Data representation: techniques and tradeoffs

1
Data representation techniques and trade-offs

• Rob Knight
• Dept. Chem. Biochem.
• CU Boulder

2
Two conflicting goals for representing data

• Machine-readable
• Dot-bracket (structures)
• Array of chars (alignments)
• Allows standardizationacross algorithms,
  high-throughput analyses, but
• Difficult to relate to expertknowledge

WTF?
3
Two conflicting goals for representing data

• Human-readable
• Secondary structure pictures
• Mutated alignments basedon hand-crafted rules
• Prohibits standardizationacross algorithms,
  high-throughput analyses, but
• Allows efficient exploitationof expert knowledge

WTF?
4
Cant we all just get along?

• NO - inherent conflicts in representation goals
• Need to either
• (a) find middle ground that is acceptable for
  bothpurposes, or
• (b) agree to separate representations that cannot
  be interconverted without substantial manual
  intervention and/or loss of data

5
Example 1 motif representation

• Conflicting goals
• Want human-readable, unique specification of
  motif
• Want to be able to recapture arbitrary
  interactions among bases in the motif

6
What are we prepared to give up?

• Example kink-turn motif and reverse-kink-turn
  motif
• If we discard out-of-order interactions, can use
  something like
• GCWWGASHGASHAGHS,GAAGCWWCGWW

Leontis et al. 2006
7
What are we prepared to give up?

• but if we include the long-range interactions,
  we must number the bases and include arbitrary
  pairs
• GiCj8WWGi1Aj7SHGi2Aj6SHAi3Gj5HS,GAAGi6
  Cj1WWCi7GjWW-Ai5Gj2SSGi6Aj6SS
• Not really human-readable or machine-readable!
• (problem representing arbitrary graphs is hard)

8
So whats the solution?

• Two proposals
• Tiered system of less complex - more complex
  linear representations depending on the type of
  motif (think of chemical nomenclature for
  substituents or cycles)
• Use common names but require deposition of formal
  motif definition with unique accession in
  central db

j
i
9
Advantages and disadvantages

• Tiered system
• More human-readable
• Difficult to parse (unless readability is
  expended),liable to incomplete or ambiguous
  specifications
• Probably wont be able to do text search anyway
  because of journal formatting
• Accession system
• Easy to parse (store machine-readable connect
  list, incl. ambiguities) so can automate
  analyses
• Can generate human-readable diagrams as output
• Can generate specification using graphical tools
  so need not require familiarity with the file
  format
• Requires central repository and enforcement of
  deposition
• Question is the community prepared to reify and
  enforce the current motif nomenclature?

Leontis et al. 2006RNA
10
Example 2 homology

• Fundamental problem systems that are homologous
  at one level are not necessarily homologous at
  other levels
• E.g. bat wings and bird wings homologous as
  pentadactyl limbs, but not homologous as wings
• Homology is hierarchical andcan partially
  overlap at any level(e.g. Griffiths 2006)

Bat forelimbs
Bird forelimbs
Frog forelimbs
Rodent forelimbs
Mammal forelimbs
Tetrapod forelimbs
Ridley Evolution 3rd ed.
11
Is one alignment enough?

• Example Lorsch Szostak (1994) evolution of
  polynucleotide kinase from ATP aptamer (15 mut.)
• Some recovered classes retained sequence
  similarity to the ATP aptamer, but formed new
  active sites that ignored thissimilarity
• Therefore, aligning thefunctional regions
  isincompatible withaligning the most
  similarsequences -- need diff.alignments for
  ancestryand functionality

12
Multi-level alignments

• Can align aptamer to Class IV and Class V by
  sequence alone, but cannot align to the others
  without structural info (except in small regions)
• General problem (for othersequences, not shown
  here)coarse-grained structurevaries too, so
  cannotmatch up the right stemsand loops w/o
  sequenceinformation

S1
B1
S2
S2
B2
S1
13
Solution iterative approach?

• Need alignment to producereasonable tree and
  structure
• Need structure and tree toproduce
  reasonablealignment
• Need sequence sim.to anchor structuresim. b/c
  structure changes
• Main drawback tocurrent techniquesassume that
  all partsof sequence areequally important for
  aln!

14
Conclusions

• Goal in both cases should be to connect expert
  knowledge with automated approaches -- too big a
  gap at present
• Motifs central database with accessions has many
  advantages, but will the community support it?
• Alignments probably need to move away from the
  annotate one alignment model towards many
  alignments for the same set of sequences
  depending on task model -- needs hierarchical
  view of homology, new techniques for connecting
  levels (not clustal!)