The document describes power graphs, a method for compressing and analyzing networks. Power graphs identify dense substructures like cliques and bi-cliques to represent the network in a compressed form. The algorithm runs in sub-quadratic time. Power graphs are useful for identifying master regulators in biological networks and assessing network quality. Completing incomplete bi-cliques in power graphs increases shared domains and binding sites. This approach shows promise for drug repositioning by allowing hypotheses about new drug-target-disease connections.
5. Network motifs
Hubs in networks
(stars)
Protein Complexes
(cliques)
Domain and motif-
based interactions
(bi-cliques)
Royer et al., PLoS Comp. Bio., 20085
6. Power graph algorithm compresses networks
Example: SWR1 & INO80 chromatin remodeling complexes
Before After
Modules in Networks
11. Network for mesenchymal to
neural stem cell conversion
Maisel et. al. Experimental Cell Research, 201011
12. Network for mesenchymal to
neural stem cell conversion
Maisel et. al. Experimental Cell Research, 201012
2010: miR-124
plays a role in
neural stem cell
conversion
13. 13
...repressing PTB via miR-
124 is sufficient to induce
trans-differentiation of
fibroblasts into functional
neurons
(Cell, 2013)
21. Complete and accurate networks
• Protein interactions are incomplete and noisy
• How about complete and accurate networks?
21
22. Complete and accurate networks
• Protein interactions are incomplete and noisy
• How about complete and accurate networks?
– Class hierarchy of Cytoscape,
– US Airports,
– US corporate ownership,
– Characters in Bible,
– Power grid,
– Internet routers, ...
22
24. Incomplete bi-cliques
• Power Graph are lossless
– A-B in G iff A-B in PG
• Idea: Accept small violations and
– Increase compression by adding new edges
– Completing incomplete bi-cliques
24
26. Algorithm
Find all edges e1 and e2 with n2 inside n1
Rank by score:
•Ratio total edges after (e3) to edges added (e4)
•Weight by ratio e1 to e2
•s = (e3 / e4) x (e1 / e2)
e1
e4
e3
e2
n1
n2
36. Daminell, et al. Intr. Bio., 2012
Binding sites are similar
(SMAP p-value 10-5
– 10-12
)
37. Conclusions
• Power graphs find meaningful modules
– enriched GO, PFAM, binding sites,...
– pinpoint master regulators
– can assess network quality
• Completing bi-cliques suitable for
hypotheses in drug repositioning
37
38. Acknowledgement
Jörg Heinrich,
Joachim Haupt,
Simone Daminelli
38
Former:
Matthias Reimann
Loic Royer
Collaborators:
Yixin Zhang, Aliz Emyei, BCUBE
Alexander Storch, MedFak
Francis Stewart, Biotec
Christian Pilarsky, MedFak
Robert Grützmann, MedFak
Dresden Supercomputer Department
Sainitin Donakonda,
Zerrin Isik,
Janine Roy,
Sebastian Salentin,
George Tsatsaronis,
Maria Kissa,
Daniel Eisinger,
Jan Mönnich,
Alina Petrova
FIX: Motivate the need to look a DNA repair proteins CADUC FIX: Put before and after DONE FIX: What are the 6 subunits? OK FIX: add legend DONE fiX: annimate DONE FIX: tell the story better DONE
This is the probability that the cluster has k or more proteins with domain or GO term X, if the cluster's contents were drawn randomly from the set of known proteins.
This is the probability that the cluster has k or more proteins with domain or GO term X, if the cluster's contents were drawn randomly from the set of known proteins.
Hif1a mouse knock out phenotype? MicroRNA-125b promotes neuronal differentiation in human cells by repressing multiple targets expression of either miR-124a or miR-125b increases the percentage of differentiated SH-SY5Y cells with neurite outgrowth embryonic carcinoma cells with those of differentiated neural stem cells showed that the expression level of 65 miRNAs changed (2-fold) after differentiation. MiR-124a was dramatically upregulated
Hif1a mouse knock out phenotype? MicroRNA-125b promotes neuronal differentiation in human cells by repressing multiple targets expression of either miR-124a or miR-125b increases the percentage of differentiated SH-SY5Y cells with neurite outgrowth embryonic carcinoma cells with those of differentiated neural stem cells showed that the expression level of 65 miRNAs changed (2-fold) after differentiation. MiR-124a was dramatically upregulated
How do we measure compressibility in networks? Because we have to account for the fact that even random networks are compressible. We measure the compressibility of a network relative to a random baseline. The network is first randomized many times in a way that preserves the topological properties. The network and its randomized variants are all compressed using the power graph algorithm. The compression rate of the network is compared to the average compressibility of the randomized networks. The relative compression rate is defined as the difference between the compression rate of the original network and The average compressibility.
FIX: socio-affinity explain FIX: mention databases and litterature derived networks We computed the absolute and relative compression rates for 29 networks 21 of which are derived from Y2H or AP/MS experiments 5 are multi-species databases that provide a view on the ‘ average signal ’ 2 are derived from manual curation of literature. 1 is a network derived from structure.
FIX: socio-affinity explain FIX: mention databases and litterature derived networks We computed the absolute and relative compression rates for 29 networks 21 of which are derived from Y2H or AP/MS experiments 5 are multi-species databases that provide a view on the ‘ average signal ’ 2 are derived from manual curation of literature. 1 is a network derived from structure.
We chose networks hat are accurately and completely known.