11
Background frequencies

• IC has a very strong preference for CG-rich
  regions for the dataset (SOX21)
• The background frequencies
• However, the areas of conservation are not always
  CG-rich

12
Background frequencies
IS an issue!!

• IC is not suitable for a strongly biased region
• Li Gang raised the similar question (a 2-letter
  case for simplification)
• log(0.9/0.8) 0.9 log(0.1/0.2) 0.1 0.0367
  log(0.5/0.8) 0.5 log(0.5/0.2) 0.5 0.2231
• IC favors the random nucleotides over the
  conserved ones!
• Further questions raised (by Cyrus)
• Is it possible the promoter regions are really
  biased in real-world problem?
• In this paper, the sequences in the datasets
  tested are 8kb-10kb long what about the promoter
  regions?

13
Different IC scores to a motif and the reverse
complement
MAY be an issue!!

• Highest scoring motif with IC
• caggcaccactcactgcccc (207.92)
• the reverse complement C is less than G in
  SOX21 ggggcagtgagtggtgcctg (189.73)
• Both score 117 (of a possible 120 (206)) with MC
• Questions
• Should the background sequences be recalculated
  for reverse complement?
• How are the motif instances aligned according to
  this motif (see the next issue of synonyms)?

14
Synonyms problem

• Different motifs correspond to the same IC score
• Have a look at how the motif instances are chosen
  and aligned for IC in this work

High MC motif
Lower-quality synonym
15
Synonyms problem

• The procedure
• First, GAMI (Consensus-led) is performed to
  locate the best motifs based on MC
• One motif may correspond to several instances in
  the sequences (they may together contribute to
  the same MC)
• ATCGATCG ATCGATGG or ATCGAACG
• When IC is tested, it is calculated only based on
  the combinations of these instances (up to 1000
  combinations)

16
Synonyms problem
NOT an issue!!

• The result is the motif instances combination
  may correspond to another consensus but are
  forced to align with the specific motif
• The representational power of consensus
• The inappropriate use of ICit is used for
  position-led representations!

17
Future work concerning IC

• Some of the above issues have to be worked on
• Background frequencieshow to estimate?
• Forward and Backward strands
• Positional background frequencies
• Pseudo-counts
• Large pseudo-counts have pronounced affect on IC,
  especially for small sample size
• Some work is proposed to estimate appropriate
  pseudo-counts
• More
• Context information of the instances
• Additional information

18
Similarity measures between PWMs 4

• Useful in EC for maintaining diversity
• For position-led representation
• Position frequency matrix (PFM)
• Position weight matrix (PWM)

19
D for PFMs

• test for each position

D (stands for the distance) is incremented by 1
20
For PFMs with different widths

• Various shifts are tried to get the minimal D
  (note that overlap is at least 6 nucleotides)

21
Further issues

• To adopt the D for PFMs in our work
• Change the actual counts to normalized
  frequencies for the matrix used in our GA
• C (Correlation Coefficient) for PWMs
• Use a random DNA sequence to compute the sum of
  weights of the PWMs and then use C to measure the
  scores (skipped, not so practical)
• The paper 4 did more than proposing D and C
• Beyond this presentation

4 Measuring similarities between transcription
factor binding sites, Sep 2005 BMC Bioinformatics
(if 3.62)
22
The End

• Thank you very much!
• Q and A?
• Conference Report
• Issues for TFBSs

My footprint at Wellcome Collection Health
is most important