1 / 38

Network analysis of complex systems

Network analysis of complex systems. Peter Andras School of Computing Science Newcastle University peter.andras@ncl.ac.uk. Overview. Network analysis background Brain area networks Protein interaction systems Ecological system Organisations Large-scale software systems. 2. Networks.

lower
Download Presentation

Network analysis of complex systems

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Network analysis of complex systems Peter Andras School of Computing Science Newcastle University peter.andras@ncl.ac.uk

  2. Overview • Network analysis background • Brain area networks • Protein interaction systems • Ecological system • Organisations • Large-scale software systems 2

  3. Networks • Molecular interaction networks • Cellular interaction networks • Human interaction networks • Software interaction networks 3

  4. Erdos-Renyi vs Scale-free networks • Erdos-Renyi networks: uniform probability of links between any two nodes  exponential distribution of connectedness (P(k)=exp(-*k))– very few highly connected nodes • Scale-free networks: more connected nodes are more likely to be linked to other nodes  power law distribution of connectedness (P(k)=k^(-)) – some very highly connected nodes 4

  5. Real networks • Many real networks (biological, social, technological) are scale-free networks • E.g. networks of brain areas are more similar to scale-free networks than to Erdos-Renyi networks (Kaiser et al, 2007) 5

  6. Implications of being scale-free • Scale-free networks are robust to random damage, but vulnerable to well-targeted damage • Scale-free networks grow through preferential attachment 6

  7. Brain area networks • CoCoMac database – connectivity of brain areas in cat and macaque (brain areas defined in histological sense) • Connectivity ~ estimate of the number / relative importance of connecting axons • E.g. V1 receives around 5% of its inputs from LGN

  8. Are these networks scale-free ? • Networks: around 60 nodes with 600 – 800 connections – small networks • Measurements of such small size networks may be misleading (Kaiser et al, 2007)

  9. Comparison of networks • Method: • measure key parameters of these networks (average clustering coefficient and average connectivity) • generate a set of scale-free networks and a set of exponential networks with the same parameters • test statistically whether the brain networks behave in the same way or not in terms of damage measures as the random sample of scale-free or exponential networks – test both random and targeted damage

  10. Determination of scale-free-ness • The analysis shows that the brain networks are more similar to scale-free networks than to exponential networks • However, in terms of the evolution of the average clustering coefficient under targeted node elimination the brain networks are more similar to exponential networks Macaque brain network with random and targeted node elimination

  11. Protein interactions networks Aim: kill bacteria 11

  12. Important nodes • Importance – contribution to structural network integrity • Key assumption: structural and functional integrity correlates well (e.g. 80% of proteins corresponding to structurally nodes in B. subtilis are essential – Idowu & Andras, 2005) • Centrality measures: • Connectedness – Hubs • Betweenness – Bottlenecks 12

  13. Measuring damage • Integrity measures: • Average shortest path length • Average clustering coefficient • Number of isolated sub-networks • Calculation of benchmark damage – average values of integrity measures after n% of randomly selected nodes are removed • Damage effect measured as equivalent average random damage – comparability (Idowu and Andras, 2005; Idowu et al, 2004) 13

  14. Searching for drug targets • Parasite and host proteomes • Network analysis to reveal important nodes (hubs and bottlenecks) • Filtering against host proteins • Combinatorial optimisation using the equivalent damage measures • Result: pairs and triples of potential targets • Note: most single targets are already known (EU & US Patents – e-Therapeutics Plc / Newcastle University) 14

  15. Ecological system • Winter wheat field • Food-web network of plant and animal species – complemented by natural physical factors (sunlight, wind, soil type), diseases (viruses, bacteria, fungi) • Human intervention – pesticides How to avoid damaging the ecological system too much, while protecting the crop ?

  16. Network of species and other factors • Data from the Boxworth project (1992) – 184 nodes in the network: 118 species, 43 diseases, 23 other factors • 82 pesticides with 61 active ingredients • Weighted links to represent supporting and damaging interactions (asymmetric graph) • Four seasons – separate networks (Andras et al, 2007)

  17. Ecological network analysis • System integrity – integrity measures • Average shortest path length • Clustering coefficient • Benchmark curves of variation of integrity damage using averaged damage effect calculated for randomly selected sets of nodes (same size sets – e.g. 1 node, 2 nodes, etc.) together with corresponding damage variances

  18. Damage evaluation • Consider a pesticide or combination of pesticides • Remove affected nodes and their links from the network, further remove nodes and links which lack enough supporting links • Calculate integrity measures and assess the equivalent random damage by statistical testing of identity of calculated damage and the average benchmark damage

