1. Functional Coherence in Domain Interaction Networks
Prof. Ananth Grama

2. Dept. of Computer Science, Purdue University
Outline
Motivation
Protein and Domain Interaction Networks
Formal framework
Properties for term-, set- similarity measures
New Similarity Metric
Results
Comparison of measures
Comparison of PPI, DDI networks
Dept. of Computer Science, Purdue University

3. Dept. of Computer Science, Purdue University
Motivation
Extracting functional information from protein-protein interactions
Noisy, incomplete, generic, static data from high throughput experiments
Typical proteins are composed of multiple domains
Independent unit (function, evolution, folding)
Behind protein-protein interactions there are protein domains interacting physically with one another.
understanding protein interactions at the domain level gives a global view of the protein interaction network and possibly of protein functions p1 d1 d2 d3 d4 p2
Domain-domain interaction
Dept. of Computer Science, Purdue University

4. Dept. of Computer Science, Purdue University
Motivation
How does functional modularity manifests itself in a network of molecular interactions?
Explore relationship between functional similarity and network proximity
Functional annotations available for domains and proteins vastly differ
Do current similarity measures work in unbiased manner? (due to incompleteness of annotation)
Are they statistically meaningful and biologically interpretable?
Annotation for domains is derived from proteins: as such its more general, scarce and incomplete
Dept. of Computer Science, Purdue University

5. Dept. of Computer Science, Purdue University
Formal Framework
C = { ci | 0 ≤ i < N } is a finite partially ordered set of concepts (Ontology).
Concepts are related by binary
relationship, denoted by
eg: c3 c1, c6 c3, c5 r
Set of Ancestors Ai = { ck | ci ck }
Two concepts (ci, cj) are comparable (~) if either ci cj or cj ci
All concepts in Ai may not be comparable as the ontology is a DAG (as opposed to a tree) C3
C0 = r C2 C1 C4 C5 C6
Dept. of Computer Science, Purdue University

6. Properties for term-similarity
Similarity (δ) of two terms based on underlying taxonomical relationship
Existing measures
Distance based:
Count the number of edges between the nodes
δE(ci,cj)=2*MAX-min[len(ci,cj)]
Fails property (4) as distance is uniform over all edges
symmetric.
more specific terms should have at least as much self-similarity as more general terms.
3. a term should not be less similar to itself than to any other term.
4. that terms with more specific common ancestors should be more similar to
each other, compared to those with less specific common ancestors.
Dept. of Computer Science, Purdue University

7. Existing metrics for term-similarity
Information Content:
If Gc be set of molecules associated with concept c, then IC(c) = - log2 (|Gc|/|Gr|)
δR(ci,cj)= max [ -log2 ( c ) ], c Є Ai and c Є Aj (c is common Ancestor)
Normalization:
δL(ci,cj)= 2 * δR(ci,cj) / (IC(ci) + IC(cj))
Hybrid approach:
δJC(ci,cj)= (1 - 2 * δR(ci,cj) + IC(ci) + IC(cj))-1
All three satisfy term-similarity properties
Dept. of Computer Science, Purdue University

8. Properties for set-similarity
Let S be set of concepts, we want a measure ρ(Si, Sj) to access the semantic similarity of two sets
Symmetric
adding a common annotation for two molecules should not decrease the similarity between these two molecules.
if new annotations are added for a new molecule, the similarity of this molecule to any other molecule should not decrease.
a set of annotations should be at least as similar to itself as it is to any other set.
Dept. of Computer Science, Purdue University

9. Existing metrics for set-similarity
Average
Violates properties (ii), (iii) and (iv)
Maximum
Weakly satisfies (ii)
Average of Maximums
Fails properties (ii), (iii) and (iv)
Dept. of Computer Science, Purdue University

10. IC based set similarity
Extend the notion of minimum common ancestor (λ) to sets of terms as
Information content of a set is defined as:
Where is set associated with all terms in MCA of Si, Sj
This satisfies all 4 properties, can be extended
and
Dept. of Computer Science, Purdue University

11. Dept. of Computer Science, Purdue University
Datasets
Protein-Protein interactions
Extract physical interactions from BioGRID database
Binary data (no reliability score)
Domain-Domain interactions
DOMINE database
Confidence score used to split dataset
Struct: Only structure based interactions
HC+NA : High Confidence (HC) and Structure based interactions
HC+MC : High Confidence (HC) and Medium Confidence (MC) interactions
Comp-2: Interactions predicted by at least two computational approaches
Comp-1: Interactions predicted by at least one computational approach
Dept. of Computer Science, Purdue University

12. Comparison of Semantic Similarity Measures
Negative relation between network distance and functional similarity
The proposed information content based measure (ρJC) provides the sharpest decline in semantic similarity for distance<4
For each network, we compute the distance between all pairs of molecules (proteins or domains) in the network. Then, we group molecule pairs according to their distance and compute the average semantic similarity for each group.
C. elegans PPI network
Dept. of Computer Science, Purdue University

13. Comparison of Semantic Similarity Measures
Proposed metric (ρJC) provides large similarity score for larger fraction of pairs at close distances (1,2),
and low similarity score for large fraction at distance>2
A comparison of the distribution of semantic similarity scores for the average information content (resnik) and
self-normalized information content (rho_JC) measures is shown.
Structural DDI network
Dept. of Computer Science, Purdue University

14. Comparison of PPI and DDI Networks
we compare the relationship between network proximity and functional similarity comprehensively, using several PPI and DDI networks.
Relation between network proximity and semantic similarity with respect to molecular function
Dept. of Computer Science, Purdue University

15. Comparison of PPI and DDI Networks
Relation between network proximity and semantic similarity with respect to biological process
Dept. of Computer Science, Purdue University

16. Comparison of PPI and DDI Networks
Immediate and Indirect neighbors perform similar functions
Functional similarity is stronger in Struct DDI network
After normalization, the relationship between functional similarity and network distance is stronger in computationally inferred DDI networks than that in PPI networks
network proximity in DDI networks is likely to be a better indicator of functional modularity, than that in PPI networks.
DDI networks that are based on structural information are relatively more reliable than PPI networks, which may come from noisy high-throughput screening.
Dept. of Computer Science, Purdue University

17. Dept. of Computer Science, Purdue University
Summary
We present necessary properties for any admissible metric for term- and set-similarity
Current metrics are not admissible, develop new metric for set-similarity
Proposed metric provides highly intuitive biological interpretation
Comprehensive comparative analysis of PPIs and DDIs validates the role of DDIs in quantifying functional coherence
Dept. of Computer Science, Purdue University