Author: Jason Weston et., al

1
Protein Ranking From Local to global structure
in protein similarity network

• Author Jason Weston et., al
• PANS
• Presented by Tie Wang

2
Outline

• Introduction
• Background
• Method
• Experiment
• Analysis

3
Introduction

• Pairwise subtle sequence similarities imply
  structural functional and evolutionary relations
  among DNA and protein seqences
• Search biosequences from online database is
  analogous to searching the WWW (search engine
  search the db for query and return a ranked
  list)
• A protein ranking algorithm is presented for
  biosequence query

4
Background

• Early algorithms only focus on pair-wise sequence
  similarity (SW LA search)
• Statistical models use multiple alignments for
  similarity search (profile based, psi-blast)
• Global similarity search can be mapped onto
  protein similarity network.

5
How to perform protein ranking?

• Underlying idea Google ranking
• Key feature Exploiting global structure by
  interring it from local hyperlink structure.
• Construct a protein similarity network
• Add query sequence
• Weight diffusion
• Rank proteins upon convergence

6
Algorithm
7
Experiment

• Use protein 3-D structure database SCOP as golden
  standard.
• Sequences have no more than 95 similarity.
• 7329 proteins are splitted into 379 superfamilies
  as training and 332 for testing
• 3 networks are generated using BLAST and
  PSI-BLAST.

8
Experiment

• Value
• Compare with other two experiments
• 1. only local structure are considered
• 2. non-local edges without weak edges
• The result shows that the second one is only
  slightly worse than our algorithm

Where Sj(i) is E value assigned to protein I
given query j.
9
Analysis
Bower et al, Science vol 306, 2004
Cluster structure
10
Motif based protein ranking by network propagation

• Author Kuang Rui et., al
• Bioinformatics
• Presented by Tie Wang

11
Outline

• Introduction
• Background
• Method
• Experiment
• Analysis

12
Background

• Direct measure of pairwise sequence is proved to
  be effective on classification.
• Performance is dropped down when detecting subtle
  remotely homology sequences.
• Those sequences share a conserved structure at
  least at some components.
• Formulate problem based on this statement.

13
Protein motif bipartite network

• Each protein contains a set of motifs.
• Each motif belongs to a set of proteins.
• Their relationship are mapped to a
• Bipartite graph as shown on the left.
• The edge weight indicates the probi-
• lity that motif x is in protein y.

14
Motifdrop Algorithm

• Set P represents protein sequences and set F
  represents motifs. H is the connectivity matrix.

is row normalized version of H.
is a vector of initial value for H.
is a vector of initial value for P.
15
MotifProp Algorithm

• The convergence of motifdrop is guranteed.
• The problem is reformulated based on the
  following rule,

is row normalized version of H.
is a vector of initial value for H.
is a vector of initial value for P.
16
Edge weighting scheme

• PSI-BLAST E-value is assigned between pair-wise
  protein nodes.
• Gaussian edge weights are calculated.
• The Gaussian weights from query to each protein
  are assigned as initial value.

17
Value estimation

• Sq(i) is the E-value of protein i and query q.
• Eq(j) is the E-value of the jth motif and ith
  protein.

(1)
???
18
Estimation on substitution score

• Substitutions score between a kmer f and sequence
  x can be estimated as,
• where
• and
• sl is a log value which implied the S score
  below threshold can be a motif hits against
  sequence x.

19
Sequential MotifProp

• Empirical experiments suggest that using a
  weighted linear combination of multiple motifs
  does not improve the results.
• Apply a simple multiple motif sets scheme.
• Motif nodes F can be divided into n set partition
  
• in which F(i) is
  a set of motif from ith motif set.
• F set represents the motifs instead of individual
  ones.

20
Motif-rich regions
21
Experiments

• 7329 protein domains with known 3D structure on
  SCOP.
• They are divided into training (4246) and testing
  (3083).
• Apply additional 10602 from swiss-prot db.
• Evaluation on ROC curve.

22
Results of classification
23
Results of classification (cont)
24
Results on Motif rich region
25
Conclusion

• Two methods are presented on protein
  classification using protein ranking methods.
• Similarity matrix and protein/motif propagation
  network are base structures.
• Simple methods but innovative formulation.
• Better results compared with current approaches.
• Analysis on results play an important roles.