Proteomics: Analyzing proteins space

1
Proteomics Analyzing proteins space
2
Protein families

• Why proteins?
• Shift of interest from Genomics to Proteomics
• Classification of proteins to groups/families -
  what is it good for?
• Explosion in biological sequence data gt need to
  organize!
• Understanding relations/hierarchy of groups is
  interesting as is, e.g. in evolutionary research.
• For applied research
• Annotation of new proteins predicting their
  function, structure, cellular localization etc.
• Looking for new folds

3
Sequence-based classification

• By sequence similarity (domains, motifs or
  complete proteins) Pfam, PROSITE, SMART,
  InterPro etc.
• InterPro Synthesizes the data from Pfam,
  PROSITE, Prints, ProDom, and SMART. Considered as
  best domain-based classification available

4
Other kinds of classification

• Global classification
• Systers, Protomap, CLUSTr
• MetaFam synthesizes global classification data
• By structure similarity SCOP etc.
• By function Albumin, RetNet, TumorGenes etc.

5
http//www.protonet.cs.huji.ac.il

• A long-term project in HUJI led by Michal Nati
  Linial.
• Provides automatic global classification of the
  known proteins.
• Performs hierarchical clustering on
  sequence-based metric space of proteins.
• Allows to place an external protein into the
  hierarchy.

6
Why clustering?

• We want to refine the similarity notion,
  compared to e.g. BLAST
• Exploit transitivity to improve grouping
• Can use a low threshold on similarity
• - uses vast information from low similarities
• - allowable because clustering filters noise

7
Why hierarchical?
Vertical Perspective
Horizontal Perspective
8
ProtoNet Pre-Computation

• All-against-all gapped BLAST using BLOSUM62
• SwissProt release 40.28 database (114,033
  proteins)
• BLAST identified 2107 relations between these
  proteins with relatively high sequence
  similarity E-Score of 100 or less
• Dont want to lose information gt very
  permissive!
• But still less then 6.5109 gt infeasible

9
Clustering Method

• First, each cluster is considered a singleton

10
Clustering Method

• Next, we iteratively merge the pairs of clusters
• We choose to merge the most similar pair of
  clusters.

11
Clustering Method

• Next, we iteratively merge the pairs of clusters
• We choose to merge the most similar pair of
  clusters.

12
Clustering Method

• Next, we iteratively merge the pairs of clusters
• We choose to merge the most similar pair of
  clusters.

13
Clustering Method

• As we progress the number of singletons drops

14
Clustering Method

• The clustering process gradually generates a tree
  of clusters
• Stop whenever we like

15
How to merge?

• The potential merging score is calculated for
  each pair of clusters relevant for merging at
  each level
• At the bottom equals
• Higher, designed to reflect the similarity of
  clusters.
• Depends on the inter-cluster similarities of
  pairs of proteins, each from a different cluster.

16
Potential Merging Score of

• Arithmetic Mean
• VI
• Geometric Mean
• VI
• Harmonic Mean

17
Missing Data Treatment

• For very low similarity pair (outside of 2107
  ), its length is defined as
• Practically, the merging process should finish,
  when the weight of the infinite lengths in
  calculation of the score between new clusters is
  very large (losing signal)

18
Results ProtoNet top 20

• Why clustering at all?
• We want to extend the range of similarity,
  compared to e.g. BLAST
• Exploit transitivity to improve grouping
• Can use a low threshold on similarity
• - uses vast information from low similarities
• - allowable because clustering filters noise

20 largest clusters in the ProtoNet (Arithmetic)
tree at a preselected level
19
Problem of result assessment what is a good
cluster?

• Contains all proteins in the family, does not
  contain proteins not in family
• But what is family? Does any keyword define a
  family?
• Stable as the merging events occur (long
  life-time)?

20
Problem of result assessment what is a good
tree?

• Should we trust the resulting forest?
• Which clustering technique is better? Combined?
• Bootstrap?
• Do the clusters correspond to meaningful families
  of proteins?
• Validation against InterPro, SCOP etc.
• Lack of will to automatically reconstruct them!!!
• What is the right level/cut to look at the
  forest?

21
Interpro Validation

• Interpro annotation allows systematic validation
  of the generated clustering
• The geometric method exhibits high cluster
  purity
• Corresponds to low FP

22
The Domain Problem

• Many proteins are composed of several domains
• The sequence similarity tools used are therefore
  local in nature
• The score of comparing two sequences is the edit
  distance of the most similar subsequences of them
• This creates a false similarity problem

23
The Modular Nature of Proteins
24
False Transitivity of Local Alignment
We ran BLASTusing default parameters
All these pairwise similarities havebetter than
1e-40 EScore
If we cluster these proteins, assuming
transitivity of local alignment scores, we will
cluster K6A1_MOUSE with MPP3_HUMAN
25
Alternative methods

• Different types of clustering
• Non-binary
• Goal-oriented gt semi-guided
• Graph theory insights
• Non-clustering ways of exploring the space of
  proteins
• Why BLAST E-score???
• Enrichment of the metric using structure