Structural Characterization of Proteins using Residues Environments

1
Structural Characterization of Proteins using
Residues Environments

• Sean D. Mooney, Mike Hsin-Ping Liang, Rob
  DeConde, and Russ B. Altman
• Stanford University, Indiana University
• PROTEINS Structure, Function, and Bioinformatics
• 2005

2
Outline

• Introduction
• Materials and Methods
• Results
• Discussions
• Conclusions

3
Introduction

• A primary challenge for structural genomics is
  the automated functional characterization of
  protein structure.
• Current structural methods for identifying
  function rely on one of the following
• Phylogenetic tree derived from sequence
  similarity (evolutionary trace)
• Hand curated molecular fingerprints (template
  based)
• Fold recognition and alignment methods (structure
  comparison)
• Sequence-based methods for functional
  characterization rely on identifying conserved
  residues within protein structures

4
Introduction (cont)

• It is important to develop sequence-independent
  methods for identifying function to complement
  sequence-based methods when they are limited by
• Lack of sequence similarity
• Small datasets
• Methods for identifying key functional residues,
  or molecular fingerprints, can classify function

5
Introduction (cont)

• Sequence-independent structure-based methods for
  function assignment are challenging for several
  reasons
• Aligning local structure is a difficult
  computational task
• Estimating the statistical significance of the
  results is challenging
• Scanning through the entire protein data bank
  (PDB) can be computational demanding
• Structural similarity and functional similarity
  are not always well correlated

6
Goal

• Develop a method for unsupervised mining of
  structural datasets and automatically identifying
  local regions within protein structures that are
  statistically associated with a given annotation
• Define the most structurally significant residues
  environments for given classification, based on
  the structural environments represented in that
  database

7
Methods

• S-BLEST (Structure-Based Local Environment Search
  Tool)
• Based on the FEATURE representation of a local
  environment
• Rapidly search databases of vectors of local
  structure properties
• S-BLEST method relies on k nearest neighbor (KNN)
  searches using a Manhattan distance metric
• Significance score (z-score)

8
Local Environment

• Local structure environment
• A set of concentric shells extending outward from
  a positions of residues Cß atom
• Glycine residues
• Cß atom position were estimated by determining
  the average of position of a Cß from serine
  protease 1DSU
• It is the simplest of the 20 standard amino
  acids its side chain is a hydrogen atom
• Each shell contains 66 properties
• of atoms associated with a given residue type
• of positively and negatively charge ions
• The van der Waals volume of the shell
• The solvent accessibility
• Four radial boundaries
• 1.875, 3.75, 5.625, 7.5 Å
• Vector dimension 264

glycine1
1DSU
1 http//en.wikipedia.org/wiki/Glycine
9
Materials

• Datasets
• ASTRAL 40 non-redundant structure database
  (ASTRAL 1.65)
• X-ray crystal structures only
• Steps for data cleaning
• Remove all hetero-atoms (PDB HETATM)
• Normalization

10
Identification of Residues Environments
Associated With a Structural of Functional
Annotation

• The performance of each residues can be
  determined by creating a receiver operator
  characteristics (ROC) plot of the ranking
• TP a protein structure that belongs to the same
  SCOP family as the query protein with a z-score
  of greater magnitude than the threshold
• FP a protein structure that doesnt belong to
  the same SCOP family but has a z-score of greater
  magnitude than the threshold
• The AUC of a residue in a query structure of
  known function indicates how well the reside
  environment classifies the SCOP family of the
  structure and can range from 0.0 to 1.0.

11
Congruence Approach for Combining S-BLEST Searches

• Congruence approaches (Shotgun) are a useful way
  to combine several searches to increase
  statistical significant.
• If input is a query with multiple residues (query
  chain)
• Each residue in the query chain --gt most similar
  residues in each dataset chain (z-score
  threshold)
• If there were n resides in the query chain, there
  would be n residues (possibly redundant) in the
  dataset chain that are identified as most similar
  to each of the n residues in the query each with
  a z-score.

