Large Scale Analysis of Proteomic Data

1
Large Scale Analysis of Proteomic Data

• A sample of the research in the Linal group

2
Three Sample Projects

• ProtoNet Automatic Clustering of Protein
  Sequences
• EVEREST Automatic Identification and
  Classification of Protein Domains
• PANDORA Annotation Based Analysis of Protein
  Sets

3
Protein classification - The Challenges

• We have
• MANY new Genomes Sequences - almost 250 complete
  genomes
• Growing very fast
• We want
• Sequence to function inference
• New discoveries (new families, hypothetical
  proteins, new connection of remote families..)

4
Clustering by Pairwise Similarity

• Our Approach
• Use sequence similarity data to define
  (hierarchical) protein families
• Validate our families
• Infer function / 3d structure of unknown proteins
  from annotation of proteins in family
• Identify families with no annotation as new
  families (candidates for further research)

5
Transitivity of Sequence Similarity

• Basic tool pairwise sequence similarity
• Limited range
• Transitivity increases range
• Pitfalls
• Amplification of noise
• False transitivity

6
Amplification of Noise

• Assume a simple world were proteins are divided
  into exact families.
• Our sequence similarity tools are noisy.
• Some edges are missed, some edges that should not
  be there are added.

• We use transitivity to recover lost edges.
• Careless use of transitivity would create one
  large cluster

7
ProtoNet Clustering Method

• First, each protein is considered a singleton (a
  cluster of its own).

8
ProtoNet Clustering Method

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

9
ProtoNet Clustering Method

• The clustering process gradually generates a tree
  of clusters

10
ProtoNet Clustering Method

• Lastly, we choose a subset of the clusters in the
  tree.

11
ProtoNet - Evaluation

• Evaluate vs. an annotation system (e.g. InterPro)
• Cluster correspondence w.r.t. a specific keyword
• TP proteins with keyword in cluster
• FN proteins with keyword out of cluster
• FP annotated proteins without keyword in
  cluster
• ScoreTP/(TPFPFN)
• For each keyword, we find the highest scoring
  cluster
• We want minimal number of clusters covering all
  keywords

12
ProtoNet - Evaluation

• 28000 clusters

13
ProtoNet - Example
14
ProtoNet Web - in brief
15
Protein card
16
Cluster card
Automatically added A NAME for a cluster (when
coherent with its annotations)
17
EVEREST Domain Classification

• Proteins are composed of domains
• Lets identify these domains and classify them
  instead of the whole sequence

C transcriptional terminal regulatory domain
Response regulator receiver domain
luxR family
Autoinducer binding domain
18
EVEREST - process
19
EVEREST - evaluation

• Testing for Domains gt more FP

• Coverage Proportion of known domain families
  reconstructed well
• Accuracy Proportion of clusters that be well
  explained by known families (excluding new)

20
PANDORA Analysing Annotations of Protein Sets

• How does one understand a cluster of proteins?
• Look at the annotations of those proteins
• Say you have 100 proteins
• Say I tell you that
• 60 of them are annotated as membrane proteins
• 40 of them are annotated as enzymes
• Is that enough information?

21
PANDORA
60 membrane
40 enzyme
?
?
?
membrane
enzyme
membrane
membrane
enzyme
enzyme
22
PANDORA

• Given
• set of proteins (e.g. ProtoNet cluster)
• keywords annotating them
• Let each combination of keywords define a subset
  of the proteins
• Show
• A graph of inclusion and intersection of the
  protein subsets defined by the keywords

23
PANDORA - example

• Take set of all 576 proteins annotated by GO
  molecular function as anion channel.
• Analyze their InterPro signatures

24
BASIC SET
InterPro
Number of proteins
Sensitivity TP/(TPFN) red FN white TP
25
InterPro
GABA A receptor
gamma subunit
alpha subunit
beta subunit
26
Summary

• ProtoNet EVEREST
• Automatic classification of large databases of
  proteins sequences (protein domains)
• Find new members of known families
• Interactions between families
• New families
• PANDORA
• Analyze annotations of a set of proteins
• Understand clusters
• Understand experiment results (e.g. expression
  chips)
• Curate annotations