Optimatization of a New Score Function for the Detection of Remote Homologs

1
Optimatization of a New Score Function for the
Detection of Remote Homologs

• Kann et al

2
Introduction

• New method to calculate a score function, aiming
  to optimize the ability to discriminate between
  homologs and non-homologs
• Existing software uses the following to compute
  an alignment score

3
Number of times AA i is aligned with AA j
Number of gaps in alignment
Number of residues in each gap beyond one
Score function / Substitution matrix Contribution
to score for AA match/mismatch
Contribution to score for gap initialization
Contribution to score for gap extension
4
Current Methods to Calculate Homology

• p(Sr gt x) probability that a random pair of
  proteins of the same length would have that score
• E expected number of random proteins in the db
  that would have at least that score
• P probability that there is at least one random
  pair with a higher score
• As p(Sr gt x), E, P increase, the likelihood that
  the given pair is homologous decreases

5
Current Score Matrices

• PAM (percent accepted mutations) Dayhoff
• GCB, JTT used to apply to larger sequence
  datasets
• BLOSUM62 Henikoff Henikoff, constructed using
  a dataset of aligned sequence blocks
• STR protein sequences aligned based on their
  observed structures

6
Limitations of Current Score Functions

• Current score functions assume independent
  evolution of each location, overlooking
  correlations
• Score functions derived from a db of properly
  aligned proteins, not on alignments between
  random sequences
• Gap penalty a priori

7
Theory

• Z score for alignment
• Characterize the significance of alignment score
  by calculating the likelihood that this score or
  higher would be obtained by a random match
• Account for variations in E with the length of
  the proteins

8
Theory

• Score function optimized by maximizing the
  confidence ltCgt over the training set
• Avoids dependence on extreme E values (easily
  detected or overly distant homologies)
• Eliminates contribution of falsely identified
  homologies (overly distant)

9
Database Preparation

• Use set of known homologs whose homology cannot
  be reliably determined with standard pairwise
  comparison, in order to optimize score function
  for detection of distant homologs
• Training set 900 pairs of protein in same COG
  with lt 25 sequence identity

10
Optimization of Score Function

• Align using BLOSOM62 matrix
• Calculate Z and C for each pair of homologs, then
  averaged over pairs in training set to yield ltCgt
• Generate initial alignments using gap penalties
  that yielded highest C values
• 10 cycles of optimization and realignments until
  score function converged

11
Results

• Small changes in gap penalties most of the
  improvement cones from refinements of
• OPTIMA resulting score function
• has significantly improved average confidence ltCgt
  value compared with other score matrices
• ltp(Sr gt x)gt, ltPgt significantly decreased

12
Summary

• Aim optimize score matrix to discriminate
  between homologs and non-homologs
• OPTIMA score function more successful at
  discriminating between homologs and non-homologs
  compared with standard score matrices
• Gap penalties treated as additional parameters to
  be optimized