  19. Pesticide application optimisation • Consider pesticide combinations that are expected to be functionally equivalent from the perspective of crop protection • Calculate the effect of considered pesticide combinations for the whole year – for the four seasonal system graphs • Choose the combination that causes minimal expected damage

  20. Organisations – 1 • Humans working together to deliver goods and services • Organisational hierarchy: units, management

  21. Organisations – 2 • Network of human interactions • Humans act as communication units • The organisation is the dynamic network of interrelated human communications that follow a set of organisation-specific rules • E.g. rules of accounting, rules of report production, rules of addressing, etc. (Andras & Charlton, 2005)

  22. Communication networks in organisations • How to capture the communication network nature of the organisation ? • E-mail network – Enron data

  23. Organisational network dynamics • Email networks calculated for time periods (e.g. months) • Network structure dynamics

  24. Organisation analysis • Network structures – clusters • How does the structure implied by the communication network match with the formally defined structure of the organisation ? • Significant mismatches may indicate organisational problems • How is the dynamics of the network – is it diverging from or converging to the formal structure ?

  25. Communication rules • Contents analysis of communications – e.g. word stem frequencies, word consecutiveness networks, word patterns • Extraction of communication rules: pattern consecutiveness or pattern transition rules  rule base of the organisation in average

  26. Decision making analysis • Decisional processes – represented as structured rule sets • Validation of the extracted decisional process by the management • Detection of deviations  indicators of potential problems

  27. Document analysis • Co-authorship networks / co-referencing networks • Document clustering using word stem frequency vectors or word consecutiveness graphs

  28. Work and group dynamics • Dynamics of clusters of authors and papers  structural dynamics and work topic dynamics • Detection of emerging work topics, merging groups, and drying up of work topics

  29. Large-scale software systems • E.g. Microsoft Windows + MS Office, Linux + Open Office, etc. • Object oriented view: classes, messages, class instantiation  object, software in action = objects sending messages and acting according to the received messages (Andras et al, 2006)

  30. Static analysis • Network of classes linked by messages (calls) • Search for vulnerability – classes or messages that are the most critical for the integrity of the static network • E.g. hubs, bottlenecks, cluster links

  31. Dynamic analysis • The execution realisation of the software system may produce an object interaction graph that is differently weighted compared to the static class graph • Search for execution time vulnerabilities

  32. Summary – 1 • Many natural networks are scale-free networks • Such networks are robust to random damage, but sensitive to targeted damage • Brain area networks of macaque and cat are more similar to scale-free networks than to exponential networks • We looked for new antibiotic targets by analysing proteomes of bacteria. Network analysis reveals pairs/triplets of proteins that are potential drug targets. 32

  33. Summary – 2 • In ecological systems exposed to agricultural intervention food-web interaction network analysis can help to choose the least damaging intervention • Social networks: determination of ‘real’ structure and structural dynamics • Software networks: static & dynamic analysis, real software reliability evaluation, experimental model for analysis of complex systems

  34. Linking together • Common feature: building blocks linked through interactions • Building blocks: neurons, proteins, species, humans, software components • Interactions: neuron spikes, temporary protein complex formation, food-web interactions, human communications, software communications • Building blocks – interacting units; interactions modify the state of interacting units on both sides of the interaction 34

  35. Complex systems • Abstract view: interaction systems – made of interactions between interacting units (Charlton & Andras, 2007a; Andras & Charlton, 2005a; Andras & Charlton, 2005b; Andras & Andras, 2005) • Roots: social systems theory of Niklas Luhmann, biological (autopoietic) systems theory of Francisco Varela and Humberto Maturana • Complex systems – Abstract interaction systems • How do they maintain themselves by processing information about themselves and their environment ? 35

  36. Abstract interaction systems • Theory of abstract interaction systems – aim: to provide a common formal framework for theories of Luhmann, Varela & Maturana and generalize these • Current work combining inspiration from category theory interpretation of computation, dynamics of networks, pattern computation, computational learning, and other areas 36

  37. Acknowledgements • Network analysis: • Marcus Kaiser, OlusolaIdowu, Malcolm P Young, Panayiotis Periorelis, Agnes Madalinski, Steven Lynden, Robert Gwyther, Hermann Moisl, Abigail Haig, Greg Maniatopoulos, Will McElderry • Complex systems theory • Bruce G Charlton 37

  38. Thank you! 38

More Related