Identifying functional residues of proteins from sequence info

1

• Identifying functional residues of proteins from
  sequence info
• Using MSA (multiple sequence alignment)
• - search for remote homologs using HMMs or
  profiles
• Remote homologs with no known structure
• Given a large, diverse superfamily
• protein may evolve different function or subtype
• different substrate specificity or activity
• proteins with similar fold but different
  function
• Past methods used phylogenetic trees
• map unknown protein to one of the branches of
  the tree produced
• but- maybe diverged to long ago to be clearly
  identified
• co-evolution of multiple features
• possible convergent evolution of molecular
  function at aa level

2

• Other methodologies
• Analysis/prediction of subtype from sequence
  alignments
• characterization of aa residues, looking for
  significant substitutions
• gathering sequences into subgroups, comparing
  each subgroup
• Principal component analysis (Casari et al, 1995)
• looks for functional residues conserved in
  protein families
• Evolutionary Trace (Lichtarge et al)
• Phylogenetic Inference (Sjolander et al)

3

• Goal identify regions conferring sub-family
  specificity
• Secondary goal predict subtypes of orphan
  sequences
• Input to algorithm
• multiple sequence alignment (MSA) of sequences
  in a protein family
• classification of subfamilies of sequences from
  above MSA
• For the given subtypes (or subfamilies) provided
• get the MSA subalignment for each subfamily
• build a HMM profile for each sub-family MSA
• Rationale generate pseudocounts and account for
  statistical bias
• For each subalignment profile
• The profile value for amino acid x at position i
  for subfamily j over all amino acids at a given
  position will sum to 1. (probability of finding
  an amino acid x at position i in the subfamily j)

4

• Relative Entropy
• measure of distance between two probability
  distributions
• Relative entropy produces a value gt 0. (value
  of 0 for two identical distributions)
• for each position i in a subfamily s
• For each position, a RE value for a subfamily s
  vs s-bar (all other subfamilies)
• Cumulative Relative Entropy
• given a set of relative entropies for each
  subfamily for each position
• To produce a CRE for a given position i in the
  MSA across all subfamilies.

5

• Given this set of cumulative relative entropy
  measures
• one for each position in MRA- you take the Z
  score.
• Standard statistical measure- the number of std
  devs above/below the mean
• tells you which residue positions vary strongly
  in aa distribution between families
• empirically, Z gt 3 correlates with functional
  residue
• For position i, which amino acid is dominant in a
  given subfamily
• find probability of observing aa x at position
  in subfamily s vs not-s

6
Subfamily data

• What exactly constitutes a family or subfamily?
• not always clear
• automated tree generation could not separate
  data into clear subfamilies
• use of PFAM alignments and SWISSPROT data
• Subfamilies are not clearly defined in databases
• divided proteins from PFAM database into
  subfamilies based on SWISSPROT data
• keyword search limited to enzymatic activity
  string in SWISSPROT
• put into groups, then checked for obvious
  mistakes
• also eliminated divisions easily discernable by
  sequence comparison
• 62 groupings from 42 alignments remained
• randomly pick 11 to produce 42 groups over 42
  alignments

7
Subfamilies

• Four very large families to test their results on
• nucleotidyl cyclases
• eukaryotic protein kinases
• lactate/malate dehydrogenases
• trypsin-like serine proteases
• Nucleotidyl cyclases
• membrane-attached or cytosolic, cyclize (GTP -gt
  cGMP) or (ATP -gt cAMP)
• found residues 1018, 938, which correlate with
  previous results
• also identified residues which have not been
  tested experimentally
• Protein kinases
• phosphorylate serine/threonine or tyrosine
  residues
• compare to experimental result- some ser/thr vs
  tyr kinase differences not detected
• inconsistency (no conservation) within the
  subfamily
• residues which were common to both ser/thr and
  tyr kinases

8
Subfamilies (cont)

• Lactate/Malate Dehydrogenases
• common to a very wide variety of organisms-
  highly divergent
• results mostly as expected- but a few residues
  identified outside of active site
• Serine Proteases
• cut protein backbone- differing specificity as
  to where (what aa precedes cut)
• specificity pocket determines where protease can
  bind
• identified 2 out of 3 of experimentally-determine
  d pocket residues
• (third had a low z-score because of tolerance in
  one protein family)
• also identified a few residues outside of the
  active site

9
Prediction of Protein Subfamily

• Sequence Similarity
• straight similarity with other sequences
  (ignoring gaps)
• BLAST
• database search, assign to nearest subfamily
  with best alignment
• HMM method
• align sequence of sub-type to all HMMs of
  subfamilies and assign it to best alignment
• will attempt to do iterative optimization of
  match
• Profile method
• take original HMM, and probability profile
• Sub-profile method
• only use residues in above formula that have a
  positive Z-score
• to reduce noise, restrict to values that have
  above average positive relative entropy

10
Casari, et al. (1995) A method to predict
functional residues in proteins

• Input a multiple-sequence alignment
• each sequence is converted to a vector of size
  (20 l) where l is length of the alignment
• Generation of of N x (20l) matrix
• one sequence produces a vector of dimensions
  20l
• N sequences to produce N vectors of dimension
  20l
• Use Principal Component Analysis
• get the covariance matrix- tells you how factors
  are correlated to one another
• eliminate covariance by finding
  eigenvectors/eigenvalues of covariance matrix
• largest eigenvalues and corresponding
  eigenvectors give you principal components
• ie the largest factors determining distribution
  of your dataset
• they take the three largest (the largest of
  which represents consensus sequence)
• project their 20l dimensional data onto those 3
  dimensions
• this can be used to predict a protein subfamily
  for a given protein

11
General Weirdness

• Construction of a comparison matrix
• take matrix x (matrix transpose)
• solve for eigenvectors and eigenvalues as before
• Columns of f represent amino acid values and
  positions
• becomes possible to examine individual amino
  acid residues and positions
• plotted on graph, shows residue correlation to
  type of protein subfamily
• does this actually work?