12
Identification of Structurally Similar Residue
Environments in ASTRAL 40 v1.65

• ASTRAL 40 v1.65 encoded 4129 crystallographically
  determined structures

distance histogram distribution (2TRXA)
13
Identification of the Residue Environments
Associated With a Structural Class
1DI9A
AUC
Yellow ATP binding site, Gray peptide binding
channel, Red Phosphorylated(???) Residue
14
ROC of the Ranked Chains Outputted from the
Congruence Approach
Of the 27 members in our dataset, the first 25
chains ranked were true positives, whereas the
method failed to recognize 1KOA and 1FMK as
structurally similar (AUC is 0.935).
15
Congruence Approach to Characterize Protein
Structures

• Goal - to show that S-BLEST finds structurally
  similar environments with potential implications
  for fold, family, and function.
• Select 100 random SCOP families in ASTRAL 40
• X-ray crystallographic structures only
• Z-score threshold for each protein is -5.5

16
ASTRAL 40 v1.65 (100 random members)
PPV Positive Predictive Value
z-score
17
Analysis of Uncharacterized PDB Structures

• 86 structures
• 86 of these structures had no significant hits
  when searched against the PDB using BLAST with
  e-value cutoff of 1e-4
• How to obtain these 86 structures search PDB
  for the phrase unknown function

18
Hit Results from the 86 Structures with Unknown
Function (1/5)
True Positive
1VGYA
1LFWA
SCOP C.56.5.4
ARG97 ARG115, HIS68 HIS87, ASP70 ASP89,
GLY98 GLY112, GLU136 GLU154 z-score -6.36,
e-value 3x10-4
19
Hit Results from the 86 Structures with Unknown
Function (2/5)
1VGYA
AUC
20
Hit Results from the 86 Structures with Unknown
Function (3/5)
Of the five true positives in our dataset, three
were the top hits, the fourth was in position
five, and the fifth was ranked 65th overall (AUC
is 0.995)
21
Hit Results from the 86 Structures with Unknown
Function (4/5)
Questionable Significant
1B3UA
1OYZA
Clearly structurally related, and the best
residues matches occur between SSEs, and are
often observed bridging the structural
elements z-score -5.21
22
Hit Results from the 86 Structures with Unknown
Function (5/5)
Possible Unknown Hit
1LJOA
1B34A
Proteins share the same fold, but their
functional relationship is not known z-score
-5.64 e-value 1e-7
23
Discussions

• This method is intended to identify statistically
  significant environments in protein structures
  and will be complementary to both sequence-based
  methods such as BLAST or HMMs and fold
  recognition methods
• Analysis of a random member from each of 100
  random families (SCOP)
• S-BLEST (threshold of -5.1, z-score) finds 28
  SCOP family members that BLAST (threshold of
  1e-5, E-value) does not find
• BLAST finds 89 family members that S-BLEST does
  not find
• Local structural variability between the proteins
• However, of 66 false-family positives, all but 13
  of which share the superfamily of the query
• for each BLAST hit, the degree of structural
  conservation of each residue environment can be
  easily determined using S-BLEST.
• Enzyme
• many residues that were annotated as being
  important for enzyme chemistry are not the ones
  that are most useful for recognizing structural
  similarities.
• The method sometimes does not select the critical
  residues (such as the catalytic triad) likely
  because the environments around those residues
  are structurally variable between members.

24
Conclusions

• We developed S-BLEST to meet a need for rapidly
  identifying similar structures to a query protein
  using local structural environment.
• S-BLEST identifies constellations of structurally
  similar residues between the query protein and
  the full database of known protein structures.
• We found that many of the structural environments
  in SCOP have statistically significant local
  environment neighbors.
• S-BLEST was able to associated 20 proteins with
  at least one local structure neighbor and
  identify the amino acid environment that are most
  similar between those neighbors.